Jeffrey Shainline: Neuromorphic Computing and Optoelectronic Intelligence

00:00:00.000 | The following is a conversation with Jeff Shainlein,

00:00:03.060 | a scientist at NIST

00:00:04.600 | interested in optoelectronic intelligence.

00:00:08.440 | We have a deep technical dive into computing hardware

00:00:12.160 | that will make Jim Keller proud.

00:00:14.080 | I urge you to hop onto this rollercoaster ride

00:00:17.440 | through neuromorphic computing

00:00:19.320 | and superconducting electronics

00:00:21.720 | and hold on for dear life.

00:00:24.360 | Jeff is a great communicator of technical information.

00:00:27.200 | And so it was truly a pleasure to talk to him

00:00:30.160 | about some physics and engineering.

00:00:32.300 | To support this podcast,

00:00:34.440 | please check out our sponsors in the description.

00:00:37.440 | This is the Lex Friedman Podcast

00:00:39.880 | and here is my conversation with Jeff Shainlein.

00:00:43.360 | I got a chance to read a fascinating paper you authored

00:00:48.920 | called Optoelectronic Intelligence.

00:00:52.040 | So maybe we can start by talking about this paper

00:00:55.200 | and start with the basic questions.

00:00:57.060 | What is optoelectronic intelligence?

00:01:00.360 | - Yeah, so in that paper,

00:01:02.040 | the concept I was trying to describe

00:01:04.280 | is sort of an architecture for building

00:01:08.420 | brain-inspired computing

00:01:10.820 | that leverages light for communication

00:01:13.380 | in conjunction with electronic circuits for computation.

00:01:17.640 | In that particular paper,

00:01:18.920 | a lot of the work we're doing right now

00:01:20.880 | in our project at NIST

00:01:22.000 | is focused on superconducting electronics for computation.

00:01:25.820 | I'll go into why that is,

00:01:27.720 | but that might make a little more sense in context

00:01:31.420 | if we first describe what that is in contrast to,

00:01:35.420 | which is semiconducting electronics.

00:01:37.880 | So is it worth taking a couple minutes

00:01:39.720 | to describe semiconducting electronics?

00:01:42.820 | - It might even be worthwhile to step back

00:01:45.960 | and talk about electricity and circuits

00:01:49.960 | and how circuits work

00:01:52.240 | before we talk about superconductivity.

00:01:54.760 | - Right, okay.

00:01:56.460 | - How does a computer work, Jeff?

00:01:58.060 | - Well, I won't go into everything

00:01:59.780 | that makes a computer work,

00:02:01.100 | but let's talk about the basic building blocks,

00:02:05.180 | a transistor, and even more basic than that,

00:02:08.460 | a semiconductor material, silicon, say.

00:02:11.500 | So in silicon, silicon is a semiconductor,

00:02:15.460 | and what that means is at low temperature,

00:02:18.460 | there are no free charges,

00:02:20.780 | no free electrons that can move around.

00:02:22.860 | So when you talk about electricity,

00:02:24.920 | you're talking about predominantly electrons

00:02:28.040 | moving to establish electrical currents,

00:02:30.780 | and they move under the influence of voltages.

00:02:33.080 | So you apply voltages, electrons move around,

00:02:36.620 | those can be measured as currents,

00:02:38.400 | and you can represent information in that way.

00:02:40.900 | So semiconductors are special

00:02:43.840 | in the sense that they are really malleable.

00:02:46.860 | So if you have a semiconductor material,

00:02:49.120 | you can change the number of free electrons

00:02:52.600 | that can move around by putting different elements,

00:02:56.020 | different atoms in lattice sites.

00:02:58.280 | So what is a lattice site?

00:03:00.200 | Well, a semiconductor is a crystal,

00:03:02.200 | which means all the atoms that comprise the material

00:03:06.480 | are at exact locations

00:03:09.000 | that are perfectly periodic in space.

00:03:10.880 | So if you started any one atom

00:03:12.200 | and you go along what are called the lattice vectors,

00:03:14.660 | you get to another atom and another atom and another atom,

00:03:17.320 | and for high-quality devices,

00:03:19.440 | it's important that it's a perfect crystal

00:03:21.920 | with very few defects,

00:03:23.800 | but you can intentionally replace a silicon atom

00:03:27.440 | with say a phosphorus atom,

00:03:29.240 | and then you can change the number of free electrons

00:03:32.160 | that are in a region of space

00:03:33.980 | that has that excess of what are called dopants.

00:03:37.140 | So picture a device that has a left terminal

00:03:40.260 | and a right terminal,

00:03:42.040 | and if you apply a voltage between those two,

00:03:44.160 | you can cause electrical current to flow between them.

00:03:47.760 | Now we add a third terminal up on top there,

00:03:52.120 | and depending on the voltage

00:03:53.800 | between the left and right terminal and that third voltage,

00:03:57.160 | you can change that current.

00:03:58.720 | So what's commonly done in digital electronic circuits

00:04:02.280 | is to leave a fixed voltage from left to right,

00:04:06.560 | and then change that voltage that's applied

00:04:09.080 | at what's called the gate, the gate of the transistor.

00:04:11.200 | So what you do is you make it to where

00:04:13.960 | there's an excess of electrons on the left,

00:04:15.680 | excess of electrons on the right,

00:04:17.360 | and very few electrons in the middle,

00:04:18.880 | and you do this by changing the concentration

00:04:21.520 | of different dopants in the lattice spatially.

00:04:24.800 | And then when you apply a voltage to that gate,

00:04:27.880 | you can either cause current to flow or turn it off.

00:04:30.680 | And so that's sort of your zero and one.

00:04:33.440 | If you apply voltage, current can flow,

00:04:35.360 | that current is representing a digital one,

00:04:38.360 | and from that, from that basic element,

00:04:41.960 | you can build up all the complexity

00:04:44.180 | of digital electronic circuits

00:04:45.520 | that have really had a profound influence on our society.

00:04:50.080 | - Now you're talking about electrons.

00:04:51.600 | Can you give a sense of what scale we're talking about

00:04:54.680 | when we're talking about in silicon

00:04:56.520 | being able to mass manufacture these kinds of gates?

00:05:01.560 | - Yeah, so scale in a number of different senses.

00:05:04.480 | Well, at the scale of the silicon lattice,

00:05:07.520 | the distance between two atoms there is half a nanometer.

00:05:10.880 | So people often like to compare these things

00:05:14.480 | to the width of a human hair.

00:05:16.520 | I think it's some six orders of magnitude smaller

00:05:20.020 | than the width of a human hair, something on that order.

00:05:24.440 | So remarkably small,

00:05:25.840 | we're talking about individual atoms here,

00:05:27.280 | and electrons are of that length scale

00:05:29.480 | when they're in that environment.

00:05:31.520 | But there's another sense that scale matters

00:05:33.400 | in digital electronics.

00:05:34.380 | This is perhaps the more important sense,

00:05:36.460 | although they're related.

00:05:37.880 | Scale refers to a number of things.

00:05:41.360 | It refers to the size of that transistor.

00:05:44.560 | So for example, I said you have a left contact,

00:05:47.280 | a right contact, and some space between them

00:05:49.800 | where the gate electrode sits.

00:05:52.240 | That's called the channel width or the channel length.

00:05:56.000 | And what has enabled what we think of as Moore's law

00:06:00.120 | or the continued increased performance

00:06:03.040 | in silicon microelectronic circuits

00:06:05.560 | is the ability to make that size,

00:06:07.940 | that feature size ever smaller, ever smaller.

00:06:10.480 | At a really remarkable pace.

00:06:14.800 | I mean, that feature size has decreased consistently

00:06:19.800 | every couple of years since the 1960s.

00:06:23.760 | And that was what Moore predicted in the 1960s.

00:06:27.160 | He thought it would continue for at least two more decades,

00:06:29.480 | and it's been much longer than that.

00:06:30.800 | And so that is why we've been able to fit ever more devices,

00:06:35.800 | ever more transistors, ever more computational power

00:06:39.180 | on essentially the same size of chip.

00:06:41.640 | So a user sits back and does essentially nothing.

00:06:44.600 | You're running the same computer program,

00:06:46.160 | but those devices are getting smaller,

00:06:47.680 | so they get faster, they get more energy efficient,

00:06:50.280 | and all of our computing performance

00:06:51.700 | just continues to improve.

00:06:53.440 | And we don't have to think too hard

00:06:56.400 | about what we're doing as say a software designer

00:07:00.280 | or something like that.

00:07:01.120 | I absolutely don't mean to say that there's no innovation

00:07:03.720 | in software or the user side of things.

00:07:06.400 | Of course there is,

00:07:07.240 | but from the hardware perspective,

00:07:09.580 | we just have been given this gift

00:07:12.460 | of continued performance improvement through this scaling

00:07:16.180 | that is ever smaller feature sizes

00:07:19.300 | with very similar, say, power consumption.

00:07:22.900 | That power consumption has not continued to scale

00:07:26.100 | in the most recent decades,

00:07:27.500 | but nevertheless, we had a really good run there for a while

00:07:31.420 | and now we're down to gates that are seven nanometers,

00:07:34.380 | which is state-of-the-art right now,

00:07:36.300 | maybe Global Foundries is trying to push it

00:07:38.940 | even lower than that.

00:07:39.780 | I can't keep up with where the predictions are

00:07:42.600 | that it's gonna end,

00:07:43.440 | but seven nanometer transistor

00:07:47.640 | has just a few tens of atoms

00:07:50.080 | along the length of the conduction pathway.

00:07:51.920 | So a naive semiconductor device physicist

00:07:56.800 | would think you can't go much further than that

00:07:58.760 | without some kind of revolution

00:08:01.060 | in the way we think about the physics of our devices.

00:08:04.000 | - Is there something to be said

00:08:05.140 | about the mass manufacture of these devices?

00:08:08.300 | - Right, right, so that's another thing.

00:08:09.500 | So how have we been able to make those transistors

00:08:13.100 | smaller and smaller?

00:08:14.020 | Well, companies like Intel, Global Foundries,

00:08:17.900 | they invest a lot of money in the lithography.

00:08:20.460 | So how are these chips actually made?

00:08:22.980 | Well, one of the most important steps

00:08:24.500 | is this what's called ion implantation.

00:08:27.840 | So you start with sort of a pristine silicon crystal

00:08:32.100 | and then using photolithography,

00:08:34.240 | which is a technique where you can pattern different shapes

00:08:37.560 | using light, you can define which regions of space

00:08:41.560 | you're going to implant with different species of ions

00:08:45.800 | that are going to change

00:08:46.760 | the local electrical properties right there.

00:08:49.700 | So by using ever shorter wavelengths of light

00:08:52.840 | and different kinds of optical techniques

00:08:54.540 | and different kinds of lithographic techniques,

00:08:56.760 | things that go far beyond my knowledge base,

00:09:01.140 | you can just simply shrink that feature size down.

00:09:03.520 | And you say you're at seven nanometers.

00:09:04.860 | Well, the wavelength of light that's being used

00:09:07.440 | is over a hundred nanometers.

00:09:08.800 | That's already deep in the UV.

00:09:10.280 | So how are those minute features patterned?

00:09:14.760 | Well, there's an extraordinary amount of innovation

00:09:16.960 | that has gone into that, but nevertheless,

00:09:19.000 | it stayed very consistent

00:09:20.320 | in this ever shrinking feature size.

00:09:21.940 | And now the question is, can you make it smaller?

00:09:24.800 | And even if you do,

00:09:25.960 | do you still continue to get performance improvements?

00:09:28.360 | But that's another kind of scaling

00:09:30.280 | where these companies have been able to...

00:09:34.320 | So, okay, you picture a chip that has a processor on it.

00:09:37.080 | Well, that chip is not made as a chip.

00:09:38.820 | It's made on a wafer.

00:09:40.920 | And using photolithography,

00:09:43.400 | you basically print the same pattern

00:09:46.200 | on different dyes all across the wafer,

00:09:48.360 | multiple layers, tens, probably a hundred some layers

00:09:53.360 | in a mature foundry process.

00:09:55.260 | And you do this on ever bigger wafers too.

00:09:57.620 | That's another aspect of scaling

00:09:59.060 | that's occurred in the last several decades.

00:10:00.940 | So now you have this 300 millimeter wafer.

00:10:03.000 | It's like as big as a pizza

00:10:04.320 | and it has maybe a thousand processors on it.

00:10:06.560 | And then you dice that up using a saw.

00:10:08.920 | And now you can sell these things so cheap

00:10:11.800 | because the manufacturing process was so streamlined.

00:10:15.200 | I think a technology as revolutionary

00:10:17.440 | as Silicon Microelectronics

00:10:19.360 | has to have that kind of manufacturing scalability,

00:10:23.760 | which I will just emphasize,

00:10:25.720 | I believe is enabled by physics.

00:10:29.220 | It's not, I mean, of course there's human ingenuity

00:10:31.780 | that goes into it, but at least from my side, where I sit,

00:10:36.100 | it sure looks like the physics of our universe

00:10:39.020 | allows us to produce that.

00:10:41.140 | And we've discovered how more so than we've invented it.

00:10:45.980 | Although of course we have invented it.

00:10:47.420 | Humans have invented it,

00:10:48.720 | but it's almost as if it was there

00:10:51.220 | waiting for us to discover it.

00:10:53.220 | - You mean the entirety of it

00:10:54.420 | or are you specifically talking about

00:10:56.180 | the techniques of photolithography,

00:10:58.620 | like the optics involved?

00:10:59.960 | I mean the entirety of the scaling down

00:11:02.860 | to the seven nanometers,

00:11:04.620 | you're able to have electrons not interfere with each other

00:11:08.500 | in such a way that you could still have gates.

00:11:11.580 | Like that's enabled to achieve that scale,

00:11:14.700 | spatial and temporal.

00:11:16.860 | It seems to be very special

00:11:18.700 | and is enabled by the physics of our world.

00:11:21.480 | - All of the things you just said.

00:11:22.760 | So starting with the silicon material itself,

00:11:25.940 | silicon is a unique semiconductor.

00:11:28.780 | It has essentially ideal properties

00:11:31.540 | for making a specific kind of transistor

00:11:33.700 | that's extraordinarily useful.

00:11:35.220 | So I mentioned that silicon has,

00:11:39.540 | well, when you make a transistor,

00:11:40.620 | you have this gate contact

00:11:42.020 | that sits on top of the conduction channel.

00:11:44.540 | And depending on the voltage you apply there,

00:11:47.260 | you pull more carriers into the conduction channel

00:11:50.020 | or push them away so it becomes more or less conductive.

00:11:53.060 | In order to have that work

00:11:54.700 | without just sucking those carriers right into that contact,

00:11:57.120 | you need a very thin insulator.

00:11:59.260 | And part of scaling has been to gradually decrease

00:12:03.220 | the thickness of that gate insulator

00:12:06.060 | so that you can use a roughly similar voltage

00:12:09.060 | and still have the same current voltage characteristics.

00:12:12.080 | So the material that's used to do that,

00:12:14.600 | or I should say was initially used to do that,

00:12:16.900 | was silicon dioxide,

00:12:18.720 | which just naturally grows on the silicon surface.

00:12:21.820 | So you expose silicon to the atmosphere that we breathe.

00:12:25.100 | And well, if you're manufacturing,

00:12:27.580 | you're gonna purify these gases,

00:12:29.740 | but nevertheless, what's called a native oxide

00:12:32.460 | will grow there.

00:12:33.500 | There are essentially no other materials

00:12:36.060 | on the entire periodic table

00:12:37.500 | that have as good of a gate insulator

00:12:41.900 | as that silicon dioxide.

00:12:43.660 | And that has to do with nothing but the physics

00:12:46.260 | of the interaction between silicon and oxygen.

00:12:49.240 | And if it wasn't that way,

00:12:51.240 | transistors could not,

00:12:53.040 | they could not perform in nearly the degree of capability

00:12:57.960 | that they have.

00:12:58.800 | And that has to do with the way that the oxide grows,

00:13:02.880 | the reduced density of defects there,

00:13:05.760 | its insulation, meaning essentially its energy gaps.

00:13:08.600 | You can apply a very large voltage there

00:13:10.320 | without having current leak through it.

00:13:12.260 | So that's physics right there.

00:13:15.760 | There are other things too.

00:13:16.880 | Silicon is a semiconductor in an elemental sense.

00:13:20.040 | You only need silicon atoms.

00:13:21.560 | A lot of other semiconductors,

00:13:22.820 | you need two different kinds of atoms,

00:13:24.880 | like a compound from group three

00:13:26.800 | and a compound from group five.

00:13:28.580 | That opens you up to lots of defects that can occur

00:13:31.680 | where one atom's not sitting quite at the lattice site it is

00:13:34.640 | and it's switched with another one.

00:13:35.960 | That degrades performance.

00:13:38.280 | But then also on the side that you mentioned

00:13:40.280 | with the manufacturing,

00:13:43.480 | we have access to light sources

00:13:45.840 | that can produce these very short wavelengths of light.

00:13:48.560 | How does photolithography occur?

00:13:50.880 | Well, you actually put this polymer on top of your wafer

00:13:54.460 | and you expose it to light.

00:13:56.200 | And then you use a aqueous chemical processing

00:14:00.080 | to dissolve away the regions that were exposed to light

00:14:03.440 | and leave the regions that were not.

00:14:05.960 | And we are blessed with these polymers

00:14:08.560 | that have the right property

00:14:09.680 | where they can cause scission events

00:14:13.600 | where the polymer splits where a photon hits.

00:14:16.200 | I mean, maybe that's not too surprising,

00:14:19.760 | but I don't know.

00:14:20.600 | It all comes together to have this

00:14:22.560 | really complex manufacturable ecosystem

00:14:26.160 | where very sophisticated technologies can be devised

00:14:30.640 | and it works quite well.

00:14:33.040 | - And amazingly, like you said,

00:14:34.440 | with a wavelength of like 100 nanometers

00:14:36.280 | or something like that,

00:14:37.120 | you're still able to achieve on this polymer

00:14:39.760 | precision of whatever we said, seven nanometers.

00:14:43.520 | I think I've heard like four nanometers

00:14:45.800 | being talked about, something like that.

00:14:48.440 | If we could just pause on this

00:14:49.880 | and we'll return to superconductivity,

00:14:52.160 | but in this whole journey from a history perspective,

00:14:56.420 | what do you think is the most beautiful

00:14:58.440 | at the intersection of engineering and physics

00:15:01.760 | to you in this whole process

00:15:03.280 | that we talked about with silicon and photolithography,

00:15:06.560 | things that people were able to achieve

00:15:08.360 | in order to push the Moore's law forward?

00:15:12.320 | Is it the early days,

00:15:13.680 | the invention of the transistor itself?

00:15:16.280 | Is it some particular cool little thing

00:15:19.320 | that maybe not many people know about?

00:15:22.000 | Like, what do you think is the most beautiful

00:15:23.400 | in this whole process journey?

00:15:26.800 | - The most beautiful is a little difficult to answer.

00:15:29.600 | Let me try and sidestep it a little bit

00:15:32.040 | and just say what strikes me

00:15:34.760 | about looking at the history of silicon microelectronics

00:15:39.760 | is that, so when quantum mechanics was developed,

00:15:44.580 | people quickly began applying it to semiconductors

00:15:47.440 | and it was broadly understood

00:15:49.360 | that these are fascinating systems

00:15:50.760 | and people cared about them for their basic physics,

00:15:52.720 | but also their utility as devices.

00:15:55.040 | And then the transistor was invented in the late 40s

00:15:59.260 | in a relatively crude experimental setup

00:16:02.040 | where you just crammed a metal electrode

00:16:04.240 | into the semiconductor and that was ingenious.

00:16:08.000 | These people were able to make it work, you know?

00:16:12.960 | But so what I wanna get to that really strikes me

00:16:16.800 | is that in those early days,

00:16:19.300 | there were a number of different semiconductors

00:16:21.160 | that were being considered.

00:16:22.080 | They had different properties, different strengths,

00:16:23.800 | different weaknesses.

00:16:24.880 | Most people thought germanium was the way to go.

00:16:28.440 | It had some nice properties related to things

00:16:33.640 | about how the electrons move inside the lattice,

00:16:37.280 | but other people thought that compound semiconductors

00:16:39.760 | with group three and group five

00:16:41.060 | also had really, really extraordinary properties

00:16:44.920 | that might be conducive to making the best devices.

00:16:50.060 | So there were different groups exploring each of these

00:16:52.520 | and that's great, that's how science works.

00:16:54.220 | You have to cast a broad net.

00:16:56.120 | But then what I find striking is why is it that silicon won?

00:17:02.120 | Because it's not that germanium is a useless material

00:17:05.240 | and it's not present in technology

00:17:06.720 | or compound semiconductors.

00:17:08.060 | They're both doing exciting and important things,

00:17:12.600 | slightly more niche applications,

00:17:14.360 | whereas silicon is the semiconductor material

00:17:18.040 | for microelectronics, which is the platform

00:17:20.060 | for digital computing, which has transformed our world.

00:17:22.720 | Why did silicon win?

00:17:24.160 | It's because of a remarkable assemblage of qualities

00:17:28.720 | that no one of them was the clear winner,

00:17:32.120 | but it made these sort of compromises

00:17:34.580 | between a number of different influences.

00:17:36.680 | It had that really excellent gate oxide

00:17:40.520 | that allowed us to make MOSFETs,

00:17:43.240 | these high-performance transistors,

00:17:45.400 | so quickly and cheaply and easily

00:17:47.200 | without having to do a lot of materials development.

00:17:49.560 | The band gap of silicon is actually...

00:17:53.400 | So in a semiconductor, there's an important parameter,

00:17:56.240 | which is called the band gap,

00:17:57.440 | which tells you if you...

00:17:59.220 | There are sort of electrons that fill up to one level

00:18:02.280 | in the energy diagram, and then there's a gap

00:18:05.560 | where electrons aren't allowed to have an energy

00:18:07.480 | in a certain range,

00:18:08.320 | and then there's another energy level above that.

00:18:11.300 | And that difference between the lower sort of filled level

00:18:14.920 | and the unoccupied level,

00:18:16.860 | that tells you how much voltage you have to apply

00:18:19.600 | in order to induce a current to flow.

00:18:22.120 | So with germanium, that's about 0.75 electron volts.

00:18:27.260 | That means you have to apply 0.75 volts

00:18:29.600 | to get a current moving.

00:18:31.960 | And it turns out that if you compare that

00:18:34.240 | to the thermal excitations that are induced

00:18:38.480 | just by the temperature of our environment,

00:18:40.660 | that gap's not quite big enough.

00:18:42.080 | You start to use it to perform computations,

00:18:45.080 | it gets a little hot,

00:18:45.940 | and you get all these accidental carriers

00:18:48.120 | that are excited into the conduction band,

00:18:50.720 | and it causes errors in your computation.

00:18:53.340 | Silicon's band gap is just a little higher,

00:18:56.180 | 1.1 electron volts,

00:18:58.960 | but you have an exponential dependence

00:19:01.300 | on the number of carriers that are present

00:19:04.220 | that can induce those errors.

00:19:06.640 | It decays exponentially with that voltage.

00:19:08.480 | So just that slight extra energy in that band gap

00:19:12.780 | really puts it in an ideal position to be operated

00:19:17.040 | in the conditions of our ambient environment.

00:19:20.240 | - It's kind of fascinating that,

00:19:21.440 | so like you mentioned,

00:19:22.480 | errors decrease exponentially with the voltage.

00:19:27.480 | So it's funny 'cause this error thing comes up

00:19:32.020 | when you start talking about quantum computing.

00:19:34.620 | It's kind of amazing that everything

00:19:36.020 | we've been talking about,

00:19:37.060 | the errors as we scale down,

00:19:39.480 | seems to be extremely low.

00:19:41.500 | - Yes.

00:19:42.340 | - And like all of our computation

00:19:45.380 | is based on the assumption that it's extremely low.

00:19:47.740 | - Yes, well it's digital computation.

00:19:49.540 | - Digital, sorry, digital computation.

00:19:51.460 | So as opposed to our biological computation,

00:19:54.260 | our brain is like,

00:19:55.700 | the assumption is stuff is gonna fail all over the place

00:19:59.100 | and we somehow have to still be robust to that.

00:20:01.740 | - That's exactly right.

00:20:03.060 | - So this also,

00:20:04.180 | this is gonna be the most controversial part

00:20:05.820 | of our conversation where you're gonna make some enemies.

00:20:07.780 | So let me ask,

00:20:09.040 | 'cause we've been talking about physics and engineering.

00:20:11.840 | Which group of people is smarter

00:20:15.740 | and more important for this one?

00:20:17.780 | Let me ask the question in a better way.

00:20:20.540 | Some of the big innovations,

00:20:22.580 | some of the beautiful things that we've been talking about,

00:20:25.740 | how much of it is physics?

00:20:26.980 | How much of it is engineering?

00:20:28.420 | My dad is a physicist

00:20:30.260 | and he talks down to all the amazing engineering

00:20:33.260 | that we're doing in the artificial intelligence

00:20:36.940 | and the computer science and the robotics

00:20:38.780 | and all that space.

00:20:39.620 | So we argue about this all the time.

00:20:41.660 | So what do you think?

00:20:42.500 | Who gets more credit?

00:20:43.920 | - I'm genuinely not trying

00:20:45.420 | to just be politically correct here.

00:20:46.780 | I don't see how you would have

00:20:49.060 | any of the, what we consider

00:20:51.180 | sort of the great accomplishments of society without both.

00:20:54.140 | You absolutely need both of those things.

00:20:55.860 | Physics tends to play a key role earlier in the development

00:20:59.660 | and then engineering, optimization, these things take over.

00:21:04.660 | And I mean, the invention of the transistor,

00:21:09.220 | or actually even before that,

00:21:10.860 | the understanding of semiconductor physics

00:21:12.860 | that allowed the invention of the transistor,

00:21:14.800 | that's all physics.

00:21:15.640 | So if you didn't have that physics,

00:21:16.940 | you don't even get to get on the field.

00:21:20.260 | But once you have understood and demonstrated

00:21:24.180 | that this is in principle possible,

00:21:26.500 | more so as engineering,

00:21:28.380 | why we have computers more powerful

00:21:32.140 | than old supercomputers in each of our phones,

00:21:36.380 | that's all engineering.

00:21:37.460 | And I think I would be quite foolish to say that,

00:21:42.160 | that that's not valuable,

00:21:44.980 | that that's not a great contribution.

00:21:46.940 | - It's a beautiful dance.

00:21:47.820 | Would you put like silicon,

00:21:49.740 | the understanding of the material properties

00:21:52.780 | in the space of engineering?

00:21:54.340 | Like how does that whole process work

00:21:55.700 | to understand that it has all these nice properties

00:21:58.000 | or even the development of photolithography?

00:22:02.240 | Is that basically,

00:22:03.820 | would you put that in a category of engineering?

00:22:06.140 | - No, I would say that it is basic physics,

00:22:09.940 | it is applied physics, it's material science,

00:22:12.740 | it's x-ray crystallography,

00:22:16.060 | it's polymer chemistry, it's everything.

00:22:18.580 | - So chemistry even is thrown in there?

00:22:20.260 | - Absolutely, yes.

00:22:21.580 | Yes, absolutely.

00:22:22.580 | - Just no biology.

00:22:23.840 | (laughing)

00:22:25.340 | - We can get to biology.

00:22:26.580 | - Well, the biology is in the humans

00:22:28.260 | that are engineering the system,

00:22:29.500 | so it's all integrated deeply.

00:22:31.220 | Okay, so let's return,

00:22:32.580 | you mentioned this word superconductivity.

00:22:35.660 | So what does that have to do with what we're talking about?

00:22:38.620 | - Right, okay, so in a semiconductor,

00:22:40.540 | as I tried to describe a second ago,

00:22:44.260 | you can sort of induce currents by applying voltages

00:22:49.260 | and those have sort of typical properties

00:22:52.280 | that you would expect from some kind of a conductor.

00:22:55.400 | Those electrons,

00:22:57.060 | they don't just flow perfectly without dissipation.

00:23:00.360 | If an electron collides with an imperfection in the lattice

00:23:03.240 | or another electron, it's gonna slow down,

00:23:05.580 | it's gonna lose its momentum.

00:23:06.860 | So you have to keep applying that voltage

00:23:09.260 | in order to keep the current flowing.

00:23:11.420 | In a superconductor, something different happens.

00:23:13.420 | If you get a current to start flowing,

00:23:16.540 | it will continue to flow indefinitely.

00:23:18.260 | There's no dissipation.

00:23:19.860 | So that's crazy.

00:23:21.660 | How does that happen?

00:23:22.500 | Well, it happens at low temperature and this is crucial.

00:23:26.820 | It has to be a quite low temperature

00:23:30.200 | and what I'm talking about there,

00:23:32.780 | for essentially all of our conversation,

00:23:35.780 | I'm gonna be talking about conventional superconductors

00:23:39.500 | sometimes called low TC superconductors,

00:23:42.220 | low critical temperature superconductors.

00:23:45.120 | And so those materials have to be at a temperature

00:23:49.820 | around, say around four Kelvin.

00:23:51.960 | I mean, their critical temperature might be 10 Kelvin,

00:23:54.560 | something like that,

00:23:55.400 | but you wanna operate them at around four Kelvin,

00:23:57.120 | four degrees above absolute zero.

00:23:59.300 | And what happens at that temperature,

00:24:01.640 | at very low temperatures in certain materials

00:24:03.900 | is that the noise of atoms moving around,

00:24:08.900 | the lattice vibrating,

00:24:11.900 | electrons colliding with each other,

00:24:13.580 | that becomes sufficiently low

00:24:15.100 | that the electrons can settle into this very special state.

00:24:18.580 | It's sometimes referred to as a macroscopic quantum state

00:24:22.260 | because if I had a piece of superconducting material here,

00:24:26.560 | let's say niobium is a very typical superconductor.

00:24:30.940 | If I had a block of niobium here

00:24:33.620 | and we cooled it below its critical temperature,

00:24:36.700 | all of the electrons in that superconducting state

00:24:40.820 | would be in one coherent quantum state.

00:24:42.860 | The wave function of that state

00:24:46.500 | is described in terms of all of the particles simultaneously

00:24:49.740 | but it extends across macroscopic dimensions,

00:24:52.220 | the size of whatever material,

00:24:53.700 | the size of whatever block of that material

00:24:56.100 | I have sitting here.

00:24:57.060 | And the way this occurs is that,

00:24:59.440 | let's try to be a little bit light on the technical details,

00:25:03.480 | but essentially the electrons coordinate with each other.

00:25:06.500 | They are able to, in this macroscopic quantum state,

00:25:10.300 | they're able to sort of,

00:25:12.260 | one can quickly take the place of the other.

00:25:14.380 | You can't tell electrons apart,

00:25:15.780 | they're what's known as identical particles.

00:25:17.740 | So if this electron runs into a defect

00:25:22.020 | that would otherwise cause it to scatter,

00:25:24.980 | it can just sort of almost miraculously avoid that defect

00:25:29.980 | because it's not really in that location,

00:25:32.220 | it's part of a macroscopic quantum state

00:25:33.960 | and the entire quantum state

00:25:35.640 | was not scattered by that defect.

00:25:37.100 | So you can get a current that flows without dissipation

00:25:40.720 | and that's called a super current.

00:25:42.920 | That's sort of just very much scratching the surface

00:25:47.400 | of superconductivity.

00:25:49.800 | There's very deep and rich physics there,

00:25:52.200 | just probably not the main subject

00:25:54.480 | we need to go into right now,

00:25:55.600 | but it turns out that when you have this material,

00:26:00.480 | you can do usual things like make wires out of it

00:26:03.540 | so you can get current to flow

00:26:05.040 | in a straight line on a chip,

00:26:06.400 | but you can also make other devices

00:26:08.880 | that perform different kinds of operations.

00:26:11.860 | Some of them are kind of logic operations

00:26:14.720 | like you'd get in a transistor.

00:26:16.680 | The most common or most, I would say,

00:26:21.680 | diverse in its utility component is a Josephson junction.

00:26:26.880 | It's not analogous to a transistor in the sense that

00:26:30.000 | if you apply a voltage here,

00:26:31.440 | it changes how much current flows from left to right,

00:26:33.840 | but it is analogous in sort of a sense of

00:26:37.160 | it's the go-to component that a circuit engineer

00:26:41.020 | is going to use to start to build up more complexity.

00:26:44.480 | - So these junctions serve as gates?

00:26:48.800 | - They can serve as gates.

00:26:50.600 | I'm not sure how concerned to be with semantics,

00:26:56.520 | but let me just briefly say what a Josephson junction is

00:26:59.440 | and we can talk about different ways that they can be used.

00:27:02.240 | Basically, if you have a superconducting wire

00:27:05.280 | and then a small gap of a different material

00:27:09.680 | that's not superconducting, an insulator or normal metal,

00:27:13.520 | and then another superconducting wire on the other side,

00:27:15.800 | that's a Josephson junction.

00:27:17.020 | So it's sometimes referred to

00:27:18.600 | as a superconducting weak link.

00:27:20.320 | So you have this superconducting state on one side

00:27:24.200 | and on the other side

00:27:25.080 | and the superconducting wave function

00:27:27.520 | actually tunnels across that gap.

00:27:30.680 | And when you create such a physical entity,

00:27:35.440 | it has very unusual current voltage characteristics.

00:27:40.440 | - Within that gap, like weird stuff happens.

00:27:44.280 | - Through the entire circuit.

00:27:45.120 | So you can imagine, suppose you had a loop setup

00:27:47.560 | that had one of those weak links in the loop.

00:27:51.200 | Current would flow in that loop

00:27:53.200 | independent even if you hadn't applied a voltage to it

00:27:55.600 | and that's called the Josephson effect.

00:27:57.000 | So the fact that there's this phase difference

00:28:00.520 | in the quantum wave function

00:28:02.160 | from one side of the tunneling barrier to the other

00:28:04.400 | induces current to flow.

00:28:05.760 | - So how does he change state?

00:28:07.720 | - Right, exactly.

00:28:08.560 | So how do you change state?

00:28:09.580 | Now picture if I have a current bias coming down

00:28:13.900 | this line of my circuit

00:28:14.920 | and there's a Josephson junction right in the middle of it.

00:28:18.120 | And now I make another wire

00:28:19.960 | that goes around the Josephson junction.

00:28:21.800 | So I have a loop here, a superconducting loop.

00:28:24.600 | I can add current to that loop

00:28:27.960 | by exceeding the critical current of that Josephson junction.

00:28:30.840 | So like any superconducting material,

00:28:33.960 | it can carry this super current that I've described,

00:28:37.680 | this current that can propagate without dissipation

00:28:40.520 | up to a certain level.

00:28:41.840 | And if you try and pass more current than that

00:28:44.240 | through the material,

00:28:45.920 | it's going to become a resistive material,

00:28:49.200 | a normal material.

00:28:51.120 | So in the Josephson junction, the same thing happens.

00:28:54.120 | I can bias it above its critical current.

00:28:57.120 | And then what it's gonna do,

00:28:58.200 | it's going to add a quantized amount of current

00:29:03.200 | into that loop.

00:29:04.360 | And what I mean by quantized is

00:29:06.320 | it's going to come in discrete packets

00:29:08.980 | with a well-defined value of current.

00:29:11.200 | So in the vernacular of some people

00:29:15.040 | working in this community,

00:29:16.860 | you would say you pop a flux on into the loop.

00:29:20.200 | So a flux on--

00:29:21.720 | - You pop a flux on into the loop.

00:29:23.680 | - Yeah, so a flux on--

00:29:24.520 | - Sounds like skateboarder talk, I love it.

00:29:26.480 | (Lex laughing)

00:29:27.440 | Sorry, go ahead.

00:29:28.920 | - A flux on is one of these quantized

00:29:31.200 | amounts of current that you can add to a loop.

00:29:35.200 | And this is a cartoon picture,

00:29:36.600 | but I think it's sufficient for our purposes.

00:29:38.440 | - So which, maybe it's useful to say,

00:29:40.780 | what is the speed at which

00:29:43.720 | these discrete packets of current travel?

00:29:47.000 | Because we'll be talking about light a little bit.

00:29:49.160 | It seems like the speed is important.

00:29:51.080 | - The speed is important, that's an excellent question.

00:29:53.560 | Sometimes I wonder how you became so astute.

00:29:57.800 | But so this--

00:30:00.800 | - Matrix 4 is coming out, so maybe that's related.

00:30:04.400 | I'm not sure.

00:30:05.240 | I'm dressed for the job.

00:30:06.920 | I was trying to become an extra on Matrix 4,

00:30:09.360 | it didn't work out.

00:30:10.680 | Anyway, so what's the speed of these packets?

00:30:13.280 | - You'll have to find another gig.

00:30:15.000 | - I know, I'm sorry.

00:30:16.600 | - So the speed of the packet is actually these flux ons,

00:30:19.400 | these sort of pulses of current

00:30:24.280 | that are generated by Joseph's injunctions.

00:30:26.160 | They can actually propagate

00:30:27.640 | very close to the speed of light,

00:30:29.720 | maybe something like a third of the speed of light.

00:30:31.880 | That's quite fast.

00:30:32.840 | So one of the reasons why Joseph's injunctions are appealing

00:30:37.240 | is because their signals can propagate quite fast

00:30:40.560 | and they can also switch very fast.

00:30:43.400 | What I mean by switch is perform that operation

00:30:46.040 | that I described where you add current to the loop.

00:30:49.400 | That can happen within a few tens of picoseconds.

00:30:53.920 | So you can get devices that operate

00:30:56.840 | in the hundreds of gigahertz range.

00:30:58.800 | And by comparison, most processors

00:31:02.040 | in our conventional computers operate closer

00:31:04.920 | to the one gigahertz range, maybe three gigahertz

00:31:08.320 | seems to be kind of where those speeds have leveled out.

00:31:12.560 | - So the gamers listening to this

00:31:14.200 | are getting really excited

00:31:15.560 | that overclocked their system to like,

00:31:17.680 | what is it like four gigahertz or something?

00:31:19.560 | 100 sounds incredible.

00:31:21.960 | Can I just, as a tiny tangent,

00:31:24.040 | is the physics of this understood well

00:31:26.880 | how to do this stably?

00:31:28.520 | - Oh yes, the physics is understood well.

00:31:30.160 | The physics of Joseph's injunctions is understood well.

00:31:32.560 | The technology is understood quite well too.

00:31:34.520 | The reasons why it hasn't displaced

00:31:37.640 | silicon microelectronics in conventional digital computing,

00:31:41.720 | I think are more related to what I was alluding to before

00:31:45.040 | about the myriad practical, almost mundane aspects

00:31:49.240 | of silicon that make it so useful.

00:31:52.080 | You can make a transistor ever smaller and smaller

00:31:55.880 | and it will still perform its digital function quite well.

00:31:58.840 | The same is not true of a Joseph's injunction.

00:32:00.760 | You really, they don't, they just,

00:32:02.400 | it's not the same thing that there's this feature

00:32:04.440 | that you can keep making smaller and smaller

00:32:06.280 | and it'll keep performing the same operations.

00:32:08.200 | This loop I described, any Joseph's in circuit,

00:32:11.480 | well, I want to be careful,

00:32:13.120 | I shouldn't say any Joseph's in circuit,

00:32:14.680 | but many Joseph's in circuits,

00:32:17.240 | the way they process information

00:32:19.440 | or the way they perform whatever function it is

00:32:21.280 | they're trying to do,

00:32:22.120 | maybe it's sensing a weak magnetic field,

00:32:24.560 | it depends on an interplay between the junction

00:32:27.480 | and that loop.

00:32:28.800 | And you can't make that loop much smaller.

00:32:30.560 | And it's not for practical reasons

00:32:32.120 | that have to do with lithography.

00:32:33.480 | It's for fundamental physical reasons

00:32:35.680 | about the way the magnetic field interacts

00:32:38.960 | with that superconducting material.

00:32:41.160 | There are physical limits that no matter

00:32:43.800 | how good our technology got,

00:32:45.920 | those circuits would, I think,

00:32:47.840 | would never be able to be scaled down

00:32:50.360 | to the densities that silicon microelectronics can.

00:32:54.380 | - I don't know if we mentioned,

00:32:55.560 | is there something interesting

00:32:56.960 | about the various superconducting materials involved?

00:33:00.200 | Or is it all--

00:33:01.040 | - There's a lot of stuff that's interesting.

00:33:02.640 | - And it's not silicon.

00:33:04.440 | - It's not silicon, no.

00:33:05.840 | - So like it's some materials

00:33:08.400 | that also require to be super cold for Kelvin and so on.

00:33:12.040 | - Yes, so let's dissect a couple of those different things.

00:33:15.320 | The super cold part, let me just mention,

00:33:18.520 | for your gamers out there that are trying

00:33:20.120 | to clock it at four gigahertz

00:33:21.360 | and would love to go to 400.

00:33:22.200 | - Yeah, what kind of cooling system can achieve four Kelvin?

00:33:24.160 | - For Kelvin, you need liquid helium.

00:33:26.320 | And so liquid helium is expensive, it's inconvenient.

00:33:29.920 | You need a cryostat that sits there

00:33:32.120 | and the energy consumption of that cryostat

00:33:36.560 | is impracticable for, it's not going in your cell phone.

00:33:40.080 | So you can picture holding your cell phone like this

00:33:42.080 | and then something the size of a keg of beer

00:33:46.160 | or something on your back to cool it,

00:33:47.840 | like that makes no sense.

00:33:49.560 | So if you're trying to make this in consumer devices,

00:33:54.140 | electronics that are ubiquitous across society,

00:33:57.040 | superconductors are not in the race for that.

00:33:59.280 | - For now, but you're saying,

00:34:01.040 | so just to frame the conversation,

00:34:03.280 | maybe the thing we're focused on is computing systems

00:34:06.560 | that serve as servers, like large systems.

00:34:10.360 | - Yes, large systems.

00:34:11.400 | So then you can contrast what's going on in your cell phone

00:34:14.800 | with what's going on at one of the supercomputers.

00:34:18.240 | Colleague Katie Schumann invited us out to Oak Ridge

00:34:22.080 | a few years ago, so we got to see Titan

00:34:24.080 | and that was when they were building Summit.

00:34:26.120 | So these are some high performance supercomputers

00:34:29.480 | out in Tennessee and those are filling entire rooms

00:34:32.480 | the size of warehouses.

00:34:33.920 | So once you're at that level,

00:34:35.920 | okay, there you're already putting a lot of power

00:34:37.760 | into cooling, cooling is part of your engineering task

00:34:42.280 | that you have to deal with.

00:34:43.600 | So there it's not entirely obvious

00:34:45.600 | that cooling to four Kelvin is out of the question.

00:34:49.480 | It has not happened yet and I can speak to why that is

00:34:53.240 | in the digital domain if you're interested.

00:34:55.520 | I think it's not going to happen.

00:34:57.640 | I don't think superconductors are gonna replace

00:35:01.240 | semiconductors for digital computation.

00:35:05.840 | There are a lot of reasons for that,

00:35:07.600 | but I think ultimately what it comes down to

00:35:09.740 | is all things considered, cooling, errors,

00:35:13.400 | scaling down to feature sizes, all that stuff,

00:35:16.040 | semiconductors work better at the system level.

00:35:19.360 | - Is there some aspect of just curious

00:35:22.680 | about the historical momentum of this?

00:35:25.500 | Is there some power to the momentum of an industry

00:35:28.160 | that's mass manufacturing using a certain material?

00:35:31.160 | Is this like a Titanic shifting?

00:35:33.640 | Like, what's your sense?

00:35:34.840 | When a good idea comes along,

00:35:37.080 | how good does that idea need to be

00:35:39.880 | for the Titanic to start shifting?

00:35:42.600 | - That's an excellent question.

00:35:44.160 | That's an excellent way to frame it.

00:35:46.500 | And you know, I don't know the answer to that,

00:35:51.280 | but what I think is, okay, so the history

00:35:54.660 | of the superconducting logic goes back to the 70s.

00:35:57.960 | IBM made a big push to do superconducting

00:36:00.520 | digital computing in the 70s.

00:36:02.360 | And they made some choices about their devices

00:36:06.080 | and their architectures and things that in hindsight

00:36:09.400 | were kind of doomed to fail.

00:36:10.980 | And I don't mean any disrespect for the people that did it.

00:36:13.120 | It was hard to see at the time.

00:36:14.260 | But then another generation of superconducting logic

00:36:17.860 | was introduced, I wanna say the 90s.

00:36:22.280 | Someone named Likorev and Seminov,

00:36:24.960 | they proposed an entire family of circuits

00:36:28.240 | based on Josephson junctions

00:36:29.920 | that are doing digital computing based on logic gates

00:36:33.420 | and or not these kinds of things.

00:36:36.560 | And they showed how it could go hundreds of times faster

00:36:41.520 | than silicon microelectronics.

00:36:43.160 | And it's extremely exciting.

00:36:45.320 | I wasn't working in the field at that time,

00:36:47.000 | but later when I went back and read the literature,

00:36:49.520 | I was just like, wow, this is so awesome.

00:36:51.980 | And so you might think, well,

00:36:55.940 | the reason why it didn't display silicon

00:36:58.240 | is because silicon already had so much momentum

00:37:00.360 | at that time.

00:37:01.680 | But that was the 90s.

00:37:02.920 | Silicon kept that momentum

00:37:04.280 | because it had the simple way to keep getting better.

00:37:06.920 | You just make features smaller and smaller.

00:37:08.680 | So it would have to be,

00:37:11.760 | I don't think it would have to be that much better

00:37:13.520 | than silicon to displace it.

00:37:15.400 | But the problem is it's just not better than silicon.

00:37:17.760 | It might be better than silicon in one metric,

00:37:19.920 | speed of a switching operation

00:37:21.400 | or power consumption of a switching operation.

00:37:24.360 | But building a digital computer

00:37:26.000 | is a lot more than just that elemental operation.

00:37:28.680 | It's everything that goes into it,

00:37:31.000 | including the manufacturing, including the packaging,

00:37:33.280 | including the various materials aspects of things.

00:37:38.280 | So the reason why,

00:37:40.560 | and even in some of those early papers,

00:37:42.760 | I can't remember which one it was,

00:37:44.080 | Likorev said something along the lines of,

00:37:47.440 | you can see how we could build an entire family

00:37:49.860 | of digital electronic circuits based on these components.

00:37:52.760 | They could go a hundred or more times faster

00:37:54.960 | than semiconductor logic gates.

00:37:59.280 | But I don't think that's the right way

00:38:00.860 | to use superconducting electronic circuits.

00:38:02.640 | He didn't say what the right way was,

00:38:04.280 | but he basically said digital logic

00:38:07.400 | trying to steal the show from silicon

00:38:11.240 | is probably not what these circuits

00:38:13.400 | are most suited to accomplish.

00:38:16.320 | - So if we can just linger on,

00:38:18.120 | you used the word computation.

00:38:20.800 | When you talk about computation, how do you think about it?

00:38:24.060 | Do you think purely on just the switching,

00:38:28.900 | or do you think something a little bit larger scale,

00:38:31.320 | a circuit taken together,

00:38:32.720 | performing the basic arithmetic operations

00:38:36.940 | that are then required to do the kind of computation

00:38:40.400 | that makes up a computer?

00:38:42.160 | 'Cause when we talk about the speed of computation,

00:38:44.400 | is it boiled down to the basic switching,

00:38:46.960 | or is there some bigger picture that you're thinking about?

00:38:49.160 | - Well, all right, so maybe we should disambiguate.

00:38:52.320 | There are a variety of different kinds of computation.

00:38:55.580 | I don't pretend to be an expert

00:38:57.140 | in the theory of computation or anything like that.

00:39:00.140 | I guess it's important to differentiate though

00:39:02.620 | between digital logic,

00:39:05.760 | which represents information as a series of bits,

00:39:09.740 | binary digits, which you can think of them

00:39:12.980 | as zeros and ones or whatever.

00:39:14.120 | Usually they correspond to a physical system

00:39:17.360 | that has two very well separated states.

00:39:21.180 | And then other kinds of computation,

00:39:22.820 | like we'll get into more the way your brain works,

00:39:25.240 | which it is, I think,

00:39:27.560 | indisputably processing information,

00:39:30.380 | but where the computation begins and ends

00:39:34.060 | is not anywhere near as well defined.

00:39:36.320 | It doesn't depend on these two levels.

00:39:39.640 | Here's a zero, here's a one.

00:39:41.280 | There's a lot of gray area

00:39:42.600 | that's usually referred to as analog computing.

00:39:45.600 | Also in conventional digital computers,

00:39:49.840 | or digital computers in general,

00:39:54.200 | you have a concept of what's called arithmetic depth,

00:39:57.280 | which is jargon that basically means

00:39:59.780 | how many sequential operations are performed

00:40:03.820 | to turn an input into an output.

00:40:07.640 | And those kinds of computations in digital systems

00:40:10.880 | are highly serial, meaning that data streams,

00:40:14.460 | they don't branch off too far to the side.

00:40:16.480 | You do, you have to pull some information over there

00:40:18.880 | and access memory from here and stuff like that.

00:40:20.880 | But by and large, the computation proceeds

00:40:24.320 | in a serial manner.

00:40:26.200 | It's not that way in the brain.

00:40:27.720 | In the brain, you're always drawing information

00:40:30.720 | from different places.

00:40:31.560 | It's much more network-based computing.

00:40:33.760 | Neurons don't wait for their turn.

00:40:35.640 | They fire when they're ready to fire.

00:40:37.160 | And so it's asynchronous.

00:40:39.160 | So one of the other things about a digital system

00:40:41.640 | is you're performing these operations on a clock.

00:40:44.480 | And that's a crucial aspect of it.

00:40:46.680 | Get rid of a clock in a digital system,

00:40:48.880 | nothing makes sense anymore.

00:40:50.440 | The brain has no clock.

00:40:51.560 | It builds its own time scales based on its internal activity.

00:40:55.300 | - So you can think of the brain as kind of,

00:40:59.120 | I like this, like network computation

00:41:00.920 | where it's actually really trivial, simple computers,

00:41:04.600 | just a huge number of them, and they're networked.

00:41:09.000 | - I would say it is complex, sophisticated,

00:41:11.980 | little processors, and there's a huge number of them.

00:41:14.400 | Neurons are not-- - No offense,

00:41:16.200 | I don't mean to offend neurons.

00:41:17.600 | They're very complicated and beautiful.

00:41:19.800 | But we often oversimplify them.

00:41:21.960 | Yes, there's computation happening within a neuron.

00:41:25.680 | - Right, so I would say to think of a transistor

00:41:29.520 | as the building block of a digital computer is accurate.

00:41:32.360 | You use a few transistors to make your logic gates.

00:41:34.680 | You build up processors from logic gates

00:41:37.840 | and things like that.

00:41:39.160 | So you can think of a transistor

00:41:40.600 | as a fundamental building block,

00:41:42.300 | or you can think of, as we get into more

00:41:44.600 | highly parallelized architectures,

00:41:46.360 | you can think of a processor

00:41:47.680 | as a fundamental building block.

00:41:49.320 | To make the analogy to the neuro side of things,

00:41:53.160 | a neuron is not a transistor.

00:41:55.320 | A neuron is a processor.

00:41:57.320 | It has synapses.

00:41:58.560 | Even synapses are not transistors,

00:42:00.220 | but they are lower on the information processing hierarchy

00:42:04.800 | in a sense.

00:42:05.640 | They do a bulk of the computation,

00:42:08.200 | but neurons are entire processors in and of themselves

00:42:13.600 | that can take in many different kinds of inputs

00:42:16.280 | on many different spatial and temporal scales

00:42:18.840 | and produce many different kinds of outputs

00:42:20.840 | so that they can perform different computations

00:42:23.840 | in different contexts.

00:42:24.880 | - So this is where it enters this distinction

00:42:27.440 | between computation and communication.

00:42:30.760 | So you can think of neuron as performing computation

00:42:34.140 | and the networking, the interconnectivity of neurons

00:42:39.000 | is communication between neurons.

00:42:40.920 | And you see this with very large server systems

00:42:43.480 | I mentioned offline, I've been talking to Jim Keller

00:42:46.160 | whose dream is to build giant computers that,

00:42:48.580 | you know, the bottom like there's often the communication

00:42:52.400 | between the different pieces of computing.

00:42:54.700 | So in this paper that we mentioned,

00:42:57.400 | Optoelectronic Intelligence,

00:42:59.680 | you say electrons excel at computation

00:43:03.200 | while light is excellent for communication.

00:43:08.200 | Maybe you can linger and say in this context,

00:43:11.040 | what do you mean by computation and communication?

00:43:13.980 | What are electrons?

00:43:16.080 | What is light?

00:43:17.440 | And why do they excel at those two tasks?

00:43:20.660 | - Yeah, just to first speak to computation

00:43:23.620 | versus communication, I would say computation

00:43:27.700 | is essentially taking in some information,

00:43:31.620 | performing operations on that information

00:43:34.680 | and producing new, hopefully more useful information.

00:43:39.040 | So for example, imagine you have a picture in front of you

00:43:44.040 | and there is a key in it and that's what you're looking for

00:43:48.960 | for whatever reason, you wanna find the key,

00:43:50.680 | we all wanna find the key.

00:43:51.560 | So the input is that entire picture

00:43:56.540 | and the output might be the coordinates where the key is.

00:43:59.040 | So you've reduced the total amount of information you have,

00:44:01.520 | but you found the useful information

00:44:03.000 | for you in that present moment,

00:44:04.380 | that's the useful information.

00:44:05.220 | - And you think about this computation

00:44:07.160 | as the controlled synchronous sequential--

00:44:10.840 | - Not necessarily, it could be.

00:44:12.680 | That could be how your system is performing the computation

00:44:16.240 | or it could be asynchronous.

00:44:19.320 | There are lots of ways to find the key.

00:44:21.420 | It depends on the nature of the data,

00:44:23.720 | it depends on, that's a very simplified example,

00:44:27.560 | a picture with a key in it.

00:44:28.720 | What about if you're in the world

00:44:30.540 | and you're trying to decide the best way

00:44:32.520 | to live your life, you know?

00:44:35.960 | It might be interactive, there might be some recurrence

00:44:38.600 | or some weird asynchrony, I got it.

00:44:41.360 | But there's an input and there's an output

00:44:43.280 | and you do some stuff in the middle

00:44:44.440 | that actually goes from the input to the output.

00:44:46.040 | - You've taken in information

00:44:47.360 | and output different information,

00:44:49.080 | hopefully reducing the total amount of information

00:44:51.800 | and extracting what's useful.

00:44:53.800 | Communication is then getting that information

00:44:57.760 | from the location in which it's stored

00:44:59.440 | because information is physical as Landauer emphasized

00:45:02.680 | and so it is in one place

00:45:05.120 | and you need to get that information to another place

00:45:07.840 | so that something else can use it

00:45:10.120 | for whatever computation it's working on.

00:45:12.020 | Maybe it's part of the same network

00:45:13.440 | and you're all trying to solve the same problem

00:45:15.000 | but neuron A over here just deduced something

00:45:20.000 | based on its inputs and it's now sending that information

00:45:24.160 | across the network to another location

00:45:26.480 | so that would be the act of communication.

00:45:28.440 | - Can you linger on Landauer

00:45:29.880 | and saying information is physical?

00:45:31.840 | - Rolf Landauer, not to be confused with Lev Landau.

00:45:35.360 | Yeah, and he made huge contributions

00:45:38.000 | to our understanding of the reversibility of information

00:45:43.000 | and this concept that energy has to be dissipated

00:45:46.680 | in computing when the computation is irreversible

00:45:50.120 | but if you can manage to make it reversible

00:45:52.120 | then you don't need to expend energy

00:45:55.080 | but if you do expend energy to perform a computation

00:45:59.680 | there's sort of a minimal amount that you have to do

00:46:02.360 | and it's KT log two.

00:46:04.480 | - And it's all somehow related

00:46:05.920 | to the second law of thermodynamics

00:46:07.680 | and that the universe is an information process

00:46:09.640 | and then we're living in a simulation.

00:46:11.560 | So, okay, sorry.

00:46:12.960 | Sorry for that tangent.

00:46:14.680 | So that's the defining the distinction

00:46:17.160 | between computation and communication.

00:46:19.600 | - Let me say one more thing just to clarify.

00:46:21.960 | Communication ideally does not change the information.

00:46:26.960 | It moves it from one place to another

00:46:28.920 | but it is preserved.

00:46:30.960 | - Got it, okay.

00:46:32.560 | All right, that is beautiful.

00:46:33.760 | So, then the electron versus light distinction

00:46:38.640 | and why are electrons good at computation

00:46:42.440 | and light good at communication?

00:46:44.560 | - Yes, there's a lot that goes into it I guess

00:46:48.860 | but just try to speak to the simplest part of it.

00:46:51.720 | Electrons interact strongly with one another.

00:46:57.040 | They're charged particles.

00:46:58.360 | So if I pile a bunch of them over here,

00:47:02.040 | they're feeling a certain amount of force

00:47:03.880 | and they wanna move somewhere else.

00:47:05.680 | They're strongly interactive.

00:47:06.920 | You can also get them to sit still.

00:47:09.560 | An electron has a mass

00:47:10.680 | so you can cause it to be spatially localized.

00:47:15.680 | So for computation, that's useful

00:47:18.120 | because now I can make these little devices

00:47:20.160 | that put a bunch of electrons over here

00:47:21.960 | and then I change the state of a gate

00:47:25.640 | like I've been describing,

00:47:26.520 | put a different voltage on this gate

00:47:28.400 | and now I move the electrons over here.

00:47:30.000 | Now they're sitting somewhere else.

00:47:31.240 | I have a physical mechanism

00:47:34.000 | with which I can represent information.

00:47:36.040 | It's spatially localized and I have knobs

00:47:38.140 | that I can adjust to change where those electrons are

00:47:41.240 | or what they're doing.

00:47:42.400 | Light by contrast, photons of light,

00:47:45.280 | which are the discrete packets of energy

00:47:48.120 | that were identified by Einstein,

00:47:50.800 | they do not interact with each other

00:47:53.440 | especially at low light levels.

00:47:56.280 | If you're in a medium

00:47:57.400 | and you have a bright, high light level,

00:48:00.400 | you can get them to interact with each other

00:48:02.560 | through the interaction with that medium that they're in

00:48:05.360 | but that's a little bit more exotic

00:48:07.840 | and for the purposes of this conversation,

00:48:10.360 | we can assume that photons don't interact with each other.

00:48:13.240 | So if you have a bunch of them

00:48:16.120 | all propagating in the same direction,

00:48:17.620 | they don't interfere with each other.

00:48:19.160 | If I wanna send, if I have a communication channel

00:48:22.960 | and I put one more photon on it,

00:48:24.500 | it doesn't screw up with those other ones.

00:48:26.040 | It doesn't change what those other ones were doing at all.

00:48:29.120 | So that's really useful for communication

00:48:31.280 | because that means you can sort of

00:48:33.320 | allow a lot of these photons to flow

00:48:35.600 | without disruption of each other

00:48:38.840 | and they can branch really easily and things like that

00:48:41.120 | but it's not good for computation

00:48:42.720 | because it's very hard for this packet of light

00:48:46.360 | to change what this packet of light is doing.

00:48:48.720 | They pass right through each other.

00:48:50.200 | So in computation, you want to change information

00:48:53.300 | and if photons don't interact with each other,

00:48:55.720 | it's difficult to get them to change

00:48:57.360 | the information represented by the others.

00:48:59.400 | - So that's the fundamental difference.

00:49:01.440 | Is there also something about the way they travel

00:49:04.800 | through different materials

00:49:06.440 | or is that just a particular engineering?

00:49:10.720 | - No, it's not.

00:49:11.560 | That's deep physics, I think.

00:49:12.600 | So this gets back to electrons interact with each other

00:49:17.080 | and photons don't.

00:49:18.160 | So say I'm trying to get a packet of information

00:49:22.440 | from me to you and we have a wire going between us.

00:49:25.840 | In order for me to send electrons across that wire,

00:49:29.040 | I first have to raise the voltage on my end of the wire

00:49:32.240 | and that means putting a bunch of charges on it

00:49:34.640 | and then that charge packet has to propagate along the wire

00:49:39.180 | and it has to get all the way over to you.

00:49:41.280 | That wire is gonna have something that's called capacitance

00:49:44.380 | which basically tells you how much charge

00:49:46.960 | you need to put on the wire

00:49:48.080 | in order to raise the voltage on it

00:49:49.960 | and the capacitance is gonna be proportional

00:49:52.520 | to the length of the wire.

00:49:54.040 | So the longer the length of the wire is,

00:49:56.880 | the more charge I have to put on it

00:49:59.120 | and the energy required to charge up that line

00:50:03.080 | and move those electrons to you

00:50:04.960 | is also proportional to the capacitance

00:50:06.840 | and goes as the voltage squared.

00:50:08.520 | So you get this huge penalty

00:50:10.560 | if you wanna send electrons across a wire

00:50:15.040 | over appreciable distances.

00:50:16.600 | - So distance is an important thing here

00:50:19.120 | when you're doing communication.

00:50:20.800 | - Distance is an important thing,

00:50:22.120 | so is the number of connections I'm trying to make.

00:50:25.360 | Me to you, okay, one, that's not so bad.

00:50:27.640 | If I want to now send it to 10,000 other friends,

00:50:31.440 | then all of those wires

00:50:33.360 | are adding tons of extra capacitance.

00:50:35.480 | Now, not only does it take forever

00:50:37.680 | to put the charge on that wire

00:50:39.560 | and raise the voltage on all those lines,

00:50:41.840 | but it takes a ton of power

00:50:43.560 | and the number 10,000 is not randomly chosen,

00:50:47.000 | that's roughly how many connections

00:50:49.120 | each neuron in your brain makes.

00:50:50.640 | So a neuron in your brain needs to send 10,000 messages

00:50:55.040 | every time it has something to say.

00:50:56.780 | You can't do that if you're trying to drive electrons

00:51:00.120 | from here to 10,000 different places.

00:51:02.080 | The brain does it in a slightly different way,

00:51:03.680 | which we can discuss.

00:51:04.880 | - How can light achieve the 10,000 connections

00:51:07.040 | and why is it better in terms of the energy use

00:51:12.000 | required to use light for the communication

00:51:14.240 | of the 10,000 connections?

00:51:15.280 | - Right, right, so now instead of trying

00:51:16.800 | to send electrons from me to you,

00:51:18.200 | I'm trying to send photons.

00:51:19.360 | So I can make what's called a waveguide,

00:51:21.520 | which is just a simple piece of material.

00:51:25.120 | It could be glass, like an optical fiber

00:51:27.060 | or silicon on a chip.

00:51:29.880 | And I just have to inject photons into that waveguide

00:51:34.160 | and independent of how long it is,

00:51:35.800 | independent of how many different connections I'm making,

00:51:39.640 | it doesn't change the voltage or anything like that

00:51:43.120 | that I have to raise up on the wire.

00:51:45.480 | So if I have one more connection,

00:51:47.920 | if I add additional connections,

00:51:49.820 | I need to add more light to the waveguide

00:51:51.760 | because those photons need to split

00:51:53.280 | and go to different paths.

00:51:55.040 | That makes sense, but I don't have a capacitive penalty.

00:51:58.840 | Sometimes these are called wiring parasitics.

00:52:01.280 | There are no parasitics associated with light

00:52:03.440 | in that same sense.

00:52:05.120 | - Well, this might be a dumb question,

00:52:07.480 | but how do I catch a photon on the other end?

00:52:11.360 | Is it material, is it the polymer stuff

00:52:14.080 | you were talking about for a different application

00:52:17.240 | for photolithography?

00:52:19.520 | How do you catch a photon?

00:52:21.000 | - There's a lot of ways to catch a photon.

00:52:22.520 | It's not a dumb question.

00:52:23.600 | It's a deep and important question

00:52:25.940 | that basically defines a lot of the work

00:52:29.180 | that goes on in our group at NIST.

00:52:31.360 | One of my group leaders, Seywounam,

00:52:34.280 | has built his career around these superconducting

00:52:37.100 | single photon detectors.

00:52:38.400 | So if you're going to try to sort of reach a lower limit

00:52:42.440 | and detect just one particle of light,

00:52:45.240 | superconductors come back into our conversation.

00:52:47.660 | And just picture a simple device

00:52:50.140 | where you have current flowing

00:52:51.600 | through a superconducting wire and--

00:52:54.600 | - A loop again or no?

00:52:56.560 | - Let's say, yes, you have a loop.

00:52:57.800 | So you have a superconducting wire

00:52:59.560 | that goes straight down like this.

00:53:01.000 | And on your loop branch, you have a little ammeter,

00:53:04.240 | something that measures current.

00:53:05.800 | There's a resistor up there too.

00:53:07.960 | Go with me here.

00:53:09.000 | So you're current biasing this,

00:53:12.000 | so there's current flowing

00:53:13.060 | through that superconducting branch.

00:53:14.400 | Since there's a resistor over here,

00:53:16.560 | all the current goes through the superconducting branch.

00:53:18.920 | Now a photon comes in, strikes that superconductor.

00:53:22.280 | We talked about this superconducting

00:53:24.280 | macroscopic quantum state.

00:53:25.760 | That's going to be destroyed by the energy of that photon.

00:53:28.480 | So now that branch of the circuit is resistive too.

00:53:32.080 | And you've properly designed your circuit

00:53:33.760 | so that the resistance on that superconducting branch

00:53:36.680 | is much greater than the other resistance.

00:53:38.400 | Now, all of your current's going to go that way.

00:53:40.760 | Your ammeter says, oh, I just got a pulse of current.

00:53:43.240 | That must mean I detected a photon.

00:53:45.120 | Then where you broke that superconductivity

00:53:47.200 | in a matter of a few nanoseconds,

00:53:49.080 | it cools back off, dissipates that energy,

00:53:51.080 | and the current flows back

00:53:52.800 | through that superconducting branch.

00:53:54.280 | This is a very powerful superconducting device

00:53:59.200 | that allows us to understand quantum states of light.

00:54:02.280 | - I didn't realize a loop like that

00:54:04.880 | could be sensitive to a single photon.

00:54:07.160 | I mean, that seems strange to me because,

00:54:12.000 | I mean, so what happens when you just barrage it with photons?

00:54:16.600 | - If you put a bunch of photons in there,

00:54:18.360 | essentially the same thing happens.

00:54:19.880 | You just drive it into the normal state.

00:54:21.680 | It becomes resistive,

00:54:23.160 | and it's not particularly interesting.

00:54:25.400 | - So you have to be careful how many photons you send.

00:54:28.000 | Like you have to be very precise with your communication.

00:54:30.120 | - Well, it depends.

00:54:31.240 | So I would say that that's actually,

00:54:32.920 | in the application that we're trying

00:54:34.840 | to use these detectors for, that's a feature.

00:54:37.960 | Because what we want is for,

00:54:39.880 | if a neuron sends one photon to a synaptic connection,

00:54:46.120 | and one of these superconducting detectors is sitting there,

00:54:49.840 | you get this pulse of current,

00:54:51.200 | and that synapse says, "Event,"

00:54:54.000 | then I'm gonna do what I do when there's a synapse event,

00:54:56.040 | I'm gonna perform computations, that kind of thing.

00:54:58.640 | But if accidentally you send two there, or three, or five,

00:55:02.120 | it does the exact same.

00:55:03.360 | - Got it.

00:55:04.200 | - And so this is how, in the system that we're devising here,

00:55:09.200 | communication is entirely binary.

00:55:12.760 | And that's what I tried to emphasize a second ago.

00:55:15.040 | Communication should not change the information.

00:55:17.800 | You're not saying, "Oh, I got this kind

00:55:20.600 | "of communication event for photons."

00:55:22.280 | No, we're not keeping track of that.

00:55:23.720 | This neuron fired, this synapse says,

00:55:25.720 | "That neuron fired, that's it."

00:55:27.040 | So that's a noise filtering property of those detectors.

00:55:31.480 | However, there are other applications

00:55:33.120 | where you'd rather know the exact number of photons.

00:55:36.120 | That can be very useful in quantum computing with light.

00:55:39.320 | And our group does a lot of work

00:55:41.800 | around another kind of superconducting sensor

00:55:44.560 | called a transition edge sensor

00:55:46.640 | that Adrian Alita in our group does a lot of work on that.

00:55:49.920 | And that can tell you,

00:55:52.560 | based on the amplitude of the current pulse you divert,

00:55:55.760 | exactly how many photons were in that pulse.

00:55:59.760 | - What's that useful for?

00:56:02.480 | - One way that you can encode information

00:56:04.700 | in quantum states of light is in the number of photons.

00:56:07.440 | You can have what are called number states.

00:56:09.320 | And a number state will have a well-defined number

00:56:11.960 | of photons, and maybe the output

00:56:14.400 | of your quantum computation encodes its information

00:56:18.760 | in the number of photons that are generated.

00:56:20.460 | So if you have a detector that is sensitive to that,

00:56:23.040 | it's extremely useful.

00:56:24.280 | - Can you achieve like a clock with photons?

00:56:29.240 | Or is that not important?

00:56:30.240 | Is there a synchronicity here?

00:56:32.440 | - In general, it can be important.

00:56:35.360 | Clock distribution is a big challenge

00:56:39.380 | in especially large computational systems.

00:56:43.280 | And so yes, optical clocks, optical clock distribution

00:56:47.960 | is a very powerful technology.

00:56:51.140 | I don't know the state of that field right now,

00:56:53.180 | but I imagine that if you're trying to distribute a clock

00:56:55.620 | across any appreciable size computational system,

00:56:58.960 | you wanna use light.

00:57:00.360 | - Yeah, I wonder how these giant systems work,

00:57:04.280 | especially like supercomputers.

00:57:07.360 | Do they need to do clock distribution

00:57:09.360 | or are they doing more ad hoc parallel,

00:57:14.240 | like concurrent programming?

00:57:15.520 | Like there's some kind of locking mechanisms or something.

00:57:18.120 | - That's a fascinating question.

00:57:20.060 | Let's zoom in at this very particular question

00:57:23.860 | of computation on a processor

00:57:28.180 | and communication between processors.

00:57:31.520 | So what does this system look like

00:57:36.420 | that you're envisioning?

00:57:38.180 | One of the places you're envisioning it

00:57:40.060 | is in the paper on optoelectronic intelligence.

00:57:43.100 | So what are we talking about?

00:57:44.700 | Are we talking about something that starts to look a lot

00:57:47.300 | like the human brain or does it still look a lot

00:57:50.160 | like a computer?

00:57:51.260 | What are the size of this thing?

00:57:52.920 | Is it going inside a smartphone?

00:57:54.380 | Or as you said, does it go inside something

00:57:56.360 | that's more like a house?

00:57:58.540 | Like what should we be imagining?

00:58:01.140 | What are you thinking about when you're thinking

00:58:02.660 | about these fundamental systems?

00:58:05.420 | - Let me introduce the word neuromorphic.

00:58:07.340 | There's this concept of neuromorphic computing

00:58:09.900 | where what that broadly refers to is computing

00:58:15.440 | based on the information processing principles of the brain.

00:58:19.220 | And as digital computing seems to be pushing

00:58:23.780 | towards some fundamental performance limits,

00:58:26.440 | people are considering architectural advances,

00:58:29.140 | drawing inspiration from the brain,

00:58:30.840 | more distributed parallel network architectures and stuff.

00:58:34.200 | And so there's this continuum of neuromorphic

00:58:37.420 | from things that are pretty similar to digital computers,

00:58:42.740 | but maybe there are more cores

00:58:45.720 | and the way they send messages is a little bit more

00:58:49.160 | like the way brain neurons send spikes.

00:58:52.780 | But for the most part, it's still digital electronics.

00:58:56.080 | And then you have some things in between

00:58:58.700 | where maybe you're using transistors,

00:59:02.060 | but now you're starting to use them

00:59:03.220 | instead of in a digital way, in an analog way.

00:59:06.240 | And so you're trying to get those circuits

00:59:08.140 | to behave more like neurons.

00:59:10.300 | And then that's a little bit,

00:59:12.100 | quite a bit more on the neuromorphic side of things.

00:59:14.920 | You're trying to get your circuits,

00:59:17.060 | although they're still based on silicon,

00:59:19.120 | you're trying to get them to perform operations

00:59:22.460 | that are highly analogous to the operations in the brain.

00:59:24.660 | That's where a great deal of work is

00:59:26.680 | in neuromorphic computing, people like Yakima Wendoveri

00:59:29.160 | and Gert Kauenberg, Jennifer Hasler, countless others.

00:59:32.940 | It's a rich and exciting field

00:59:35.540 | going back to Carver Mead in the late 1980s.

00:59:39.440 | And then all the way on the other extreme of the continuum

00:59:44.220 | is where you say, I'll give up anything related

00:59:48.540 | to transistors or semiconductors or anything like that.

00:59:51.380 | I'm not starting with the assumption

00:59:53.580 | that I'm gonna use any kind

00:59:55.020 | of conventional computing hardware.

00:59:57.220 | And instead, what I wanna do is try and understand

00:59:59.260 | what makes the brain powerful

01:00:00.660 | at the kind of information processing it does.

01:00:03.300 | And I wanna think from first principles

01:00:05.500 | about what hardware is best going to enable us

01:00:10.500 | to capture those information processing principles

01:00:13.980 | in an artificial system.

01:00:15.980 | And that's where I live.

01:00:17.380 | That's where I'm doing my exploration these days.

01:00:21.700 | - So what are the first principles

01:00:25.780 | of brain-like computation and communication?

01:00:29.920 | - Right, yeah, this is so important.

01:00:32.560 | And I'm glad we booked 14 hours for this

01:00:34.460 | because-- - I only have 13, I'm sorry.

01:00:37.240 | (laughing)

01:00:38.360 | - Okay, so the brain is notoriously complicated.

01:00:41.480 | And I think that's an important part

01:00:44.040 | of why it can do what it does.

01:00:46.300 | But okay, let me try to break it down.

01:00:48.680 | Starting with the devices, neurons, as I said before,

01:00:54.620 | they're sophisticated devices in and of themselves.

01:00:57.080 | And synapses are too.

01:00:58.220 | They can change their state based on the activity.

01:01:03.040 | So they adapt over time.

01:01:04.880 | That's crucial to the way the brain works.

01:01:07.000 | They don't just adapt on one time scale.

01:01:09.360 | They can adapt on myriad timescales

01:01:12.440 | from the spacing between pulses,

01:01:16.080 | the spacing between spikes that come from neurons,

01:01:18.720 | all the way to the age of the organism.

01:01:21.540 | Also relevant, perhaps, I think the most important thing

01:01:28.100 | that's guided my thinking is the network structure

01:01:32.320 | of the brain.

01:01:33.880 | - Which can also be adjusted on different scales.

01:01:36.560 | - Absolutely, yeah.

01:01:37.400 | So you're changing the strength of contacts,

01:01:41.360 | you're changing the spatial distribution of them,

01:01:44.120 | although spatial distribution doesn't change that much

01:01:46.880 | once you're a mature organism.

01:01:49.400 | But that network structure is really crucial.

01:01:52.880 | So let me dwell on that for a second.

01:01:55.380 | You can't talk about the brain without emphasizing

01:01:58.960 | that most of the neurons in the neocortex

01:02:02.880 | or the prefrontal cortex, the part of the brain

01:02:04.840 | that we think is most responsible for high-level reasoning

01:02:08.080 | and things like that,

01:02:09.080 | those neurons make thousands of connections.

01:02:11.400 | So you have this network that is highly interconnected.

01:02:15.560 | And I think it's safe to say that one of the primary reasons

01:02:19.880 | that they make so many different connections

01:02:23.200 | is that allows information to be communicated very rapidly

01:02:26.880 | from any spot in the network to any other spot

01:02:29.240 | in the network.

01:02:30.320 | So that's a sort of spatial aspect of it.

01:02:33.920 | You can quantify this in terms of concepts

01:02:38.480 | that are related to fractals and scale invariance,

01:02:41.020 | which I think is a very beautiful concept.

01:02:43.260 | So what I mean by that is kind of no matter

01:02:46.720 | what spatial scale you're looking at in the brain

01:02:50.720 | within certain bounds,

01:02:52.320 | you see the same general statistical pattern.

01:02:54.960 | So if I draw a box around some region of my cortex,

01:02:59.280 | most of the connections that those neurons

01:03:02.280 | within that box make are gonna be within the box

01:03:04.480 | to each other in their local neighborhood.

01:03:06.200 | And that's sort of called clustering, loosely speaking.

01:03:09.280 | But a non-negligible fraction

01:03:10.920 | is gonna go outside of that box.

01:03:12.600 | And then if I draw a bigger box,

01:03:14.060 | the pattern is gonna be exactly the same.

01:03:16.400 | So you have this scale invariance

01:03:18.400 | and you also have a non-vanishing probability

01:03:22.680 | of a neuron making connection very far away.

01:03:25.360 | So suppose you wanna plot the probability

01:03:28.680 | of a neuron making a connection as a function of distance.

01:03:32.400 | If that were an exponential function,

01:03:34.200 | it would go e to the minus radius

01:03:36.680 | over some characteristic radius,

01:03:38.660 | and it would drop off up to some certain radius.

01:03:41.820 | The probability would be reasonable close to one.

01:03:44.880 | And then beyond that characteristic length, R zero,

01:03:49.120 | it would drop off sharply.

01:03:51.280 | And so that would mean that the neurons in your brain

01:03:53.560 | are really localized, and that's not what we observe.

01:03:58.560 | Instead, what you see is that the probability

01:04:00.680 | of making a longer distance connection,

01:04:02.700 | it does drop off, but it drops off as a power law.

01:04:05.760 | So the probability that you're gonna have a connection

01:04:08.400 | at some radius R goes as R to the minus some power.

01:04:13.160 | And that's what we see with forces in nature,

01:04:16.800 | like the electromagnetic force

01:04:18.440 | between two particles or gravity

01:04:20.440 | goes as one over the radius squared.

01:04:22.960 | - You can see this in fractals.

01:04:24.400 | I love that there's a fractal dynamics of the brain

01:04:28.600 | that if you zoom out, you draw the box,

01:04:31.440 | and you increase that box by certain step sizes,

01:04:35.020 | you're gonna see the same statistics.

01:04:36.720 | - I think that's probably very important

01:04:39.960 | to the way the brain processes information.

01:04:41.840 | It's not just in the spatial domain,

01:04:43.560 | it's also in the temporal domain.

01:04:45.600 | And what I mean by that is-

01:04:47.080 | (imitates explosion)

01:04:48.600 | - That's incredible that this emerged

01:04:50.600 | through the evolutionary process

01:04:52.320 | that potentially somehow connected

01:04:54.880 | to the way the physics of the universe works.

01:04:57.800 | - Yeah, I couldn't agree more.

01:04:59.320 | That it's a deep and fascinating subject

01:05:01.400 | that I hope to be able to spend the rest of my life studying.

01:05:04.160 | - You think you need to solve, understand this,

01:05:07.280 | this fractal nature in order to understand

01:05:09.480 | intelligence and communication.

01:05:10.720 | - I do think so.

01:05:11.560 | I think they're deeply intertwined.

01:05:13.320 | Yes, I think power laws are right at the heart of it.

01:05:16.880 | So just to push that one through,

01:05:19.400 | the same thing happens in the temporal domain.

01:05:21.480 | So suppose you had,

01:05:23.080 | suppose your neurons in your brain

01:05:26.000 | were always oscillating at the same frequency.

01:05:28.160 | Then the probability of finding a neuron

01:05:30.760 | oscillating as a function of frequency

01:05:32.520 | would be this narrowly peaked function

01:05:34.520 | around that certain characteristic frequency.

01:05:36.520 | That's not at all what we see.

01:05:37.880 | The probability of finding neurons oscillating

01:05:40.240 | or pulsing, producing spikes at a certain frequency

01:05:43.880 | is again a power law,

01:05:45.200 | which means there's no defined scale

01:05:49.640 | of the temporal activity in the brain.

01:05:51.680 | At what speed do your thoughts occur?

01:05:56.040 | Well, there's a fastest speed they can occur

01:05:58.280 | and that is limited by communication and other things.

01:06:01.520 | But there's not a characteristic scale.

01:06:03.960 | We have thoughts on all temporal scales

01:06:06.880 | from a few tens of milliseconds,

01:06:10.840 | which is physiologically limited by our devices.

01:06:13.400 | Compare that to tens of picoseconds

01:06:15.760 | that I talked about in superconductors,

01:06:17.160 | all the way up to the lifetime of the organism.

01:06:19.280 | You can still think about things

01:06:20.760 | that happened to you when you were a kid.

01:06:22.600 | - Or if you wanna be really trippy,

01:06:24.080 | then across multiple organisms

01:06:25.880 | in the entirety of human civilization,

01:06:27.480 | you have thoughts that span organisms, right?

01:06:29.440 | - Yes, taking it to that level, yes.

01:06:31.280 | - If you're willing to see the entirety

01:06:33.680 | of the human species as a single organism

01:06:35.720 | with a collective intelligence,

01:06:37.200 | then that too on a spatial and temporal scale,

01:06:39.920 | there's thoughts occurring.

01:06:41.120 | And then if you look at not just the human species,

01:06:44.040 | but the entirety of life on earth

01:06:46.480 | as an organism with thoughts that are occurring,

01:06:49.640 | that are greater and greater sophisticated thoughts,

01:06:51.640 | there's a different spatial and temporal scale there.

01:06:54.680 | This is getting very suspicious.

01:06:57.240 | - Well, hold on though.

01:06:58.120 | Before we're done, I just wanna just tie the bow

01:07:01.000 | and say that the spatial and temporal aspects

01:07:04.320 | are intimately interrelated with each other.

01:07:06.480 | So activity between neurons

01:07:09.120 | that are very close to each other

01:07:10.360 | is more likely to happen on this faster timescale

01:07:13.600 | and information is gonna propagate

01:07:15.320 | and encompass more of the brain,

01:07:17.200 | more of your cortices, different modules in the brain

01:07:20.320 | are gonna be engaged in information processing

01:07:23.760 | on longer timescales.

01:07:25.320 | So there's this concept of information integration

01:07:27.960 | where most neurons are specialized.

01:07:31.960 | Any given neuron or any cluster of neuron

01:07:34.000 | has its specific purpose,

01:07:35.760 | but they're also very much integrated.

01:07:39.920 | So you have neurons that specialize,

01:07:41.920 | but share their information.

01:07:43.680 | And so that happens through these fractal nested oscillations

01:07:47.600 | that occur across spatial and temporal scales.

01:07:49.440 | I think capturing those dynamics in hardware,

01:07:53.680 | to me, that's the goal of neuromorphic computing.

01:07:57.080 | - So does it need to look,

01:07:58.720 | so first of all, that's fascinating.

01:08:00.840 | We've stated some clear principles here.

01:08:03.960 | Now, does it have to look like the brain

01:08:08.120 | outside of those principles as well?

01:08:09.800 | Like what other characteristics

01:08:11.320 | have to look like the human brain?

01:08:13.080 | Or can it be something very different?

01:08:15.840 | - Well, it depends on what you're trying to use it for.

01:08:18.000 | And so I think a lot of the community

01:08:20.760 | asks that question a lot.

01:08:23.080 | What are you gonna do with it?

01:08:24.360 | And I completely get it.

01:08:26.600 | I think that's a very important question.

01:08:28.040 | And it's also sometimes not the most helpful question.

01:08:31.840 | What if what you wanna do with it is study it?

01:08:33.760 | What if you just wanna see,

01:08:35.120 | what do you have to build into your hardware

01:08:39.240 | in order to observe these dynamical principles?

01:08:41.880 | - And also, I ask myself that question every day,

01:08:47.520 | and I'm not sure I'm able to answer that.

01:08:49.840 | It's like, what are you gonna do

01:08:51.200 | with this particular neuromorphic machine?

01:08:53.480 | - So suppose what we're trying to do with it

01:08:55.320 | is build something that thinks.

01:08:56.960 | We're not trying to get it to make us any money

01:08:59.120 | or drive a car.

01:09:00.200 | Maybe we'll be able to do that, but that's not our goal.

01:09:02.600 | Our goal is to see if we can get the same types

01:09:06.240 | of behaviors that we observe in our own brain.

01:09:08.880 | And by behaviors, in this sense,

01:09:10.440 | what I mean, the behaviors of the components,

01:09:14.280 | the neurons, the network, that kind of stuff.

01:09:15.960 | I think there's another element that I didn't really hit on

01:09:19.080 | that you also have to build into this,

01:09:21.160 | and those are architectural principles.

01:09:22.920 | They have to do with the hierarchical modular construction

01:09:26.640 | of the network.

01:09:27.480 | And without getting too lost in jargon,

01:09:30.200 | the main point that I think is relevant there,

01:09:33.680 | let me try and illustrate it with a cartoon picture

01:09:35.720 | of the architecture of the brain.

01:09:38.200 | So in the brain, you have the cortex,

01:09:41.080 | which is sort of this outer sheet.

01:09:43.520 | It's actually, it's a layered structure.

01:09:47.280 | If you could take it out of your brain,

01:09:48.440 | you could unroll it on the table,

01:09:50.640 | and it would be about the size of a pizza sitting there.

01:09:53.500 | And that's a module.

01:09:57.240 | It does certain things.

01:09:58.640 | It processes, as Yorgi Buzsaki would say,

01:10:01.640 | it processes the what of what's going on around you.

01:10:04.440 | But you have another really crucial module

01:10:07.120 | that's called the hippocampus,

01:10:08.920 | and that network is structured entirely differently.

01:10:11.400 | First of all, this cortex that I described,

01:10:13.740 | 10 billion neurons in there, so numbers matter here.

01:10:17.760 | And they're organized in that sort of power law distribution

01:10:21.320 | where the probability of making a connection

01:10:23.140 | drops off as a power law in space.

01:10:25.240 | The hippocampus is another module

01:10:26.880 | that's important for understanding how,

01:10:30.160 | where you are, when you are,

01:10:34.080 | keeping track of your position in space and time.

01:10:37.960 | And that network is very much random.

01:10:39.960 | So the probability of making a connection,

01:10:42.080 | it almost doesn't even drop off as a function of distance.

01:10:45.500 | It's the same probability

01:10:46.600 | that you'll make it here to over there.

01:10:48.420 | But there are only about 100 million neurons there.

01:10:51.440 | So you can have that huge, densely connected module

01:10:55.120 | because it's not so big.

01:10:57.720 | And the neocortex, or the cortex and the hippocampus,

01:11:02.480 | they talk to each other constantly.

01:11:05.360 | And that communication is largely facilitated

01:11:08.340 | by what's called the thalamus.

01:11:10.160 | I'm not a neuroscientist here.

01:11:11.320 | I'm trying to do my best to recite things.

01:11:13.520 | - Cartoon picture of the brain, I got you.

01:11:15.120 | - Yeah, something like that.

01:11:16.020 | So this thalamus is coordinating the activity

01:11:19.080 | between the neocortex and the hippocampus

01:11:21.160 | and making sure that they talk to each other

01:11:24.000 | at the right time and send messages

01:11:25.720 | that will be useful to one another.

01:11:27.260 | So this all taken together

01:11:29.080 | is called the thalamocortical complex.

01:11:32.040 | And it seems like building something like that

01:11:35.320 | is going to be crucial to capturing

01:11:37.720 | the types of activity we're looking for

01:11:40.540 | because those responsibilities, those separate modules,

01:11:45.280 | they do different things.

01:11:46.160 | That's gotta be central to achieving these states

01:11:51.160 | of efficient information integration across space and time.

01:11:55.260 | - By the way, I am able to achieve this state

01:11:59.400 | by watching simulations, visualizations

01:12:02.280 | of the thalamocortical complex.

01:12:04.240 | There's a few people, I forget from where,

01:12:06.920 | they've created these incredible visual illustrations

01:12:10.360 | of visual stimulation from the eye or something like that

01:12:15.360 | in this image flowing through the brain.

01:12:18.520 | - Wow, I haven't seen that.

01:12:20.040 | I gotta check that out.

01:12:20.880 | - So it's one of those things,

01:12:22.120 | you find this stuff in the world

01:12:24.280 | and you see on YouTube, it has 1,000 views,

01:12:28.120 | these visualizations of the human brain

01:12:30.760 | processing information.

01:12:32.140 | 'Cause there's chemistry there

01:12:36.440 | because this is from actual human brains.

01:12:38.880 | I don't know how they're doing the coloring

01:12:40.720 | but they're able to actually trace

01:12:42.600 | the chemical and the electrical signals

01:12:46.680 | throughout the brain and the visual thing.

01:12:48.880 | It's like, whoa, because it looks kind of like the universe.

01:12:51.800 | I mean, the whole thing is just incredible.

01:12:53.800 | I recommend it highly.

01:12:55.080 | I'll probably post a link to it.

01:12:56.640 | But you can just look for,

01:12:58.120 | one of the things they simulate

01:13:00.960 | is the thalamocortical complex and just visualization.

01:13:05.840 | You can find that yourself on YouTube, but it's beautiful.

01:13:09.520 | The other question I have for you is

01:13:11.320 | how does memory play into all of this?

01:13:14.440 | Because all the signals sending back and forth,

01:13:17.120 | that's computation and communication,

01:13:20.880 | but that's kind of like processing of inputs and outputs

01:13:26.240 | to produce outputs in the system.

01:13:27.560 | That's kind of like maybe reasoning,

01:13:29.000 | maybe there's some kind of recurrence.

01:13:30.920 | But is there a storage mechanism that you think about

01:13:33.920 | in the context of neuromorphic computing?

01:13:35.840 | - Yeah, absolutely.

01:13:36.680 | So that's gotta be central.

01:13:37.760 | You have to have a way that you can store memories.

01:13:41.520 | And there are a lot of different kinds of memory

01:13:44.720 | in the brain.

01:13:45.560 | That's yet another example of how it's not a simple system.

01:13:49.160 | So there's one kind of memory,

01:13:53.000 | one way of talking about memory.

01:13:56.040 | Usually starts in the context of Hopfield networks.

01:13:59.040 | You were lucky to talk to John Hopfield on this program.

01:14:02.440 | But the basic idea there is working memory

01:14:05.840 | is stored in the dynamical patterns

01:14:07.800 | of activity between neurons.

01:14:10.360 | And you can think of a certain pattern of activity

01:14:14.720 | as an attractor, meaning if you put in some signal

01:14:19.640 | that's similar enough to other previously experienced

01:14:24.000 | signals like that, then you're going to converge

01:14:27.760 | to the same network dynamics and you will see these neurons

01:14:31.720 | participate in the same network patterns of activity

01:14:36.200 | that they have in the past.

01:14:37.600 | So you can talk about the probability that different inputs

01:14:41.260 | will allow you to converge to different basins of attraction

01:14:44.240 | and you might think of that as, oh, I saw this face

01:14:48.000 | and then I excited this network pattern of activity

01:14:50.920 | because last time I saw that face,

01:14:53.080 | I was at some movie and that's a famous person

01:14:56.960 | that's on the screen or something like that.

01:14:58.120 | So that's one memory storage mechanism,

01:15:00.560 | but crucial to the ability to imprint those memories

01:15:04.420 | in your brain is the ability to change the strength

01:15:08.560 | of connection between one neuron and another,

01:15:11.380 | that synaptic connection between them.

01:15:13.300 | So synaptic weight update is a massive field

01:15:16.620 | of neuroscience and neuromorphic computing as well.

01:15:19.580 | So there are two poles to that on that spectrum.

01:15:24.580 | Okay, so more in the language of machine learning,

01:15:28.900 | we would talk about supervised and unsupervised learning.

01:15:32.020 | And when I'm trying to tie that down

01:15:33.980 | to neuromorphic computing, I will use a definition

01:15:37.040 | of supervised learning, which basically means

01:15:40.660 | the external user, the person who's controlling

01:15:44.140 | this hardware has some knob that they can tune

01:15:48.380 | to change each of the synaptic weights,

01:15:50.420 | depending on whether or not the network's doing

01:15:52.460 | what you want it to do.

01:15:53.580 | Whereas what I mean in this conversation,

01:15:56.140 | when I say unsupervised learning is that

01:15:58.380 | those synaptic weights are dynamically changing

01:16:02.340 | in your network based on nothing that the user's doing,

01:16:05.040 | nothing that there's no wire from the outside

01:16:07.100 | going into any of those synapses.

01:16:09.080 | The network itself is reconfiguring those synaptic weights

01:16:12.100 | based on physical properties

01:16:15.780 | that you've built into the devices.

01:16:17.640 | So if the synapse receives a pulse from here

01:16:21.420 | and that causes the neuron to spike,

01:16:23.400 | some circuit built in there with no help

01:16:26.700 | from me or anybody else, adjust the weight

01:16:29.220 | in a way that makes it more likely to store

01:16:33.580 | the useful information and excite

01:16:35.020 | the useful network patterns and makes it less likely

01:16:37.720 | that random noise, useless communication events

01:16:41.460 | will have an important effect on the network activity.

01:16:45.340 | - So there's memory encoded in the weights,

01:16:48.280 | the synaptic weights.

01:16:49.600 | - Yes. - What about the formation

01:16:50.940 | of something that's not often done in machine learning,

01:16:53.680 | the formation of new synaptic connections?

01:16:56.320 | - Right, well, that seems to, so again,

01:16:58.200 | not a neuroscientist here,

01:17:00.160 | but my reading of the literature is that

01:17:02.200 | that's particularly crucial in early stages

01:17:05.000 | of brain development, where a newborn is born

01:17:09.160 | with tons of extra synaptic connections

01:17:11.680 | and it's actually pruned over time.

01:17:13.880 | So the number of synapses decreases

01:17:16.820 | as opposed to growing new long distance connections.

01:17:19.700 | It is possible in the brain to grow new neurons

01:17:22.300 | and assign new synaptic connections,

01:17:26.100 | but it doesn't seem to be the primary mechanism

01:17:29.140 | by which the brain is learning.

01:17:31.820 | So for example, like right now,

01:17:34.260 | sitting here talking to you,

01:17:35.740 | you say lots of interesting things

01:17:37.020 | and I learn from you and I can remember things

01:17:40.500 | that you just said.

01:17:41.340 | And I didn't grow new axonal connections

01:17:44.720 | down to new synapses to enable those.

01:17:47.380 | It's plasticity mechanisms in the synaptic connections

01:17:52.180 | between neurons that enable me to learn on that time scale.

01:17:55.980 | - So at the very least,

01:17:57.540 | you can sufficiently approximate that

01:17:59.860 | with just weight updates.

01:18:01.360 | You don't need to form new connections.

01:18:02.980 | - I would say weight updates are a big part of it.

01:18:05.060 | I also think there's more because broadly speaking,

01:18:08.620 | when we're doing machine learning,

01:18:10.440 | our networks, say we're talking about feed forward,

01:18:12.500 | deep neural networks,

01:18:14.140 | the temporal domain is not really part of it.

01:18:16.940 | Okay, you're gonna put in an image

01:18:18.220 | and you're gonna get out a classification

01:18:20.400 | and you're gonna do that as fast as possible.

01:18:22.000 | So you care about time,

01:18:23.140 | but time is not part of the essence of this thing really.

01:18:26.600 | Whereas in spiking neural networks,

01:18:30.020 | what we see in the brain,

01:18:31.740 | time is as crucial as space

01:18:33.340 | and they're intimately intertwined as I've tried to say.

01:18:35.980 | And so adaptation on different time scales

01:18:40.260 | is important, not just in memory formation,

01:18:44.080 | although it plays a key role there,

01:18:45.340 | but also in just keeping the activity

01:18:48.220 | in a useful dynamic range.

01:18:50.300 | So you have other plasticity mechanisms,

01:18:52.500 | not just weight update,

01:18:54.200 | or at least not on the time scale of many action potentials,

01:18:58.740 | but even on the shorter time scale.

01:19:00.180 | So a synapse can become much less efficacious.

01:19:04.700 | It can transmit a weaker signal

01:19:07.200 | after the second, third, fourth.

01:19:08.780 | That can second, third, fourth action potential

01:19:11.980 | to occur in a sequence.

01:19:13.080 | So that's what's called short-term synaptic plasticity,

01:19:15.980 | which is a form of learning.

01:19:17.620 | You're learning that I'm getting too much stimulus

01:19:19.660 | from looking at something bright right now.

01:19:21.660 | So I need to tone that down.

01:19:23.400 | There's also another really important mechanism in learning.

01:19:28.580 | It's called metaplasticity.

01:19:30.540 | What that seems to be is a way

01:19:33.580 | that you change not the weights themselves,

01:19:37.420 | but the rate at which the weights change.

01:19:40.260 | So when I am in, say, a lecture hall,

01:19:44.580 | and my, this is a potentially terrible cartoon example,

01:19:48.380 | but let's say I'm in a lecture hall

01:19:49.660 | and it's time to learn, right?

01:19:51.960 | So my brain will release more, perhaps, dopamine

01:19:55.300 | or some neuromodulator that's gonna change

01:19:58.380 | the rate at which synaptic plasticity occurs.

01:20:02.260 | So that can make me more sensitive

01:20:03.820 | to learning at certain times,

01:20:05.340 | more sensitive to overriding previous information,

01:20:08.380 | and less sensitive at other times.

01:20:10.340 | And finally, as long as I'm rattling off the list,

01:20:13.180 | I think another concept that falls in the category

01:20:16.480 | of learning or memory adaptation is homeostasis

01:20:20.540 | or homeostatic adaptation,

01:20:22.420 | where neurons have the ability

01:20:24.940 | to control their firing rate.

01:20:27.780 | So if one neuron is just like blasting way too much,

01:20:31.180 | it will naturally tone itself down.

01:20:33.000 | Its threshold will adjust so that it stays

01:20:37.380 | in a useful dynamical range.

01:20:38.500 | And we see that's captured in deep neural networks

01:20:41.660 | where you don't just change the synaptic weights,

01:20:43.300 | but you can also move the thresholds

01:20:44.780 | of simple neurons in those models.

01:20:47.580 | - And so to achieve the spiking neural networks,

01:20:52.580 | you want to use, like,

01:20:56.400 | you want to implement the first principles

01:21:01.200 | that you mentioned of the temporal

01:21:03.380 | and the spatial fractal dynamics here.

01:21:07.020 | So you can communicate locally,

01:21:09.260 | you can communicate across much greater distances

01:21:13.300 | and do the same thing in space

01:21:15.980 | and do the same thing in time.

01:21:18.020 | Now, you have, like, a chapter called

01:21:21.020 | superconducting hardware for neuromorphic computing.

01:21:24.340 | So what are some ideas that integrate

01:21:27.740 | some of the things we've been talking about

01:21:29.060 | in terms of the first principles

01:21:30.440 | of neuromorphic computing and the ideas

01:21:33.460 | that you outline in optoelectronic intelligence?

01:21:38.040 | - Yeah, so let me start, I guess,

01:21:40.920 | on the communication side of things,

01:21:42.500 | 'cause that's what led us down this track

01:21:46.260 | in the first place.

01:21:47.100 | By us, I'm talking about my team of colleagues at NIST,

01:21:51.900 | you know, Sayeed Han, Bryce Primavera,

01:21:54.020 | Sonia Buckley, Jeff Childs, Adam McConn,

01:21:56.500 | to name Alex Tate, to name a few,

01:21:58.700 | our group leader, Sebu Nam and Rich Miren.

01:22:01.200 | We've all contributed to this,

01:22:02.440 | so this is not me saying necessarily

01:22:05.840 | just the things that I've proposed,

01:22:07.520 | but sort of where our team's thinking

01:22:09.540 | has evolved over the years.

01:22:11.520 | - Can I quickly ask, what is NIST

01:22:14.640 | and where is this amazing group of people located?

01:22:18.040 | - NIST is the National Institute of Standards and Technology.

01:22:21.700 | The larger facility is out in Gaithersburg, Maryland.

01:22:26.660 | Our team is located in Boulder, Colorado.

01:22:29.680 | NIST is a federal agency under the Department of Commerce.

01:22:36.200 | We do a lot with, by we, I mean other people at NIST,

01:22:40.160 | do a lot with standards, you know,

01:22:42.620 | making sure that we understand the system of units,

01:22:46.080 | international system of units, precision measurements.

01:22:49.320 | There's a lot going on in electrical engineering,

01:22:53.540 | material science.

01:22:54.760 | - And it's historic.

01:22:55.720 | I mean, it's like, it's one of those,

01:22:57.520 | it's like MIT or something like that.

01:22:59.040 | It has a reputation over many decades

01:23:00.960 | of just being this really,

01:23:03.400 | a place where there's a lot of brilliant people

01:23:05.640 | have done a lot of amazing things.

01:23:07.600 | But in terms of the people in your team,

01:23:09.760 | in this team of people involved

01:23:12.740 | in the concept we're talking about now,

01:23:14.580 | I'm just curious, what kind of disciplines

01:23:16.400 | are we talking about?

01:23:17.240 | What is it?

01:23:18.080 | - Mostly physicists and electrical engineers,

01:23:20.240 | some material scientists,

01:23:23.000 | but I would say, yeah,

01:23:25.040 | I think physicists and electrical engineers,

01:23:27.240 | my background is in photonics,

01:23:29.480 | the use of light for technology.

01:23:31.040 | So coming from there,

01:23:33.360 | I tend to have found colleagues

01:23:36.840 | that are more from that background.

01:23:38.260 | Although Adam McConn,

01:23:40.280 | more of a superconducting electronics background,

01:23:42.720 | we need a diversity of folks.

01:23:44.300 | This project is sort of cross-disciplinary.

01:23:46.840 | I would love to be working more

01:23:48.300 | with neuroscientists and things,

01:23:50.800 | but we haven't reached that scale yet.

01:23:53.920 | But yeah.

01:23:54.760 | - You're focused on the hardware side,

01:23:56.480 | which requires all the disciplines that you mentioned.

01:23:59.200 | And then of course,

01:24:00.040 | neuroscience may be a source of inspiration

01:24:02.120 | for some of the long-term vision.

01:24:04.360 | - I would actually call it more than inspiration.

01:24:06.240 | I would call it sort of a roadmap.

01:24:11.080 | We're not trying to build exactly the brain,

01:24:15.000 | but I don't think it's enough to just say,

01:24:17.520 | oh, neurons kind of work like that,

01:24:19.240 | let's kind of do that thing.

01:24:20.760 | I mean, we're very much following the concepts

01:24:25.360 | that the cognitive sciences have laid out for us,

01:24:27.440 | which I believe is a really robust roadmap.

01:24:30.520 | I mean, just on a little bit of a tangent,

01:24:33.060 | it's often stated that we just don't understand the brain,

01:24:36.100 | and so it's really hard to replicate it

01:24:37.960 | because we just don't know what's going on there.

01:24:40.400 | And maybe five or seven years ago,

01:24:43.560 | I would have said that,

01:24:44.800 | but as I got more interested in the subject,

01:24:47.880 | I read more of the neuroscience literature,

01:24:50.480 | and I was just taken by the exact opposite sense.

01:24:53.640 | I can't believe how much they know about this.

01:24:55.880 | I can't believe how mathematically rigorous

01:24:59.280 | and sort of theoretically complete a lot of the concepts are.

01:25:04.200 | That's not to say we understand consciousness

01:25:06.660 | or we understand the self or anything like that,

01:25:08.560 | but what is the brain doing,

01:25:11.020 | and why is it doing those things?

01:25:13.560 | Neuroscientists have a lot of answers to those questions.

01:25:15.980 | So there's a lot, if you're a hardware designer

01:25:17.840 | that just wants to get going,

01:25:19.440 | whoa, it's pretty clear which direction to go in, I think.

01:25:23.000 | - Okay, so (laughs)

01:25:25.280 | I love the optimism behind that,

01:25:28.260 | but in the implementation of these systems

01:25:31.120 | that uses superconductivity,

01:25:37.040 | how do you make it happen?

01:25:38.360 | - So to me, it starts with thinking

01:25:41.880 | about the communication network.

01:25:43.400 | You know for sure that the ability of each neuron

01:25:47.540 | to communicate to many thousands of colleagues

01:25:50.560 | across the network is indispensable.

01:25:52.360 | I take that as a core principle of my architecture,

01:25:56.280 | my thinking on the subject.

01:25:57.640 | So coming from a background in photonics,

01:26:02.280 | it was very natural to say,

01:26:03.560 | okay, we're gonna use light for communication.

01:26:05.340 | Just in case listeners may not know,

01:26:08.720 | light is often used in communication.

01:26:10.840 | I mean, if you think about radio, that's light.

01:26:12.720 | It's long wavelengths, but it's electromagnetic radiation.

01:26:15.320 | It's the same physical phenomenon

01:26:17.620 | obeying exactly the same Maxwell's equations.

01:26:20.300 | And then all the way down to fiber optics,

01:26:24.900 | now you're using visible

01:26:26.220 | or near infrared wavelengths of light,

01:26:27.780 | but the way you send messages across the ocean

01:26:30.340 | is now contemporary over optical fibers.

01:26:33.180 | So using light for communication is not a stretch.

01:26:37.460 | It makes perfect sense.

01:26:38.940 | So you might ask, well, why don't you use light

01:26:41.500 | for communication in a conventional microchip?

01:26:45.260 | And the answer to that is, I believe, physical.

01:26:49.240 | If we had a light source on a silicon chip

01:26:53.060 | that was as simple as a transistor,

01:26:55.280 | there would not be a processor in the world

01:26:58.200 | that didn't use light for communication,

01:26:59.720 | at least above some distance.

01:27:01.760 | - How many light sources are needed?

01:27:04.000 | Oh, you need a light source at every single point.

01:27:06.800 | - A light source per neuron.

01:27:08.400 | - Per neuron, per little,

01:27:09.920 | but then if you could have a really small

01:27:13.040 | and nice light source, your definition of neuron

01:27:16.600 | could be flexible.

01:27:17.940 | - Could be, yes, yes.

01:27:19.160 | Sometimes it's helpful to me to say,

01:27:21.720 | in this hardware, a neuron is that entity

01:27:24.560 | which has a light source.

01:27:25.760 | That is, and I can explain--

01:27:28.000 | - And then there was light.

01:27:29.520 | - I mean, I can explain more about that, but--

01:27:32.280 | - Somehow this rhymes with consciousness,

01:27:34.680 | 'cause people often say the light of consciousness.

01:27:38.240 | So that consciousness is that which is conscious.

01:27:41.680 | I got it.

01:27:42.520 | (laughing)

01:27:43.600 | - That's not my quote.

01:27:44.840 | - That's me, that's my quote.

01:27:47.000 | See, that quote comes from my background.

01:27:49.480 | Yours is in optics, mine in light, mine's in darkness.

01:27:53.600 | (laughing)

01:27:55.400 | So go ahead, what were you saying?

01:27:57.600 | - The point I was making there is that

01:27:59.600 | if it was easy to manufacture light sources

01:28:02.920 | along with transistors on a silicon chip,

01:28:05.720 | they would be everywhere.

01:28:07.200 | And it's not easy.

01:28:08.840 | People have been trying for decades

01:28:10.120 | and it's actually extremely difficult.

01:28:11.920 | I think an important part of our research

01:28:13.680 | is dwelling right at that spot there.

01:28:17.360 | - Is it physics or engineering?

01:28:18.720 | - It's physics.

01:28:19.560 | So, okay, so it's physics, I think.

01:28:22.240 | So what I mean by that is, as we discussed,

01:28:26.920 | silicon is the material of choice for transistors

01:28:29.400 | and it's very difficult to imagine

01:28:33.240 | that that's gonna change anytime soon.

01:28:35.200 | Silicon is notoriously bad at emitting light,

01:28:39.120 | and that has to do with the immutable properties

01:28:43.280 | of silicon itself, the way that the energy bands

01:28:46.240 | are structured in silicon.

01:28:47.880 | You're never going to make silicon efficient

01:28:50.480 | as a light source at room temperature

01:28:53.520 | without doing very exotic things that degrade its ability

01:28:57.440 | to interface nicely with those transistors

01:28:59.160 | in the first place.

01:29:00.000 | So that's like one of these things where it's,

01:29:02.600 | why is nature dealing us that blow?

01:29:05.200 | You give us these beautiful transistors

01:29:07.000 | and you give us all the motivation

01:29:08.640 | to use light for communication,

01:29:10.120 | but then you don't give us a light source.

01:29:11.520 | So, well, okay, you do give us a light source,

01:29:14.000 | compound semiconductors, like we talked about

01:29:16.120 | back at the beginning, an element from group three

01:29:18.080 | and an element from group five form an alloy

01:29:20.320 | where every other lattice site switches which element it is.

01:29:23.680 | Those have much better properties for generating light.

01:29:27.280 | You put electrons in, light comes out.

01:29:30.060 | Almost 100% of the electron hold,

01:29:33.140 | it can be made efficient.

01:29:36.080 | - I'll take your word for it, okay.

01:29:37.440 | - However, I say it's physics, not engineering

01:29:39.600 | because it's very difficult to get those

01:29:43.160 | compound semiconductor light sources

01:29:45.280 | situated with your silicon.

01:29:47.560 | In order to do that ion implantation

01:29:49.440 | that I talked about at the beginning,

01:29:50.800 | high temperatures are required.

01:29:52.760 | So you got to make all of your transistors first

01:29:55.700 | and then put the compound semiconductors on top of there.

01:29:58.600 | You can't grow them afterwards

01:30:00.860 | because that requires high temperature.

01:30:02.400 | It screws up all your transistors.

01:30:04.060 | You try and stick them on there.

01:30:05.840 | They don't have the same lattice constant.

01:30:08.000 | The spacing between atoms is different enough

01:30:10.360 | that it just doesn't work.

01:30:11.680 | So nature does not seem to be telling us that,

01:30:15.520 | hey, go ahead and combine light sources

01:30:17.960 | with your digital switches

01:30:19.960 | for conventional digital computing.

01:30:22.680 | - And conventional digital computing

01:30:24.720 | will often require smaller scale, I guess,

01:30:27.640 | in terms of like a smartphone.

01:30:29.580 | So in which kind of systems does nature hint

01:30:35.440 | that we can use light and photons for communication?

01:30:40.440 | - Well, so let me just try and be clear.

01:30:43.560 | You can use light for communication in digital systems.

01:30:46.880 | Just the light sources are not intimately integrated

01:30:50.320 | with the silicon.

01:30:51.320 | You manufacture all the silicon,

01:30:52.680 | you have your microchip, plunk it down.

01:30:55.280 | And then you manufacture your light sources,

01:30:57.260 | separate chip, completely different process,

01:30:59.300 | made in a different foundry.

01:31:01.020 | And then you put those together at the package level.

01:31:03.760 | So now you have some,

01:31:06.720 | I would say a great deal of architectural limitations

01:31:10.080 | that are introduced by that sort of

01:31:14.480 | package level integration as opposed to monolithic

01:31:17.080 | on the same chip integration.

01:31:18.520 | But it's still a very useful thing to do.

01:31:20.120 | And that's where I had done some work previously

01:31:23.660 | before I came to NIST.

01:31:24.640 | There's a project led by Vladimir Stoyanovich

01:31:28.440 | that now spun out into a company called IR Labs,

01:31:31.360 | led by Mark Wade and Chen Sun,

01:31:33.720 | where they're doing exactly that.

01:31:35.000 | So you have your light source chip,

01:31:36.960 | your silicon chip, whatever it may be doing,

01:31:39.760 | maybe it's digital electronics,

01:31:40.880 | maybe it's some other control purpose, something.

01:31:43.620 | And the silicon chip drives the light source chip

01:31:47.760 | and modulates the intensity of the light.

01:31:49.760 | So you can get data out of the package on an optical fiber.

01:31:52.640 | And that still gives you tremendous advantages in bandwidth

01:31:56.560 | as opposed to sending those signals out

01:31:58.820 | over electrical lines.

01:32:00.760 | But it is somewhat peculiar to my eye

01:32:05.160 | that they have to be integrated at this package level.

01:32:07.880 | And those people, I mean, they're so smart.

01:32:09.760 | Those are my colleagues that I respect a great deal.

01:32:12.880 | So it's very clear that it's not just

01:32:15.520 | they're making a bad choice.

01:32:18.760 | This is what physics is telling us.

01:32:20.480 | It just wouldn't make any sense

01:32:22.280 | to try to stick them together.

01:32:24.240 | - Yeah, so even if it's difficult,

01:32:28.200 | it's easier than the alternative, unfortunately.

01:32:30.880 | - I think so, yes.

01:32:31.720 | And again, I need to go back

01:32:33.280 | and make sure that I'm not taking the wrong way.

01:32:35.100 | I'm not saying that the pursuit

01:32:36.800 | of integrating compound semiconductors with silicon

01:32:39.400 | is fruitless and shouldn't be pursued.

01:32:41.140 | It should, and people are doing great work.

01:32:43.440 | Kai May Lau and John Bowers, others,

01:32:45.680 | they're doing it and they're making progress.

01:32:48.240 | But to my eye, it doesn't look like

01:32:50.500 | that's ever going to be

01:32:53.660 | just the standard monolithic light source on silicon process.

01:32:58.600 | I just don't see it.

01:33:00.020 | - Yeah, so nature kind of points the way usually.

01:33:02.860 | And if you resist nature,

01:33:04.500 | you're going to have to do a lot more work.

01:33:05.700 | - And it's going to be expensive and not scalable.

01:33:07.700 | - Got it.

01:33:08.540 | But okay, so let's go far into the future.

01:33:11.300 | Let's imagine this gigantic neuromorphic computing system

01:33:14.760 | that simulates all of our realities.

01:33:17.340 | It currently is, we mentioned matrix four.

01:33:19.060 | So this thing, this powerful computer,

01:33:23.200 | how does it operate?

01:33:24.880 | So what are the neurons?

01:33:27.520 | What is the communication?

01:33:29.060 | What's your sense?

01:33:30.020 | - All right, so let me now,

01:33:32.460 | after spending 45 minutes

01:33:34.180 | trashing light source integration with silicon,

01:33:36.140 | let me now say why I'm basing my entire life,

01:33:40.140 | professional life,

01:33:41.060 | on integrating light sources with electronics.

01:33:44.960 | I think the game is completely different

01:33:47.020 | when you're talking about superconducting electronics.

01:33:49.580 | For several reasons,

01:33:52.740 | let me try to go through them.

01:33:54.240 | One is that, as I mentioned,

01:33:56.480 | it's difficult to integrate

01:33:57.960 | those compound semiconductor light sources with silicon.

01:34:01.280 | With silicon is a requirement

01:34:03.600 | that is introduced by the fact

01:34:05.240 | that you're using semiconducting electronics.

01:34:07.280 | In superconducting electronics,

01:34:08.840 | you're still going to start with a silicon wafer,

01:34:10.840 | but it's just the bread for your sandwich in a lot of ways.

01:34:13.880 | You're not using that silicon

01:34:15.800 | in precisely the same way for the electronics.

01:34:17.720 | You're now depositing superconducting materials

01:34:20.460 | on top of that.

01:34:21.840 | The prospects for integrating light sources

01:34:24.520 | with that kind of an electronic process

01:34:27.400 | are certainly less explored,

01:34:30.460 | but I think much more promising

01:34:31.940 | because you don't need those light sources

01:34:34.260 | to be intimately integrated with the transistors.

01:34:36.580 | That's where the problems come up.

01:34:37.900 | They don't need to be lattice matched to the silicon,

01:34:39.900 | all that kind of stuff.

01:34:41.140 | Instead, it seems possible

01:34:43.600 | that you can take those compound semiconductor light sources,

01:34:47.300 | stick them on the silicon wafer,

01:34:49.140 | and then grow your superconducting electronics

01:34:51.380 | on the top of that.

01:34:52.280 | It's at least not obviously going to fail.

01:34:55.760 | - So the computation would be done

01:34:57.260 | on the superconductive material as well?

01:35:00.100 | - Yes, the computation is done

01:35:01.940 | in the superconducting electronics,

01:35:03.920 | and the light sources receive signals that say,

01:35:06.720 | hey, a neuron reached threshold,

01:35:08.200 | produce a pulse of light,

01:35:09.800 | send it out to all your downstream synaptic connections.

01:35:12.460 | Those are, again, superconducting electronics.

01:35:16.220 | Perform your computation, and you're off to the races.

01:35:19.620 | Your network works.

01:35:20.780 | - So then if we can rewind real quick,

01:35:22.620 | so what are the limitations or the challenges

01:35:25.980 | of superconducting electronics

01:35:28.940 | when we think about constructing these kinds of systems?

01:35:31.500 | - So actually, let me say one other thing

01:35:35.620 | about the light sources.

01:35:36.940 | - Yes, please.

01:35:37.780 | - And then I'll move on, I promise,

01:35:39.840 | 'cause this is probably tedious for some.

01:35:42.340 | - This is super exciting.

01:35:44.020 | - Okay, one other thing about the light sources.

01:35:45.740 | I said that silicon is terrible at emitting photons.

01:35:48.900 | It's just not what it's meant to do.

01:35:50.660 | However, the game is different

01:35:52.820 | when you're at low temperature.

01:35:54.060 | If you're working with superconductors,

01:35:55.700 | you have to be at low temperature

01:35:56.980 | 'cause they don't work otherwise.

01:35:58.700 | When you're at four Kelvin,

01:36:00.260 | silicon is not obviously a terrible light source.

01:36:03.420 | It's still not as efficient as compound semiconductors,

01:36:05.860 | but it might be good enough for this application.

01:36:08.680 | The final thing that I'll mention about that is, again,

01:36:11.300 | leveraging superconductors, as I said,

01:36:13.980 | in a different context,

01:36:15.380 | superconducting detectors can receive one single photon.

01:36:19.540 | In that conversation, I failed to mention

01:36:21.260 | that semiconductors can also receive photons.

01:36:23.460 | That's the primary mechanism by which it's done.

01:36:26.220 | A camera in your phone that's receptive to visible light

01:36:29.180 | is receiving photons.

01:36:31.100 | It's based on silicon,

01:36:32.220 | or you can make it in different semiconductors

01:36:34.380 | for different wavelengths,

01:36:36.580 | but it requires on the order of 1,000,

01:36:39.580 | a few thousand photons to receive a pulse.

01:36:43.580 | Now, when you're using a superconducting detector,

01:36:46.160 | you need one photon, exactly one.

01:36:48.360 | I mean, one or more.

01:36:50.860 | So the fact that your synapses can now be based

01:36:54.840 | on superconducting detectors

01:36:56.660 | instead of semiconducting detectors

01:36:58.800 | brings the light levels that are required

01:37:00.700 | down by some three orders of magnitude.

01:37:03.140 | So now you don't need good light sources.

01:37:06.540 | You can have the world's worst light sources.

01:37:08.780 | As long as they spit out maybe a few thousand photons

01:37:11.820 | every time a neuron fires,

01:37:13.780 | you have the hardware principles in place

01:37:17.640 | that you might be able to perform

01:37:19.620 | this optoelectronic integration.

01:37:21.700 | To me, optoelectronic integration, it's just so enticing.

01:37:25.020 | We want to be able to leverage

01:37:26.980 | electronics for computation, light for communication,

01:37:30.380 | working with silicon microelectronics at room temperature

01:37:32.820 | that has been exceedingly difficult.

01:37:35.020 | And I hope that when we move to the superconducting domain,

01:37:39.980 | target a different application space

01:37:41.820 | that is neuromorphic instead of digital

01:37:44.900 | and use superconducting detectors,

01:37:47.660 | maybe optoelectronic integration comes to us.

01:37:50.120 | - Okay, so there's a bunch of questions.

01:37:51.720 | So one is temperature.

01:37:53.680 | So in these kind of hybrid heterogeneous systems,

01:37:58.320 | what's the temperature?

01:37:59.560 | What are some of the constraints of the operation here?

01:38:01.600 | Does it all have to be at four Kelvin as well?

01:38:03.600 | - Four Kelvin.

01:38:04.440 | Everything has to be at four Kelvin.

01:38:06.840 | - Okay, so what are the other engineering challenges

01:38:09.720 | of making this kind of optoelectronic systems?

01:38:14.320 | - Let me just dwell on that four Kelvin for a second

01:38:16.540 | because some people hear four Kelvin

01:38:18.300 | and they just get up and leave.

01:38:19.260 | They just say, "I'm not doing it."

01:38:21.540 | And to me, that's very earth-centric, species-centric.

01:38:25.340 | We live in 300 Kelvin,

01:38:27.260 | so we want our technologies to operate there too.

01:38:29.140 | I totally get it.

01:38:30.100 | - Yeah, what's zero Celsius?

01:38:31.900 | - Zero Celsius is 273 Kelvin.

01:38:34.500 | So we're talking very, very cold here.

01:38:37.620 | - Not even Boston cold.

01:38:39.540 | (laughing)

01:38:41.340 | This is real cold.

01:38:42.380 | - Yeah.

01:38:43.220 | - Siberia cold.

01:38:44.040 | - Okay, so just for reference,

01:38:45.660 | the temperature of the cosmic microwave background

01:38:47.900 | is about 2.7 Kelvin.

01:38:49.380 | So we're still warmer than deep space.

01:38:51.660 | - Yeah, good.

01:38:52.980 | So when the universe dies out, it'll be colder than 4K.

01:38:57.940 | - It's already colder than 4K.

01:38:59.380 | In the expanses, you don't have to get that far

01:39:03.860 | away from the earth in order to drop down

01:39:06.940 | to not far from four Kelvin.

01:39:07.780 | - So what you're saying is the aliens that live

01:39:09.820 | at the edge of the observable universe

01:39:13.280 | are using superconducting material for their computation.

01:39:16.420 | - They don't have to live at the edge of the universe.

01:39:17.860 | The aliens that are more advanced than us

01:39:21.040 | in their solar system are doing this in their asteroid belt.

01:39:26.040 | We can get to that.

01:39:27.780 | - Oh, because they can get to that temperature easier there?

01:39:31.660 | - Sure, yeah, all you have to do is reflect the sunlight

01:39:33.780 | away and you have a huge head start.

01:39:36.060 | - Oh, so the sun is the problem here.

01:39:37.660 | Like it's warm here on earth.

01:39:39.020 | Got it.

01:39:39.860 | - Yeah.

01:39:40.700 | - Okay, so how do we get to 4K?

01:39:42.900 | - Well, okay, so what I want to say about temperature

01:39:47.900 | is that if you can swallow that, if you can say,

01:39:53.940 | all right, I give up applications that have to do

01:39:56.760 | with my cell phone and the convenience of a laptop

01:40:01.420 | on a train and you instead, for me,

01:40:04.300 | I'm very much in the scientific headspace.

01:40:06.420 | I'm not looking at products.

01:40:07.940 | I'm not looking at what this will be useful

01:40:09.880 | to sell to consumers.

01:40:11.000 | Instead, I'm thinking about scientific questions.

01:40:13.380 | Well, it's just not that bad to have to work at 4K.

01:40:16.540 | We do it all the time in our labs at NIST.

01:40:18.940 | And so does, I mean, for reference,

01:40:21.400 | the entire quantum computing sector

01:40:25.540 | usually has to work at something like 100 millikelvin,

01:40:28.620 | 50 millikelvin.

01:40:29.640 | So now you're talking of another factor of 100

01:40:32.100 | even colder than that, a fraction of a degree.

01:40:35.100 | And everybody seems to think quantum computing

01:40:37.340 | is gonna take over the world.

01:40:39.160 | It's so much more expensive to have to get

01:40:41.580 | that extra factor of 10 or whatever colder.

01:40:46.580 | And yet it's not stopping people from investing

01:40:49.260 | in that area.

01:40:50.260 | And by investing, I mean, putting their research into it

01:40:53.840 | as well as venture capital or whatever.

01:40:55.740 | - So based on the energy of what you're commenting on,

01:40:59.580 | I'm getting a sense that's one of the criticism

01:41:01.940 | of this approach is 4K, 4Kelvin is a big negative.

01:41:06.700 | It is the showstopper for a lot of people.

01:41:10.680 | They just, I mean, and understandably,

01:41:12.880 | I'm not saying that that's not a consideration.

01:41:16.880 | Of course it is.

01:41:17.840 | For some, okay, so different motivations

01:41:20.480 | for different people.

01:41:21.480 | In the academic world, suppose you spent your whole life

01:41:24.440 | learning about silicon microelectronic circuits.

01:41:26.800 | You send a design to a foundry, they send you back a chip

01:41:30.560 | and you go test it at your tabletop.

01:41:33.020 | And now I'm saying, here, now learn how to use

01:41:35.720 | all these cryogenics so you can do that at 4Kelvin.

01:41:38.460 | No, come on, man, I don't wanna do that.

01:41:41.220 | That sounds bad.

01:41:42.060 | - It's the old momentum, the Titanic of the turning.

01:41:44.540 | - Yeah, kind of.

01:41:45.640 | - But you're saying that's not too much of a,

01:41:48.380 | when we're looking at large systems and the gain

01:41:51.140 | you can potentially get from them,

01:41:52.380 | that's not that much of a cost.

01:41:53.460 | - And when you wanna answer the scientific question

01:41:55.180 | about what are the physical limits of cognition?

01:41:58.160 | Well, the physical limits, they don't care

01:42:00.500 | if you're at 4Kelvin.

01:42:01.460 | If you can perform cognition at a scale,

01:42:04.620 | orders of magnitude beyond any room temperature technology,

01:42:07.680 | but you gotta get cold to do it, you're gonna do it.

01:42:10.440 | And to me, that's the interesting application space.

01:42:14.600 | It's not even an application space.

01:42:16.120 | That's the interesting scientific paradigm.

01:42:18.920 | So I personally am not going to let low temperature

01:42:22.560 | stop me from realizing a technological domain

01:42:26.560 | or realm that is achieving in most ways

01:42:33.060 | everything else that I'm looking for in my hardware.

01:42:36.160 | - So that, okay, that's a big one.

01:42:37.640 | Is there other kind of engineering challenges

01:42:40.000 | that you envision?

01:42:40.840 | - Yeah, yeah, yeah.

01:42:41.660 | So let me take a moment here

01:42:43.120 | because I haven't really described what I mean

01:42:45.760 | by a neuron or a network in this particular hardware.

01:42:49.000 | - Yeah, do you wanna talk about loop neurons?

01:42:51.640 | And there's so many fascinating,

01:42:53.680 | but you just have so many amazing papers

01:42:55.960 | that people should definitely check out.

01:42:57.720 | And the titles alone are just killer.

01:42:59.840 | So anyway, go ahead.

01:43:01.080 | - Right, so let me say big picture based on optics,

01:43:06.040 | photonics for communication,

01:43:07.740 | superconducting electronics for computation.

01:43:10.140 | How does this all work?

01:43:11.500 | So a neuron in this hardware platform

01:43:16.500 | can be thought of as circuits

01:43:19.540 | that are based on Joseph's injunctions,

01:43:21.100 | like we talked about before,

01:43:22.700 | where every time a photon comes in,

01:43:25.300 | so let's start by talking about a synapse.

01:43:27.140 | A synapse receives a photon, one or more,

01:43:29.700 | from a different neuron,

01:43:31.440 | and it converts that optical signal

01:43:33.600 | to an electrical signal.

01:43:35.340 | The amount of current that that adds to a loop

01:43:38.940 | is controlled by the synaptic weight.

01:43:40.820 | So as I said before,

01:43:42.360 | you're popping fluxons into a loop, right?

01:43:44.520 | So a photon comes in,

01:43:46.440 | it hits a superconducting single photon detector,

01:43:49.200 | one photon, the absolute physical minimum

01:43:52.100 | that you can communicate

01:43:53.040 | from one place to another with light.

01:43:54.520 | - And that detector then converts that

01:43:56.120 | into an electrical signal,

01:43:57.380 | and the amount of the signal is correlated

01:44:00.000 | with some kind of weight.

01:44:01.200 | - Yeah, so the synaptic weight will tell you

01:44:02.920 | how many fluxons you pop into the loop.

01:44:05.480 | It's an analog number.

01:44:06.600 | We're doing analog computation now.

01:44:08.160 | - Well, can you just linger on that?

01:44:09.760 | What the heck is a fluxon?

01:44:11.040 | Are we supposed to know this?

01:44:12.000 | Or is this a funny, it's like the Big Bang.

01:44:15.640 | Is this a funny word for something deeply technical?

01:44:18.880 | - No, let's try to avoid using the word fluxon

01:44:21.080 | 'cause it's not actually necessary.

01:44:23.000 | When a photon--

01:44:23.840 | - It's fun to say, though.

01:44:25.480 | So it's very necessary, I would say.

01:44:29.040 | - When a photon hits that superconducting

01:44:31.260 | single photon detector,

01:44:32.940 | current is added to a superconducting loop,

01:44:36.580 | and the amount of current that you add

01:44:39.220 | is an analog value.

01:44:40.340 | It can have eight-bit equivalent resolution,

01:44:43.020 | something like that, 10 bits, maybe.

01:44:45.460 | - That's amazing, by the way.

01:44:46.980 | This is starting to make a lot more sense.

01:44:48.540 | - When you're using superconductors for this,

01:44:50.660 | the energy of that circulating current

01:44:54.740 | is less than the energy of that photon.

01:44:58.640 | So your energy budget is not destroyed

01:45:01.800 | by doing this analog computation.

01:45:04.040 | So now, in the language of a neuroscientist,

01:45:07.120 | you would say that's your postsynaptic signal.

01:45:09.200 | You have this current being stored in a loop.

01:45:11.720 | You can decide what you wanna do with it.

01:45:13.520 | Most likely, you're gonna have it decay exponentially.

01:45:16.200 | So every single synapse is gonna have

01:45:18.920 | some given time constant,

01:45:21.020 | and that's determined by putting some resistor

01:45:25.420 | in that superconducting loop.

01:45:27.100 | So a synapse event occurs when a photon

01:45:30.140 | strikes a detector, adds current to that loop,

01:45:32.480 | it decays over time.

01:45:33.880 | That's the postsynaptic signal.

01:45:35.580 | Then you can process that in a dendritic tree.

01:45:38.420 | Bryce Primavera and I have a paper

01:45:41.100 | that we've submitted about that.

01:45:43.420 | For the more neuroscience-oriented people,

01:45:45.380 | there's a lot of dendritic processing,

01:45:47.060 | a lot of plasticity mechanisms you can implement

01:45:49.460 | with essentially exactly the same circuits.

01:45:51.480 | You have this one simple building block circuit

01:45:54.460 | that you can use for a synapse, for a dendrite,

01:45:57.260 | for the neuron cell body, for all the plasticity functions.

01:46:00.320 | It's all based on the same building block,

01:46:02.360 | just tweaking a couple parameters.

01:46:03.780 | - So this basic building block has both an optical

01:46:05.980 | and an electrical component,

01:46:07.260 | and then you can just build

01:46:09.500 | arbitrary large systems with that.

01:46:11.660 | - Close.

01:46:12.500 | You're not at fault for thinking that that's what I meant.

01:46:14.980 | What I should say is that if you want it to be a synapse,

01:46:18.780 | you tack a superconducting detector onto the front of it.

01:46:22.260 | And if you want it to be anything else,

01:46:23.460 | there's no optical component.

01:46:25.260 | - Got it.

01:46:26.100 | So at the front, (laughs)

01:46:28.380 | optics in the front, electrical stuff in the back.

01:46:32.140 | - Electrical, yeah, in the processing

01:46:34.140 | and in the output signal that it sends

01:46:36.460 | to the next stage of processing further.

01:46:39.380 | - So the dendritic trees is electrical.

01:46:41.500 | - It's all electrical.

01:46:42.500 | It's all electrical in the superconducting domain.

01:46:44.900 | For anybody who's up on their superconducting circuits,

01:46:48.540 | it's just based on a DC squid, the most ubiquitous,

01:46:52.460 | which is a circuit composed of two Josephson junctions.

01:46:55.140 | So it's a very bread and butter kind of thing.

01:46:58.700 | And then the only place where you go beyond that

01:47:00.900 | is the neuron cell body itself.

01:47:03.020 | It's receiving all these electrical inputs

01:47:05.300 | from the synapses or dendrites,

01:47:06.860 | or however you've structured that particular unique neuron.

01:47:09.820 | And when it reaches its threshold,

01:47:12.300 | which occurs by driving a Josephson junction

01:47:14.380 | above its critical current,

01:47:15.740 | it produces a pulse of current,

01:47:17.220 | which starts an amplification sequence,

01:47:19.460 | voltage amplification,

01:47:21.380 | that produces light out of a transmitter.

01:47:24.500 | So one of our colleagues, Adam McConn,

01:47:26.860 | and Sonia Buckley as well,

01:47:27.980 | did a lot of work on the light sources

01:47:30.980 | and the amplifiers that drive the current

01:47:34.660 | and produce sufficient voltage to drive current

01:47:37.540 | through that now semiconducting part.

01:47:39.180 | So that light source is the semiconducting part of a neuron.

01:47:43.180 | And that, so the neuron has reached threshold.

01:47:45.620 | It produces a pulse of light.

01:47:47.660 | That light then fans out across a network of wave guides

01:47:51.580 | to reach all the downstream synaptic terminals

01:47:54.540 | that perform this process themselves.

01:47:57.300 | So it's probably worth explaining

01:47:59.860 | what a network of wave guides is,

01:48:02.340 | 'cause a lot of listeners aren't gonna know that.

01:48:04.780 | Look up the papers by Jeff Chiles on this one.

01:48:07.140 | But basically, light can be guided in a simple,

01:48:11.380 | basically wire of usually an insulating material.

01:48:14.860 | So silicon, silicon nitride,

01:48:18.260 | different kinds of glass,

01:48:20.020 | just like in a fiber optic, it's glass, silicon dioxide.

01:48:23.420 | That makes it a little bit big.

01:48:24.780 | We wanna bring these down.

01:48:26.100 | So we use different materials like silicon nitride,

01:48:28.020 | but basically just imagine a rectangle of some material

01:48:32.940 | that just goes and branches,

01:48:35.540 | forms different branch points

01:48:39.940 | that target different sub regions of the network.

01:48:43.020 | You can transition between layers of these.

01:48:44.980 | So now we're talking about building in the third dimension,

01:48:47.140 | which is absolutely crucial.

01:48:48.820 | So that's what wave guides are.

01:48:50.700 | - Yeah, that's great.

01:48:52.060 | Why the third dimension is crucial?

01:48:54.660 | - Okay, so yes, you were talking about

01:48:56.580 | what are some of the technical limitations.

01:48:58.740 | One of the things that I believe we have to grapple with

01:49:04.260 | is that our brains are miraculously compact.

01:49:08.740 | For the number of neurons that are in our brain,

01:49:11.540 | it sure does fit in a small volume,

01:49:13.500 | as it would have to if we're gonna be biological organisms

01:49:16.180 | that are resource limited and things like that.

01:49:19.220 | Any kind of hardware neuron

01:49:20.940 | is almost certainly gonna be much bigger than that

01:49:23.420 | if it is of comparable complexity,

01:49:26.140 | even whether it's based on silicon transistors.

01:49:28.500 | Okay, a transistor, seven nanometers,

01:49:30.700 | that doesn't mean a semiconductor-based neuron

01:49:33.700 | is seven nanometers.

01:49:34.900 | They're big.

01:49:35.740 | They require many transistors,

01:49:38.740 | different other things like capacitors

01:49:40.260 | and things that store charge.

01:49:41.540 | They end up being on the order of 100 microns

01:49:44.660 | by 100 microns,

01:49:45.900 | and it's difficult to get them down any smaller than that.

01:49:48.260 | The same is true for superconducting neurons,

01:49:50.860 | and the same is true

01:49:52.100 | if we're trying to use light for communication.

01:49:54.020 | Even if you're using electrons for communication,

01:49:56.860 | you have these wires where,

01:50:00.060 | okay, the size of an electron might be angstroms,

01:50:03.380 | but the size of a wire is not angstroms,

01:50:05.580 | and if you try and make it narrower,

01:50:07.140 | the resistance just goes up,

01:50:08.500 | so you don't actually win.

01:50:10.740 | To communicate over long distances,

01:50:12.340 | you need your wires to be microns wide,

01:50:15.660 | and it's the same thing for waveguides.

01:50:17.180 | Waveguides are essentially limited

01:50:18.940 | by the wavelength of light,

01:50:20.300 | and that's gonna be about a micron,

01:50:21.900 | so whereas compare that to an axon,

01:50:24.700 | the analogous component in the brain,

01:50:26.780 | which is 10 nanometers in diameter, something like that,

01:50:31.100 | they're bigger when they need to communicate

01:50:33.660 | over long distances,

01:50:34.620 | but grappling with the size of these structures

01:50:37.980 | is inevitable and crucial,

01:50:39.820 | and so in order to make systems

01:50:43.820 | of comparable scale to the human brain,

01:50:46.020 | by scale here I mean number of interconnected neurons,

01:50:49.820 | you absolutely have to be using

01:50:51.740 | the third spatial dimension,

01:50:53.660 | and that means on the wafer,

01:50:55.940 | you need multiple layers

01:50:57.380 | of both active and passive components.

01:50:59.580 | Active, I mean superconducting electronic circuits

01:51:03.380 | that are performing computations,

01:51:05.380 | and passive, I mean these waveguides

01:51:07.380 | that are routing the optical signals to different places,

01:51:10.140 | you have to be able to stack those.

01:51:11.700 | If you can get to something like 10 planes of each of those,

01:51:15.260 | or maybe not even 10, maybe five, six, something like that,

01:51:18.860 | then you're in business.

01:51:19.740 | Now you can get millions of neurons on a wafer,

01:51:22.860 | but that's not anywhere close to the brain scale.

01:51:26.380 | In order to get to the scale of the human brain,

01:51:27.980 | you're gonna have to also use the third dimension

01:51:30.020 | in the sense that entire wafers

01:51:32.580 | need to be stacked on top of each other

01:51:34.140 | with fiber optic communication between them,

01:51:36.100 | and we need to be able to fill a space

01:51:38.620 | the size of this table with stacked wafers,

01:51:42.020 | and that's when you can get to some 10 billion neurons

01:51:44.260 | like your human brain.

01:51:45.100 | And I don't think that's specific

01:51:46.660 | to the optoelectronic approach that we're taking.

01:51:48.780 | I think that applies to any hardware

01:51:51.400 | where you're trying to reach commensurate scale

01:51:53.540 | and complexity as the human brain.

01:51:55.020 | - So you need that fractal stacking.

01:51:57.540 | So stacking on the wafer, and stacking of the wafers,

01:52:01.100 | and then whatever the system,

01:52:03.820 | this stacking of the tables with the wafers.

01:52:06.380 | - And it has to be fractal all the way.

01:52:07.700 | You're exactly right, because that's the only way

01:52:10.060 | that you can efficiently get information

01:52:12.500 | from a small point to across that whole network.

01:52:15.060 | It has to have the power law connected.

01:52:17.500 | - And photons all like optics throughout.

01:52:20.660 | - Yeah, absolutely.

01:52:21.480 | Once you're at this scale, to me it's just obvious.

01:52:23.700 | Of course you're using light for communication.

01:52:25.580 | You have fiber optics given to us from nature, so simple.

01:52:31.660 | The thought of even trying to do this,

01:52:33.860 | any kind of electrical communication just doesn't,

01:52:37.140 | it doesn't make sense to me.

01:52:38.100 | I'm not saying it's wrong, I don't know,

01:52:40.220 | but that's where I'm coming from.

01:52:41.820 | - So let's return to loop neurons.

01:52:44.800 | Why are they called loop neurons?

01:52:47.360 | - Yeah, the term loop neurons comes from the fact,

01:52:49.780 | like we've been talking about,

01:52:50.880 | that they rely heavily on these superconducting loops.

01:52:54.140 | So even in a lot of forms of digital computing

01:52:58.340 | as superconductors, storing a signal

01:53:01.980 | in a superconducting loop is a primary technique.

01:53:05.900 | In this particular case,

01:53:07.460 | it's just loops everywhere you look.

01:53:09.420 | So the strength of a synaptic weight is gonna be set

01:53:13.580 | by the amount of current circulating in a loop

01:53:16.580 | that is coupled to the synapse.

01:53:18.620 | So memory is implemented as current circulating

01:53:23.300 | in a superconducting loop.

01:53:24.900 | The coupling between, say, a synapse and a dendrite

01:53:27.960 | or a synapse in the neuron cell body

01:53:29.940 | occurs through loop coupling through transformers.

01:53:33.900 | So current circulating in a synapse

01:53:35.560 | is gonna induce current in a different loop,

01:53:38.300 | a receiving loop in the neuron cell body.

01:53:41.540 | So since all of the computation is happening

01:53:45.220 | in these flux storage loops,

01:53:47.420 | and they play such a central role

01:53:49.140 | in how the information is processed,

01:53:50.980 | how memories are formed, all that stuff,

01:53:53.560 | I didn't think too much about it.

01:53:54.700 | I just called them loop neurons,

01:53:56.280 | 'cause it rolls off the tongue a little bit

01:53:57.940 | better than superconducting optoelectronic neurons.

01:54:01.240 | - Okay, so how do you design circuits

01:54:06.020 | for these loop neurons?

01:54:08.540 | - That's a great question.

01:54:09.740 | There's a lot of different scales of design.

01:54:12.340 | So at the level of just one synapse,

01:54:16.360 | you can use conventional methods.

01:54:18.980 | They're not that complicated

01:54:21.100 | as far as superconducting electronics goes.

01:54:23.220 | It's just four Josephson junctions or something like that,

01:54:27.100 | depending on how much complexity you wanna add.

01:54:29.180 | So you can just directly simulate each component in SPICE.

01:54:34.180 | - What's SPICE?

01:54:35.860 | - It's Standard Electrical Simulation Software, basically.

01:54:39.220 | So you're just explicitly solving

01:54:41.480 | the differential equations that describe the circuit elements.

01:54:44.340 | - And then you can stack these things together

01:54:46.200 | in that simulation software to then build circuits.

01:54:48.860 | - You can, but that becomes computationally expensive.

01:54:51.560 | So one of the things, when COVID hit,

01:54:54.100 | we knew we had to turn some attention

01:54:55.760 | to more things you can do at home in your basement

01:54:59.380 | or whatever, and one of them was computational modeling.

01:55:02.700 | So we started working on adapting,

01:55:07.700 | abstracting out the circuit performance

01:55:10.160 | so that you don't have to explicitly solve

01:55:12.540 | the circuit equations, which for Josephson junctions

01:55:15.860 | usually needs to be done on like a picosecond timescale

01:55:18.460 | and you have a lot of nodes in your circuit.

01:55:21.220 | So it results in a lot of differential equations

01:55:24.820 | that need to be solved simultaneously.

01:55:26.320 | We were looking for a way to simulate these circuits

01:55:29.400 | that is scalable up to networks of millions or so neurons

01:55:33.720 | is sort of where we're targeting right now.

01:55:36.600 | So we were able to analyze the behavior of these circuits.

01:55:40.780 | And as I said, it's based on these simple building blocks.

01:55:43.600 | So you really only need to understand

01:55:45.120 | this one building block.

01:55:46.320 | And if you get a good model of that, boom, it tiles.

01:55:48.960 | And you can change the parameters in there

01:55:51.120 | to get different behaviors and stuff,

01:55:52.800 | but it's all based on now it's one differential equation

01:55:56.160 | that you need to solve.

01:55:57.100 | So one differential equation for every synapse,

01:56:00.740 | dendrite or neuron in your system.

01:56:03.700 | And for the neuroscientists out there,

01:56:05.300 | it's just a simple leaky integrate and fire model,

01:56:08.380 | leaky integrator basically.

01:56:09.940 | The synapse is a leaky integrator,

01:56:11.820 | a dendrite is a leaky integrator.

01:56:13.420 | So I'm really fascinated by how this one simple component

01:56:18.420 | can be used to achieve lots of different types

01:56:22.200 | of dynamical activity.

01:56:24.420 | And to me, that's where scalability comes from.

01:56:27.540 | And also complexity as well.

01:56:29.180 | Complexity is often characterized

01:56:30.880 | by relatively simple building blocks

01:56:35.420 | connected in potentially simple

01:56:37.820 | or sometimes complicated ways

01:56:39.360 | and then emergent new behavior that was hard to predict

01:56:41.940 | from those simple elements.

01:56:44.700 | And that's exactly what we're working with here.

01:56:46.620 | So it's a very exciting platform,

01:56:49.040 | both from a modeling perspective

01:56:50.380 | and from a hardware manifestation perspective

01:56:52.740 | where we can hopefully start to have this test bed

01:56:57.360 | where we can explore things,

01:56:58.860 | not just related to neuroscience,

01:57:00.820 | but also related to other things

01:57:03.500 | that connected to other physics like critical phenomenon,

01:57:07.140 | Ising models, things like that.

01:57:08.660 | So you were asking how we simulate these circuits.

01:57:13.060 | It's at different levels

01:57:14.540 | and we've got the simple spice circuit stuff.

01:57:18.340 | That's no problem.

01:57:19.540 | And now we're building these network models

01:57:21.740 | based on this more efficient leaky integrator.

01:57:23.620 | So we can actually reduce every element

01:57:26.220 | to one differential equation.

01:57:27.460 | And then we can also step through it

01:57:28.860 | on a much coarser time grid.

01:57:30.700 | So it ends up being something like a factor

01:57:32.380 | of a thousand to 10,000 speed improvement,

01:57:35.380 | which allows us to simulate,

01:57:37.760 | but hopefully up to millions of neurons.

01:57:40.520 | Whereas before we would have been limited

01:57:42.780 | to tens, a hundred, something like that.

01:57:45.780 | - And just like simulating quantum mechanical systems

01:57:48.820 | with a quantum computer.

01:57:49.860 | So the goal here is to understand such systems.

01:57:53.380 | - For me, the goal is to study this

01:57:55.900 | as a scientific physical system.

01:57:58.940 | I'm not drawn towards turning this into an enterprise

01:58:02.780 | at this point.

01:58:03.620 | I feel-

01:58:04.440 | - Short-term applications that obviously make a lot of money

01:58:07.460 | is not necessarily a curiosity driver

01:58:10.500 | for you at the moment.

01:58:11.400 | - Absolutely not.

01:58:12.240 | If you're interested in short-term making money,

01:58:14.020 | go with deep learning, use silicon microelectronics.

01:58:16.500 | If you wanna understand things like the physics

01:58:20.880 | of a fascinating system,

01:58:23.020 | or if you wanna understand something more along the lines

01:58:26.220 | of the physical limits of what can be achieved,

01:58:29.480 | then I think single photon communication,

01:58:32.680 | superconducting electronics, extremely exciting.

01:58:35.940 | - What if I wanna use superconducting hardware

01:58:39.620 | at four Kelvin to mine Bitcoin?

01:58:42.020 | That's my main interest.

01:58:43.260 | That's the reason I wanted to talk to you today.

01:58:45.860 | I wanna, no, I don't know.

01:58:47.620 | - What's Bitcoin?

01:58:48.620 | (Lex laughing)

01:58:50.300 | - It's a, look it up on the internet.

01:58:53.140 | Somebody told me about it.

01:58:54.780 | I'm not sure exactly what it is.

01:58:56.380 | But let me ask nevertheless

01:58:59.060 | about applications to machine learning.

01:59:01.500 | - Okay.

01:59:02.340 | - So what, like if you look at the scale of five, 10, 20 years

01:59:07.340 | is it possible to, before we understand the nature

01:59:12.100 | of human intelligence and general intelligence,

01:59:14.620 | do you think we'll start falling out of this exploration

01:59:19.260 | of neuromorphic systems ability to solve some

01:59:23.260 | of the problems that the machine learning systems

01:59:25.360 | of today can't solve?

01:59:26.580 | - I'm really hesitant to over promise.

01:59:31.620 | So I really don't know.

01:59:34.140 | Also, I don't really understand machine learning

01:59:36.780 | in a lot of senses.

01:59:37.600 | I mean, machine learning from my perspective

01:59:42.480 | appears to require that you know precisely

01:59:47.420 | what your input is and also what your goal is.

01:59:51.740 | You usually have some objective function

01:59:53.340 | or something like that.

01:59:54.180 | And that's just, that's very limiting.

01:59:57.300 | I mean, of course, a lot of times that's the case,

02:00:00.180 | you know, there's a picture and there's a horse in it,

02:00:03.100 | so you're done.

02:00:03.940 | But that's not a very interesting problem.

02:00:06.520 | I think when I think about intelligence,

02:00:09.500 | it's almost defined by the ability to handle problems

02:00:13.200 | where you don't know what your inputs are going to be

02:00:15.960 | and you don't even necessarily know

02:00:17.400 | what you're trying to accomplish.

02:00:18.600 | I mean, I'm not sure what I'm trying to accomplish

02:00:21.640 | in this world.

02:00:22.560 | - Yeah, at all scales.

02:00:24.560 | - Yeah, at all scales, right.

02:00:25.920 | I mean, so it's sometimes, so I'm more drawn

02:00:29.400 | to the underlying phenomena,

02:00:33.720 | the critical dynamics of this system,

02:00:37.960 | trying to understand how elements

02:00:41.340 | that you build into your hardware

02:00:43.620 | result in emergent, fascinating activity

02:00:48.620 | that was very difficult to predict, things like that.

02:00:51.740 | So, but I gotta be really careful

02:00:53.620 | because I think a lot of other people who,

02:00:55.540 | if they found themselves working on this project

02:00:57.660 | in my shoes, they would say, all right,

02:00:59.260 | what are all the different ways we can use this

02:01:01.620 | for machine learning?

02:01:02.460 | Actually, let me just definitely mention

02:01:04.860 | colleague at NIST, Mike Schneider.

02:01:06.620 | He's also very much interested,

02:01:09.300 | particularly in the superconducting side of things,

02:01:11.680 | using the incredible speed, power efficiency,

02:01:14.880 | also Ken Segal at Colgate, other people working

02:01:17.680 | on specifically the superconducting side of this

02:01:20.480 | for machine learning and deep feed-forward neural networks.

02:01:25.040 | There, the advantages are obvious.

02:01:27.360 | It's extremely fast.

02:01:28.760 | - Yeah, so that's less on the nature of intelligence

02:01:31.840 | and more on various characteristics of this hardware.

02:01:35.720 | - Right, yes, absolutely. - That you can use

02:01:37.040 | for the basic computations we know today.

02:01:39.440 | - Yeah, one of the things-- - And communication.

02:01:40.640 | - One of the things that Mike Schneider's working on

02:01:42.280 | right now is an image classifier.

02:01:44.640 | At a relatively small scale, I think he's targeting

02:01:47.040 | that nine pixel problem where you can have

02:01:48.800 | three different characters and you just,

02:01:52.340 | you put in a nine pixel image and you classify it

02:01:54.880 | as one of these three categories.

02:01:58.400 | And that's gonna be really interesting

02:02:00.120 | to see what happens there because if you can show

02:02:04.160 | that even at that scale, you just put these images in

02:02:06.940 | and you get it out and he thinks he can do it.

02:02:09.520 | I forgot if it's a nanosecond

02:02:11.160 | or some extremely fast classification time.

02:02:14.000 | It's probably less, it's probably 100 picoseconds

02:02:16.360 | or something.

02:02:17.400 | There you have challenges though

02:02:18.800 | because the Josephson junctions themselves,

02:02:21.560 | the electronic circuit is extremely power efficient.

02:02:24.520 | Some orders of magnitude for something more

02:02:26.840 | than a transistor doing the same thing.

02:02:29.240 | But when you have to cool it down to four Kelvin,

02:02:31.440 | you pay a huge overhead just for keeping it cold,

02:02:33.860 | even if it's not doing anything.

02:02:35.240 | So it has to work at large scale

02:02:40.240 | in order to overcome that power penalty,

02:02:43.680 | but that's possible.

02:02:45.120 | It's just, it's gonna have to get that performance.

02:02:48.000 | And this is sort of what you were asking about before

02:02:49.720 | is like how much better than silicon would it need to be?

02:02:52.760 | And the answer is, I don't know.

02:02:54.120 | I think if it's just overall better than silicon

02:02:57.240 | at a problem that a lot of people care about,

02:03:00.200 | maybe it's image classification,

02:03:02.640 | maybe it's face recognition,

02:03:03.960 | maybe it's monitoring credit transactions, I don't know.

02:03:07.540 | Then I think it will have a place.

02:03:09.000 | It's not gonna be in your cell phone,

02:03:10.440 | but it could be in your data center.

02:03:12.240 | - So what about in terms of the data center?

02:03:16.160 | I don't know if you're paying attention

02:03:17.680 | to the various systems like Tesla recently announced DOJO,

02:03:22.680 | which is a large scale machine learning training system.

02:03:27.160 | That again, the bottleneck there

02:03:28.940 | is probably going to be communication

02:03:30.920 | between those systems.

02:03:32.800 | Is there something from your work

02:03:34.880 | on everything we've been talking about

02:03:39.120 | in terms of superconductive hardware

02:03:41.640 | that could be useful there?

02:03:43.520 | - Oh, I mean, okay.

02:03:45.360 | Tomorrow, no.

02:03:46.720 | In the long term, it could be the whole thing.

02:03:49.000 | It could be nothing.

02:03:49.840 | I don't know, but definitely, definitely.

02:03:52.120 | When you look at the,

02:03:55.160 | so I don't know that much about DOJO.

02:03:56.720 | My understanding is that that's new, right?

02:03:58.820 | That's just coming online.

02:04:01.320 | - Well, I don't even know where.

02:04:03.940 | It hasn't come online.

02:04:06.960 | And when you announce big, sexy,

02:04:09.560 | so let me explain to you the way things work in--

02:04:12.040 | - In the world out there.

02:04:13.160 | - In the world of business and marketing,

02:04:15.680 | it's not always clear where you are

02:04:18.400 | on the coming online part of that.

02:04:20.560 | So I don't know where they are exactly,

02:04:22.680 | but the vision is from ground up to build up

02:04:26.440 | a very, very large scale modular machine learning ASIC,

02:04:31.080 | basically hardware that's optimized

02:04:32.640 | for training neural networks.

02:04:34.120 | And of course, there's a lot of companies

02:04:36.060 | that are small and big working on this kind of problem.

02:04:39.360 | The question is how to do it in a modular way

02:04:42.680 | that has very fast communication.

02:04:45.640 | The interesting aspect of Tesla is you have a company

02:04:49.520 | that, at least at this time, is so singularly focused

02:04:54.660 | on solving a particular machine learning problem

02:04:57.740 | and is making obviously a lot of money doing so

02:05:00.900 | because the machine learning problem

02:05:02.220 | happens to be involved with autonomous driving.

02:05:05.180 | And so you have a system that's driven by an application.

02:05:09.780 | And that's really interesting because you have maybe Google

02:05:14.780 | working on TPUs and so on.

02:05:17.740 | You have all these other companies with ASICs.

02:05:21.460 | They're usually more kind of always thinking general.

02:05:25.980 | So I like it when it's driven by a particular application

02:05:29.180 | because then you can really get to the,

02:05:32.220 | it's like, it's somehow,

02:05:34.940 | if you just talk broadly about intelligence,

02:05:37.300 | you may not always get to the right solutions.

02:05:40.300 | It's nice to couple that sometimes with specific,

02:05:42.820 | clear illustration of something

02:05:46.140 | that requires general intelligence,

02:05:47.680 | which for me, driving is one such case.

02:05:49.900 | - I think you're exactly right.

02:05:51.140 | Sometimes just having that focus on that application

02:05:54.420 | brings a lot of people focus

02:05:55.900 | as their energy and attention.

02:05:57.700 | I think that, so one of the things that's appealing

02:06:00.220 | about what you're saying is not just

02:06:02.460 | that the application is specific,

02:06:03.900 | but also that the scale is big.

02:06:05.700 | - Big.

02:06:06.540 | - And that the benefit is also huge.

02:06:10.660 | - Financial anti-humanity.

02:06:12.340 | - Right, right, right.

02:06:13.180 | Yeah, so I guess, let me just try to understand,

02:06:15.520 | is the point of this Dojo system

02:06:17.980 | to figure out the parameters

02:06:21.860 | that then plug into neural networks

02:06:23.860 | and then you don't need to retrain,

02:06:26.500 | you just make copies of a certain chip

02:06:29.060 | that has all the parameters established or?

02:06:32.320 | - No, it's straight up retraining a large neural network

02:06:36.740 | over and over and over.

02:06:38.580 | - So you have to do it once for every new car?

02:06:41.660 | - No, no, no, you have to,

02:06:43.120 | so they do this interesting process,

02:06:44.820 | which I think is a process for machine learning,

02:06:47.020 | supervised machine learning systems,

02:06:49.260 | you're going to have to do,

02:06:50.740 | which is you have a system,

02:06:53.700 | you train your network once,

02:06:55.140 | it takes a long time, I don't know how long,

02:06:57.140 | but maybe a week to train.

02:06:59.720 | And then you deploy it on,

02:07:03.180 | let's say about a million cars,

02:07:05.020 | I don't know what the number is.

02:07:05.860 | - But that part you just write software

02:07:07.820 | that updates some weights in a table and yeah, okay.

02:07:10.660 | - But there's a loop back.

02:07:12.620 | - Yeah, yeah, okay.

02:07:13.460 | - And each of those cars run into trouble rarely,

02:07:17.260 | but like they catch the edge cases

02:07:23.700 | of the performance of that particular system

02:07:26.060 | and then send that data back

02:07:28.300 | and either automatically or by humans

02:07:31.420 | that weird edge case data is annotated

02:07:34.740 | and then the network has to become smart enough

02:07:37.540 | to now be able to perform in those edge cases

02:07:40.140 | so it has to get retrained.

02:07:41.820 | There's clever ways of retraining different parts

02:07:43.980 | of that network, but for the most part,

02:07:46.260 | I think they prefer to retrain the entire thing.

02:07:49.300 | So you have this giant monster

02:07:51.340 | that kind of has to be retrained regularly.

02:07:54.900 | I think the vision with Dojo

02:07:58.580 | is to have a very large machine learning focused,

02:08:02.300 | driving focused supercomputer

02:08:05.180 | that then is sufficiently modular

02:08:07.580 | that could be scaled

02:08:08.460 | to other machine learning applications.

02:08:10.340 | But like, so they're not limiting themselves completely

02:08:12.740 | to this particular application,

02:08:13.980 | but is this application is the way

02:08:16.180 | they kind of test this iterative process

02:08:18.980 | of machine learning is you make a system

02:08:21.500 | that's very dumb, deploy it,

02:08:25.140 | get the edge cases where it fails,

02:08:27.180 | make it a little smarter,

02:08:28.340 | it becomes a little less dumb

02:08:29.980 | and that iterative process achieves something

02:08:32.980 | that you can call intelligent

02:08:34.540 | or is smart enough to be able

02:08:36.460 | to solve this particular application.

02:08:37.980 | So it has to do with training neural networks fast

02:08:42.980 | and training neural networks that are large.

02:08:45.940 | - But also based on an extraordinary amount

02:08:48.540 | of diverse input.

02:08:49.900 | - Data, yeah.

02:08:50.740 | - And that's one of the things,

02:08:51.900 | so this does seem like one of those spaces

02:08:54.540 | where the scale of superconducting optoelectronics,

02:08:58.940 | the way that, so when you talk about the weaknesses,

02:09:02.540 | like I said, okay, well, you have to cool it down.

02:09:04.140 | At this scale, that's fine.

02:09:05.780 | Because that's not too much of an added cost.

02:09:09.460 | Most of your power is being dissipated

02:09:10.980 | by the circuits themselves, not the cooling.

02:09:12.980 | And also you have one centralized kind of cognitive hub,

02:09:17.980 | if you will.

02:09:20.780 | And so when, if we're talking about putting

02:09:24.340 | a superconducting system in a car, that's questionable.

02:09:28.620 | Do you really want a cryostat in the trunk

02:09:30.660 | of every one of your car?

02:09:31.500 | It'll fit, it's not that big of a deal,

02:09:32.900 | but hopefully there's a better way, right?

02:09:35.700 | But since this is sort of a central supreme intelligence

02:09:39.060 | or something like that, and it needs to really have

02:09:42.600 | this massive data acquisition, massive data integration,

02:09:47.100 | I would think that that's where large-scale

02:09:49.140 | spiking neural networks with vast communication

02:09:51.260 | and all these things would have something

02:09:53.140 | pretty tremendous to offer.

02:09:54.260 | It's not gonna happen tomorrow.

02:09:55.740 | There's a lot of development that needs to be done.

02:09:58.240 | But we have to be patient with self-driving cars

02:10:01.420 | for a lot of reasons.

02:10:02.260 | We were all optimistic that they would be here by now.

02:10:04.580 | And okay, they are to some extent,

02:10:06.460 | but if we're thinking five or 10 years down the line,

02:10:09.540 | it's not unreasonable.

02:10:12.060 | One other thing, let me just mention

02:10:14.780 | that getting into self-driving cars and technologies

02:10:17.380 | that are using AI out in the world,

02:10:19.700 | this is something NIST cares a lot about.

02:10:21.520 | Elham Tabassi is leading up a much larger effort

02:10:25.020 | in AI at NIST than my little project.

02:10:29.320 | And really central to that mission

02:10:32.700 | is this concept of trustworthiness.

02:10:35.020 | So when you're going to deploy this neural network

02:10:38.900 | in every single automobile with so much on the line,

02:10:43.340 | you have to be able to trust that.

02:10:45.260 | So now how do we know that we can trust that?

02:10:48.020 | How do we know that we can trust the self-driving car

02:10:50.100 | or the supercomputer that trained it?

02:10:53.700 | There's a lot of work there

02:10:54.820 | and there's a lot of that going on at NIST.

02:10:56.980 | And it's still early days.

02:10:58.200 | I mean, you're familiar with the problem and all that.

02:11:01.820 | But there's a fascinating dance in engineering

02:11:04.460 | with like safety critical systems.

02:11:06.700 | There's a desire in computer science,

02:11:08.820 | just recently talked to Don Knuth,

02:11:10.620 | for algorithms and for systems,

02:11:14.800 | for them to be provably correct or provably safe.

02:11:17.440 | And this is one other difference

02:11:20.300 | between humans and biological systems

02:11:22.340 | is we're not provably anything.

02:11:24.860 | And so there's some aspect of imperfection

02:11:29.760 | that we need to have built in,

02:11:32.180 | like robustness to imperfection, be part of our systems,

02:11:37.180 | which is a difficult thing for engineers to contend with.

02:11:40.700 | They're very uncomfortable with the idea

02:11:42.820 | that you have to be okay with failure

02:11:46.900 | and almost engineer failure into the system.

02:11:49.740 | - Mathematicians hate it too.

02:11:51.020 | But I think it was Turing who said something

02:11:53.660 | along the lines of, I can give you an intelligent system

02:11:57.100 | or I can give you a flawless system,

02:11:59.540 | but I can't give you both.

02:12:00.900 | And it's in sort of creativity and abstract thinking

02:12:04.140 | seem to rely somewhat on stochasticity

02:12:07.860 | and not having components

02:12:11.360 | that perform exactly the same way every time.

02:12:13.780 | - This is where like the disagreement I have with,

02:12:16.020 | not disagreement, but a different view on the world.

02:12:18.580 | I'm with Turing.

02:12:19.700 | When I talk to robotic, robot colleagues,

02:12:24.620 | that sounds like I'm talking to robots,

02:12:26.460 | colleagues that are roboticists, the goal is perfection.

02:12:31.460 | And to me is like, no, I think the goal should be

02:12:36.660 | imperfection that's communicated.

02:12:40.980 | And through the interaction between humans and robots,

02:12:44.180 | that imperfection becomes a feature, not a bug.

02:12:49.180 | Like together as a scene, as a system,

02:12:52.180 | the human and the robot together

02:12:53.740 | are better than either of them individually,

02:12:56.420 | but the robot itself is not perfect in any way.

02:13:00.580 | Of course, there's a bunch of disagreements,

02:13:02.780 | including with Mr. Elon about,

02:13:04.940 | to me autonomous driving is fundamentally

02:13:08.700 | a human robot interaction problem, not a robotics problem.

02:13:12.380 | To Elon, it's a robotics problem.

02:13:14.340 | That's actually an open and fascinating question,

02:13:18.620 | whether humans can be removed from the loop completely.

02:13:22.300 | (inhales)

02:13:24.460 | We've talked about a lot of fascinating chemistry

02:13:27.700 | and physics and engineering,

02:13:31.260 | and we're always running up against this issue

02:13:33.700 | that nature seems to dictate what's easy and what's hard.

02:13:37.580 | So you have this cool little paper

02:13:40.100 | that I'd love to just ask you about.

02:13:42.760 | It's titled,

02:13:44.500 | "Does Cosmological Evolution Select for Technology?"

02:13:48.180 | So in physics, there's parameters

02:13:53.180 | that seem to define the way our universe works,

02:13:56.220 | that physics works, that if it worked any differently,

02:13:59.340 | we would get a very different world.

02:14:01.740 | So it seems like the parameters are very fine tuned

02:14:04.220 | to the kind of physics that we see.

02:14:06.460 | All the beautiful E equals MC squared,

02:14:08.580 | that we get these nice, beautiful laws.

02:14:10.420 | It seems like very fine tuned for that.

02:14:13.140 | So what you argue in this article is,

02:14:16.820 | it may be that the universe has also fine tuned

02:14:20.580 | its parameters that enable the kind

02:14:24.340 | of technological innovation that we see,

02:14:27.280 | that technology that we see.

02:14:29.580 | Can you explain this idea?

02:14:31.500 | - Yeah, I think you've introduced it nicely.

02:14:33.420 | Let me just try to say a few things in my language.

02:14:38.420 | Leah, what is this fine tuning problem?

02:14:41.660 | So physicists have spent centuries trying to understand

02:14:46.540 | the system of equations that govern the way nature behaves,

02:14:51.540 | the way particles move and interact with each other.

02:14:55.060 | And as that understanding has become more clear over time,

02:15:00.060 | it became sort of evident that it's all well adjusted

02:15:05.820 | to allow a universe like we see, very complex,

02:15:13.060 | this large long lived universe.

02:15:16.500 | And so one answer to that is, well, of course it is,

02:15:19.960 | because we wouldn't be here otherwise.

02:15:21.540 | But I don't know, that's not very satisfying.

02:15:24.420 | That's sort of, that's what's known

02:15:25.600 | as the weak anthropic principle.

02:15:27.260 | It's a statement of selection bias.

02:15:29.220 | We can only observe a universe that is fit for us to live in.

02:15:33.660 | So what does it mean for a universe

02:15:35.000 | to be fit for us to live in?

02:15:36.140 | Well, the pursuit of physics,

02:15:38.420 | it is based partially on coming up with equations

02:15:42.620 | that describe how things behave

02:15:44.660 | and interact with each other.

02:15:46.340 | But in all those equations you have,

02:15:48.520 | so there's the form of the equation,

02:15:49.980 | sort of how different fields or particles

02:15:54.240 | move in space and time.

02:15:56.500 | But then there are also the parameters

02:15:58.520 | that just tell you sort of the strength

02:16:01.140 | of different couplings, how strongly does a charged particle

02:16:05.200 | couple to the electromagnetic field or masses?

02:16:07.660 | How strongly does a particle couple to the Higgs field

02:16:11.540 | or something like that?

02:16:13.000 | And those parameters that define,

02:16:16.980 | not the general structure of the equations,

02:16:19.780 | but the relative importance of different terms,

02:16:23.620 | they seem to be every bit as important

02:16:25.260 | as the structure of the equations themselves.

02:16:27.780 | And so I forget who it was,

02:16:29.380 | somebody when they were working through this

02:16:31.200 | and trying to see, okay, if I adjust the parameter,

02:16:34.020 | this parameter over here, call it the,

02:16:35.260 | say the fine structure constant,

02:16:36.780 | which tells us the strength

02:16:37.940 | of the electromagnetic interaction.

02:16:40.420 | Oh boy, I can't change it very much,

02:16:42.300 | otherwise nothing works.

02:16:43.740 | The universe sort of doesn't,

02:16:45.380 | it just pops into existence and goes away

02:16:47.220 | in a nanosecond or something like that.

02:16:48.940 | And somebody had the phrase, this looks like a put-up job,

02:16:52.920 | meaning every one of these parameters was dialed in.

02:16:57.060 | It's arguable how precisely they have to be dialed in,

02:17:00.700 | but dialed in to some extent,

02:17:02.740 | not just in order to enable our existence,

02:17:05.360 | that's a very anthropocentric view,

02:17:07.120 | but to enable a universe like this one.

02:17:10.020 | So, okay, maybe I think the majority position

02:17:14.040 | of working physicists in the field is,

02:17:17.020 | it has to be that way in order for us to exist.

02:17:18.940 | We're here, we shouldn't be surprised

02:17:20.380 | that that's the way the universe is.

02:17:22.780 | And I don't know, for a while, that never sat well with me,

02:17:26.100 | but I just kind of moved on because there are things to do

02:17:29.440 | and a lot of exciting work.

02:17:31.140 | It doesn't depend on resolving this puzzle,

02:17:33.780 | but as I started working more with technology,

02:17:39.340 | getting into the more recent years of my career,

02:17:41.820 | particularly when I started,

02:17:43.740 | after having worked with silicon for a long time,

02:17:46.520 | which was kind of eerie on its own,

02:17:49.020 | but then when I switched over to superconductors,

02:17:51.180 | just like, this is crazy.

02:17:53.580 | It's just absolutely astonishing

02:17:57.380 | that our universe gives us superconductivity.

02:18:00.580 | It's one of the most beautiful physical phenomena,

02:18:02.500 | and it's also extraordinarily useful for technology.

02:18:06.580 | So you can argue that the universe

02:18:07.940 | has to have the parameters it does for us to exist

02:18:11.300 | because we couldn't be here otherwise.

02:18:13.020 | But why does it give us technology?

02:18:14.800 | Why does it give us silicon that has this ideal oxide

02:18:18.860 | that allows us to make a transistor

02:18:20.940 | without trying that hard?

02:18:22.500 | That can't be explained by the same anthropic reasoning.

02:18:26.760 | - Yeah, so it's asking the why question.

02:18:30.380 | I mean, a slight natural extension of that question is,

02:18:34.700 | I wonder if the parameters were different,

02:18:39.420 | if we would simply have just another set of paint brushes

02:18:44.240 | to create totally other things

02:18:46.900 | that wouldn't look like anything

02:18:49.220 | like the technology of today,

02:18:50.900 | but would nevertheless have incredible complexity,

02:18:54.540 | which is if you sort of zoom out and start defining things,

02:18:57.180 | not by how many batteries it needs

02:19:01.460 | and whether it can make toast,

02:19:03.580 | but more like how much complexity is within the system

02:19:06.460 | or something like that.

02:19:07.300 | - Well, yeah, you can start to quantify things.

02:19:10.060 | You're exactly right.

02:19:10.900 | So nowhere am I arguing that

02:19:13.740 | in all of the vast parameter space

02:19:15.860 | of everything that could conceivably exist

02:19:18.060 | in the multiverse of nature,

02:19:20.660 | there is this one point in parameter space

02:19:23.300 | where complexity arises.

02:19:25.180 | I doubt it.

02:19:26.660 | That would be a shameful waste of resources, it seems.

02:19:31.140 | But it might be that we reside at one place

02:19:34.700 | in parameter space that has been adapted

02:19:38.460 | through an evolutionary process

02:19:40.060 | to allow us to make certain technologies

02:19:43.460 | that allow our particular kind of universe to arise

02:19:47.100 | and sort of achieve the things it does.

02:19:49.780 | - See, I wonder if nature in this kind of discussion,

02:19:52.740 | if nature is a catalyst for innovation

02:19:55.720 | or if it's a ceiling for innovation.

02:19:57.700 | So like, is it going to always limit us,

02:20:00.780 | like you're talking about Silicon,

02:20:03.980 | is it just make it super easy to do awesome stuff

02:20:06.620 | in a certain dimension,

02:20:08.020 | but we could still do awesome stuff in other ways,

02:20:10.260 | it'll just be harder?

02:20:11.580 | Or does it really set the maximum we can do?

02:20:15.380 | - That's a good subject to discuss.

02:20:19.420 | I guess I feel like we need to lay

02:20:20.960 | a little bit more groundwork.

02:20:23.140 | So I want to make sure that

02:20:24.780 | I introduce this in the context

02:20:29.240 | of Lee Smolin's previous idea.

02:20:31.740 | - So who is Lee Smolin and what kind of ideas does he have?

02:20:35.620 | - Okay, Lee Smolin is a theoretical physicist

02:20:39.020 | who back in the late 1980s published a paper

02:20:42.420 | in the early 1990s introduced this idea

02:20:45.020 | of cosmological natural selection,

02:20:46.980 | which argues that the universe did evolve.

02:20:51.380 | So his paper was called "Did the Universe Evolve?"

02:20:54.460 | And I gave myself the liberty of titling my paper,

02:20:59.460 | "Does Cosmological Evolution Select for Technology?"

02:21:03.920 | in reference to that.

02:21:04.980 | So he introduced that idea decades ago,

02:21:08.180 | now he primarily works on quantum gravity,

02:21:12.160 | loop quantum gravity, other approaches to

02:21:14.620 | unifying quantum mechanics with general relativity,

02:21:19.280 | as you can read about in his most recent book, I believe,

02:21:22.340 | and he's been on your show as well.

02:21:24.280 | So, but I want to introduce this idea

02:21:27.780 | of cosmological natural selection,

02:21:29.380 | because I think that is one of the core ideas

02:21:32.640 | that could change our understanding

02:21:35.300 | of how the universe got here, our role in it,

02:21:37.800 | what technology is doing here.

02:21:39.840 | But there's a couple more pieces

02:21:41.260 | that need to be set up first.

02:21:42.340 | So the beginning of our universe

02:21:44.240 | is largely accepted to be the Big Bang.

02:21:47.280 | And what that means is if you look back in time

02:21:49.840 | by looking far away in space,

02:21:52.540 | you see that everything used to be at one point,

02:21:56.880 | and it expanded away from there.

02:21:58.820 | There was an era in the evolutionary process of our universe

02:22:03.820 | that was called inflation,

02:22:05.420 | and this idea was developed primarily by Alan Guth

02:22:08.820 | and others, Andre Linde and others in the '80s.

02:22:13.140 | And this idea of inflation is basically that

02:22:16.020 | when a singularity begins this process,

02:22:21.020 | begins this process of growth,

02:22:25.220 | there can be a temporary stage

02:22:27.580 | where it just accelerates incredibly rapidly.

02:22:30.860 | And based on quantum field theory,

02:22:33.740 | this tells us that this should produce matter

02:22:35.700 | in precisely the proportions that we find

02:22:37.840 | of hydrogen and helium in the Big Bang,

02:22:39.980 | lithium too, lithium also, and other things too.

02:22:44.820 | So the predictions that come out

02:22:46.540 | of Big Bang inflationary cosmology

02:22:48.500 | have stood up extremely well to empirical verification,

02:22:52.540 | the cosmic microwave background, things like this.

02:22:55.720 | So most scientists working in the field

02:22:59.540 | think that the origin of our universe is the Big Bang.

02:23:03.740 | And I base all my thinking on that as well.

02:23:08.040 | I'm just laying this out there so that people understand

02:23:11.420 | that where I'm coming from is an extension,

02:23:14.180 | not a replacement of existing well-founded ideas.

02:23:19.100 | In a paper, I believe it was 1986,

02:23:21.940 | with Alan Guth and another author, Farhi,

02:23:26.580 | they wrote that a Big Bang,

02:23:30.300 | I don't remember the exact quote,

02:23:31.540 | a Big Bang is inextricably linked with a black hole.

02:23:35.260 | This singularity that we call our origin

02:23:39.240 | is mathematically indistinguishable from a black hole.

02:23:42.020 | They're the same thing.

02:23:44.340 | And Lee Smolin based his thinking on that idea, I believe.

02:23:49.340 | I don't mean to speak for him, but this is my reading of it.

02:23:52.100 | So what Lee Smolin will say is that

02:23:54.980 | a black hole in one universe

02:23:57.780 | is a Big Bang in another universe.

02:24:00.700 | And this allows us to have progeny, offspring.

02:24:04.660 | So a universe can be said

02:24:07.900 | to have come before another universe.

02:24:10.500 | And very crucially, Smolin argues,

02:24:14.180 | I think this is potentially one of the great ideas

02:24:16.900 | of all time, that's my opinion,

02:24:18.620 | that when a black hole forms, it's not a classical entity,

02:24:22.380 | it's a quantum gravitational entity.

02:24:24.500 | So it is subject to the fluctuations

02:24:27.260 | that are inherent in quantum mechanics.

02:24:29.800 | The properties, what we're calling the parameters

02:24:35.440 | that describe the physics of that system

02:24:38.620 | are subject to slight mutations.

02:24:40.620 | So that the offspring universe

02:24:42.580 | does not have the exact same parameters

02:24:45.160 | defining its physics as its parent universe.

02:24:48.380 | They're close, but they're a little bit different.

02:24:50.420 | And so now you have a mechanism for evolution,

02:24:55.140 | for natural selection.

02:24:57.260 | - So there's mutations, so there's,

02:24:59.660 | and then if you think about the DNA of the universe

02:25:03.260 | are the basic parameters that govern its laws.

02:25:05.860 | - Exactly.

02:25:06.700 | So what Smolin said is our universe results

02:25:11.540 | from an evolutionary process that can be traced back

02:25:14.500 | some, he estimated, 200 million generations.

02:25:17.660 | Initially, there was something like a vacuum fluctuation

02:25:20.660 | that produced through random chance,

02:25:24.220 | a universe that was able to reproduce just once.

02:25:27.340 | So now it had one offspring.

02:25:28.860 | And then over time, it was able to make more and more

02:25:31.020 | until it evolved into a highly structured universe

02:25:35.300 | with a very long lifetime, with a great deal of complexity.

02:25:40.220 | And importantly, especially importantly for Lee Smolin,

02:25:44.140 | stars, stars make black holes.

02:25:47.220 | Therefore, we should expect our universe to be optimized,

02:25:50.940 | have its physical parameters optimized

02:25:53.220 | to make very large numbers of stars

02:25:55.660 | because that's how you make black holes

02:25:57.980 | and black holes make offspring.

02:25:59.500 | So we expect the physics of our universe to have evolved

02:26:03.780 | to maximize fecundity, the number of offspring.

02:26:06.900 | And the way Lee Smolin argues you do that

02:26:09.260 | is through stars that the biggest ones

02:26:11.900 | die in these core collapse supernova

02:26:13.460 | that make a black hole and a child.

02:26:15.740 | - Okay, first of all, I agree with you that

02:26:19.820 | this is back to our fractal view of everything

02:26:24.700 | from intelligence to our universe.

02:26:27.180 | That is very compelling and a very powerful idea

02:26:31.060 | that unites the origin of life

02:26:36.060 | and perhaps the origin of ideas and intelligence.

02:26:39.700 | So from a Dawkins perspective here on earth,

02:26:42.180 | the evolution of those and then the evolution

02:26:45.300 | of the laws of physics that led to us.

02:26:49.640 | I mean, it's beautiful.

02:26:52.300 | And then you stacking on top of that,

02:26:54.860 | that maybe we are one of the offspring.

02:26:57.460 | - Right, okay, so before getting into

02:27:01.060 | where I'd like to take that idea,

02:27:03.060 | let me just a little bit more groundwork.

02:27:05.140 | There is this concept of the multiverse

02:27:07.060 | and it can be confusing.

02:27:08.580 | Different people use the word multiverse in different ways.

02:27:11.500 | In the multiverse that I think is relevant to picture

02:27:16.900 | when trying to grasp Lee Smolin's idea,

02:27:20.820 | essentially every vacuum fluctuation

02:27:24.260 | can be referred to as a universe.

02:27:26.460 | It occurs, it borrows energy from the vacuum

02:27:28.860 | for some finite amount of time

02:27:30.200 | and it evanesces back into the quantum vacuum.

02:27:33.980 | And ideas of Guth before that and Andre Linde

02:27:38.980 | with eternal inflation aren't that different,

02:27:42.820 | that you would expect nature

02:27:44.820 | due to the quantum properties of the vacuum,

02:27:47.260 | which we know exist.

02:27:48.700 | They're measurable through things

02:27:50.300 | like the Casimir effect and others.

02:27:52.540 | You know that there are these fluctuations

02:27:54.540 | that are occurring.

02:27:55.420 | What Smolin is arguing is that

02:27:57.820 | there is this extensive multiverse that we,

02:28:00.760 | this universe, what we can measure and interact with

02:28:04.860 | is not unique in nature.

02:28:07.380 | It's just our residence.

02:28:09.660 | It's where we reside.

02:28:10.780 | And there are countless, potentially infinity,

02:28:13.600 | other universes, other entire evolutionary trajectories

02:28:17.260 | that have evolved into things

02:28:19.140 | like what you were mentioning a second ago

02:28:20.960 | with different parameters and different ways

02:28:24.060 | of achieving complexity and reproduction and all that stuff.

02:28:27.000 | So it's not that the evolutionary process

02:28:29.460 | is a funnel towards this end point, not at all.

02:28:34.220 | Just like the biological evolutionary process

02:28:36.980 | that has occurred within our universe

02:28:39.340 | is not a unique route toward achieving

02:28:42.820 | one specific chosen kind of species.

02:28:44.980 | No, we have extraordinary diversity around us.

02:28:49.180 | That's what evolution does.

02:28:50.500 | - And for any one species like us,

02:28:52.180 | it might feel like we're at the center of this process.

02:28:54.860 | We're the destination of this process,

02:28:57.140 | but we're just one of the many

02:28:59.460 | nearly infinite branches of this process.

02:29:02.220 | - And I suspect it is exactly infinite.

02:29:04.220 | I mean, I just can't understand how with this idea,

02:29:09.020 | you can ever draw a boundary around and say,

02:29:11.020 | no, the universe, I mean, the multiverse

02:29:13.460 | has 10 to the one quadrillion components,

02:29:17.680 | but not infinity.

02:29:18.860 | I don't know.

02:29:20.140 | - Well, yeah, I have cognitively incapable,

02:29:24.060 | as I think all of us are,

02:29:25.660 | in truly understanding the concept of infinity.

02:29:29.100 | - And the concept of nothing as well.

02:29:31.020 | - And nothing.

02:29:31.940 | But also the concept of a lot is pretty difficult.

02:29:35.440 | I run out of fingers at a certain point,

02:29:39.980 | and then you're screwed.

02:29:40.820 | - And when you're wearing shoes

02:29:41.820 | and you can't even get down to your toes, it's like.

02:29:44.460 | - It's like, all right, 1,000, fine, a million.

02:29:47.100 | Is that, what?

02:29:48.100 | And then it gets crazier and crazier.

02:29:50.100 | - Right, right.

02:29:51.740 | - So this particular, so when we say technology, by the way,

02:29:54.860 | I mean, there's some, not to over-romanticize the thing,

02:30:00.660 | but there is some aspect about this branch of ours

02:30:04.180 | that allows us to, for the universe to know itself.

02:30:08.060 | - Yes, yes.

02:30:08.980 | - So to have like little conscious cognitive fingers

02:30:13.980 | that are able to feel, like to scratch the head.

02:30:18.340 | - Right, right, right.

02:30:19.660 | - To be able to construct E equals MC squared

02:30:22.060 | and to introspect, to start to gain some understanding

02:30:25.660 | of the laws that govern it.

02:30:27.460 | Isn't that kind of amazing?

02:30:32.060 | Okay, I'm just human, but it feels like that,

02:30:35.620 | if I were to build a system that does this kind of thing,

02:30:39.260 | that evolves laws of physics, that evolves life,

02:30:42.100 | that evolves intelligence, that my goal would be

02:30:45.500 | to come up with things that are able

02:30:47.020 | to think about itself, right?

02:30:49.300 | Aren't we kind of close to the design specs,

02:30:53.380 | the destination?

02:30:54.700 | - We're pretty close, I don't know.

02:30:56.060 | I mean, I'm spending my career designing things

02:30:58.380 | that I hope will think about themselves,

02:30:59.780 | so that you and I aren't too far apart on that one.

02:31:02.700 | - But then maybe that problem is a lot harder

02:31:05.500 | than we imagine.

02:31:06.380 | Maybe we need to--

02:31:07.820 | - Let's not get too far, because I want to emphasize

02:31:10.840 | something that what you're saying is, isn't it fascinating

02:31:14.620 | that the universe evolved something that can be conscious,

02:31:18.660 | reflect on itself, but Lee Smolin's idea

02:31:22.420 | didn't take us there, remember?

02:31:23.780 | It took us to stars.

02:31:25.740 | Lee Smolin has argued, I think, right on almost

02:31:30.340 | every single way that cosmological natural selection

02:31:35.340 | could lead to a universe with rich structure,

02:31:38.780 | and he argued that the structure,

02:31:41.140 | the physics of our universe is designed

02:31:43.180 | to make a lot of stars so that they can make black holes.

02:31:46.140 | But that doesn't explain what we're doing here.

02:31:48.380 | In order for that to be an explanation of us,

02:31:51.380 | what you have to assume is that once you made that universe

02:31:55.380 | that was capable of producing stars, life, planets,

02:31:59.220 | all these other things, we're along for the ride.

02:32:01.300 | They got lucky.

02:32:02.180 | We're kind of arising, growing up in the cracks,

02:32:05.600 | but the universe isn't here for us.

02:32:07.220 | We're still kind of a fluke in that picture.

02:32:09.540 | And I don't necessarily have a philosophical opposition

02:32:14.060 | to that stance.

02:32:14.900 | It's just not...

02:32:17.220 | Okay, so I don't think it's complete.

02:32:20.220 | - So it seems like whatever we got going on here to you,

02:32:22.980 | it seems like whatever we have here on Earth

02:32:25.660 | seems like a thing you might want to select for

02:32:28.500 | in this whole big process.

02:32:29.700 | - Exactly.

02:32:30.540 | So if what you are truly,

02:32:32.100 | if your entire evolutionary process

02:32:34.900 | only cares about fecundity,

02:32:36.820 | it only cares about making offspring universes

02:32:39.800 | because then there's gonna be the most of them

02:32:41.780 | in that local region of hyperspace,

02:32:45.160 | which is the set of all possible universes, let's say.

02:32:49.360 | You don't care how those universes are made.

02:32:52.860 | You know they have to be made by black holes.

02:32:54.900 | This is what inflationary theory tells us.

02:32:57.980 | The Big Bang tells us that black holes make universes.

02:33:02.120 | But what if there was a technological means

02:33:04.260 | to make universes?

02:33:06.000 | Stars require a ton of matter

02:33:09.380 | because they're not thinking very carefully

02:33:11.700 | about how you make a black hole.

02:33:12.860 | They're just using gravity, you know?

02:33:15.080 | But if we devise technologies

02:33:19.260 | that can efficiently compress matter into a singularity,

02:33:23.380 | it turns out that if you can compress about 10 kilograms

02:33:26.300 | into a very small volume,

02:33:28.660 | that will make a black hole that is likely,

02:33:30.580 | highly probable to inflate into its own offspring universe.

02:33:34.820 | This is according to calculations done by other people

02:33:37.180 | who are professional quantum theorists,

02:33:38.820 | quantum field theorists,

02:33:40.260 | and I hope I am grasping what they're telling me correctly.

02:33:44.580 | I'm somewhat of a translator here.

02:33:47.620 | But so that's the position

02:33:50.460 | that is particularly intriguing to me,

02:33:52.700 | which is that what might have happened is that,

02:33:56.300 | okay, this particular branch on the vast tree of evolution,

02:34:01.160 | cosmological evolution that we're talking about,

02:34:03.240 | not biological evolution within our universe,

02:34:05.740 | but cosmological evolution,

02:34:07.780 | went through exactly the process

02:34:09.600 | that Lee Smolin described.

02:34:10.940 | Got to the stage where stars were making lots of black holes,

02:34:15.840 | but then continued to evolve and somehow bridge that gap

02:34:19.540 | and made intelligence,

02:34:21.220 | and intelligence capable of devising technologies

02:34:24.120 | because technologies,

02:34:25.680 | intelligent species working in conjunction with technologies

02:34:29.660 | could then produce even more--

02:34:32.140 | - Yeah, more efficiently, more faster and better

02:34:35.860 | and more different.

02:34:36.860 | Then you start to have different kind of mechanisms

02:34:38.820 | and mutation perhaps, all that kind of stuff.

02:34:40.820 | - And so if you do a simple calculation that says,

02:34:43.100 | all right, if I want to,

02:34:45.300 | we know roughly how many core collapse supernovae

02:34:50.300 | have resulted in black holes in our galaxy

02:34:54.460 | since the beginning of the universe,

02:34:55.980 | and it's something like a billion.

02:34:57.780 | So then you would have to estimate

02:35:00.620 | that it would be possible for a technological civilization

02:35:04.260 | to produce more than a billion black holes

02:35:07.420 | with the energy and matter at their disposal.

02:35:09.980 | And so one of the calculations in that paper,

02:35:12.660 | back of the envelope,

02:35:14.060 | but I think revealing nonetheless,

02:35:15.660 | is that if you take a relatively common asteroid,

02:35:20.660 | something that's about a kilometer in diameter,

02:35:23.660 | what I'm thinking of is just scrap material

02:35:26.700 | laying around in our solar system

02:35:29.020 | and break it up into 10 kilogram chunks

02:35:31.420 | and turn each of those into a universe,

02:35:33.360 | then you would have made at least a trillion black holes

02:35:38.140 | outpacing the star production rate

02:35:41.540 | by some three orders of magnitude.

02:35:43.380 | That's one asteroid.

02:35:44.820 | So now if you envision an intelligent species

02:35:46.880 | that would potentially have been devised initially by humans,

02:35:51.620 | but then based on superconducting

02:35:53.180 | up to electronic networks, no doubt,

02:35:55.220 | and they go out and populate,

02:35:57.140 | they don't have to fill the galaxy.

02:35:58.980 | They just have to get out to the asteroid belt.

02:36:01.700 | They could potentially dramatically outpace

02:36:05.280 | the rate at which stars are producing offspring universes.

02:36:07.880 | And then wouldn't you expect that

02:36:09.800 | that's where we came from instead of a star?

02:36:13.120 | - Yeah, so you have to somehow become masters of gravity?

02:36:16.560 | So like, or generate--

02:36:17.400 | - John, this is really gravity.

02:36:18.720 | So stars make black holes with gravity,

02:36:20.640 | but any force that can make the energy density

02:36:25.640 | can compactify matter to produce

02:36:28.320 | a great enough energy density can form a singularity.

02:36:31.140 | It doesn't, it would not likely be gravity.

02:36:33.520 | It's the weakest force.

02:36:34.520 | You're more likely to use something like

02:36:37.540 | the technologies that we're developing

02:36:39.200 | for fusion, for example.

02:36:40.500 | So I don't know, the Large Ignition Facility

02:36:44.240 | recently blasted a pellet with 100 really bright lasers

02:36:49.240 | and caused that to get dense enough

02:36:53.360 | to engage in nuclear fusion.

02:36:55.280 | So something more like that,

02:36:56.560 | or a tokamak with a really hot plasma.

02:36:58.720 | I'm not sure, something.

02:37:00.320 | I don't know exactly how it would be done.

02:37:02.140 | - I do like the idea that,

02:37:04.540 | especially just been reading a lot

02:37:05.900 | about gravitational waves and the fact that us humans

02:37:09.620 | with our technological capabilities,

02:37:11.720 | one of the most impressive technological accomplishments

02:37:15.820 | of human history is LIGO,

02:37:17.380 | being able to precisely detect gravitational waves.

02:37:20.700 | I'm particularly find appealing the idea

02:37:25.020 | that other alien civilizations from very far distances

02:37:29.540 | communicate with gravity, with gravitational waves.

02:37:34.340 | Because as you become greater and greater master of gravity,

02:37:37.780 | which seems way out of reach for us right now,

02:37:40.580 | maybe that seems like a effective way of sending signals,

02:37:44.060 | especially if your job is to manufacture black holes.

02:37:48.380 | - Right.

02:37:49.220 | - So let me ask there,

02:37:51.820 | whatever, I mean, broadly thinking,

02:37:56.420 | 'cause we tend to think other alien civilizations

02:37:58.900 | would be very human-like,

02:37:59.980 | but if we think of alien civilizations out there

02:38:04.020 | as basically generators of black holes,

02:38:06.600 | however they do it, 'cause they get stars,

02:38:10.000 | do you think there's a lot of them

02:38:12.740 | in our particular universe out there?

02:38:15.720 | - In our universe?

02:38:19.160 | - Well, okay, let me ask, okay, this is great.

02:38:23.340 | Let me ask a very generic question,

02:38:26.660 | and then let's see how you answer it,

02:38:29.180 | which is how many alien civilizations are out there?

02:38:34.080 | - If the hypothesis that I just described

02:38:38.380 | is on the right track, it would mean that

02:38:42.300 | the parameters of our universe have been selected

02:38:45.620 | so that intelligent civilizations

02:38:48.440 | will occur in sufficient numbers,

02:38:51.100 | so that if they reach something like

02:38:54.900 | supreme technological maturity,

02:38:56.460 | let's define that as the ability to produce black holes,

02:39:00.060 | then that's not a highly improbable event.

02:39:03.440 | It doesn't need to happen often,

02:39:05.420 | because as I just described,

02:39:06.660 | if you get one of them in a galaxy,

02:39:09.180 | you're gonna make more black holes

02:39:10.500 | than the stars in that galaxy.

02:39:12.700 | But there's also not a super strong motivation, well,

02:39:17.100 | it's not obvious that you need them

02:39:21.560 | to be ubiquitous throughout the galaxy.

02:39:24.060 | So one of the things that I try to emphasize in that paper

02:39:27.540 | is that given this idea of how our parameters

02:39:32.540 | might have been selected,

02:39:35.100 | it's clear that it's a series of trade-offs, right?

02:39:39.260 | If you make, I mean, in order for intelligent life

02:39:42.020 | of our variety or anything resembling us to occur,

02:39:45.740 | you need a bunch of stuff.

02:39:46.740 | You need stars, so that's right back

02:39:49.420 | to Smolin's roots of this idea,

02:39:51.060 | but you also need water to have certain properties.

02:39:54.460 | You need things like the rocky planets,

02:39:58.740 | like the Earth to be within the habitable zone,

02:40:00.620 | all these things that you start talking about

02:40:02.700 | in the field of astrobiology,

02:40:06.780 | trying to understand life in the universe,

02:40:08.820 | but you can't overemphasize,

02:40:10.500 | you can't tune the parameters so precisely

02:40:13.620 | to maximize the number of stars

02:40:15.180 | or to give water exactly the properties

02:40:18.980 | or to make rocky planets like Earth the most numerous.

02:40:22.260 | You have to compromise on all these things.

02:40:24.500 | And so I think the way to test this idea

02:40:27.380 | is to look at what parameters are necessary

02:40:30.260 | for each of these different subsystems,

02:40:32.780 | and I've laid out a few that I think are promising,

02:40:34.860 | there could be countless others,

02:40:36.660 | and see how changing the parameters

02:40:40.860 | makes it more or less likely that stars would form

02:40:43.580 | and have long lifetimes,

02:40:44.620 | or that rocky planets in the habitable zone

02:40:47.300 | are likely to form, all these different things.

02:40:49.380 | So we can test how much these things

02:40:52.180 | are in a tug of war with each other.

02:40:54.580 | And the prediction would be that we kind of sit

02:40:57.140 | at this central point where if you move the parameters

02:41:00.580 | too much, stars aren't stable,

02:41:03.140 | or life doesn't form, or technology's infeasible,

02:41:07.740 | 'cause life alone, at least the kind of life

02:41:10.780 | that we know of, cannot make black holes.

02:41:14.140 | We don't have the, well, I'm speaking for myself,

02:41:17.020 | you're a very fit, strong person,

02:41:18.540 | but it might be possible for you,

02:41:20.600 | but not for me to compress matter.

02:41:22.380 | So we need these technologies,

02:41:23.820 | but we don't know, we have not been able to quantify yet

02:41:28.820 | how finely adjusted the parameters would need to be

02:41:33.820 | in order for silicon to have the properties it does.

02:41:35.580 | Okay, this is not directly speaking to what you're saying,

02:41:37.820 | you're getting to the Fermi paradox,

02:41:39.500 | which is where are they, where are the life forms out there,

02:41:42.840 | how numerous are they, that sort of thing.

02:41:44.580 | What I'm trying to argue is that if this framework

02:41:49.100 | is on the right track, a potentially correct explanation

02:41:52.480 | for our existence, it doesn't necessarily predict

02:41:56.340 | that intelligent civilizations are just everywhere,

02:41:59.300 | because even if you just get one of them in a galaxy,

02:42:02.540 | which is quite rare, it could be enough

02:42:05.760 | to dramatically increase the fecundity

02:42:08.940 | of the universe as a whole.

02:42:10.140 | - Yeah, and I wonder, once you start generating

02:42:12.460 | the offspring for universes, black holes,

02:42:15.420 | how that has effect on the, what kind of effect does it have

02:42:19.940 | on the other candidates, civilizations within that universe?

02:42:24.940 | Maybe as a destructive aspect,

02:42:28.420 | or there could be some arguments about

02:42:30.140 | once you have a lot of offspring,

02:42:32.060 | that that just quickly accelerates

02:42:34.060 | to where the other ones can't even catch up.

02:42:35.800 | - It could, but I guess if you want me to put my chips

02:42:41.300 | on the table or whatever, I think I come down more

02:42:44.380 | on the side that intelligent life civilizations are rare.

02:42:49.380 | And I guess I follow Max Tegmark here.

02:42:57.140 | And also, there's a lot of papers coming out recently

02:43:00.980 | in the field of astrobiology that are seeming to say,

02:43:04.340 | all right, you just work through the numbers

02:43:05.980 | on some modified Drake equation or something like that.

02:43:09.660 | And it looks like it's not improbable.

02:43:12.940 | You shouldn't be surprised that an intelligent species

02:43:16.180 | has arisen in our galaxy.

02:43:17.940 | But if you think there's one the next solar system over,

02:43:20.260 | it's highly improbable.

02:43:21.500 | So I can see that the number,

02:43:23.860 | the probability of finding a civilization in a galaxy,

02:43:28.180 | maybe it's most likely that you're gonna find

02:43:30.980 | one to 100 or something.

02:43:32.780 | But okay, now it's really important

02:43:34.860 | to put a time window on that, I think.

02:43:36.600 | Because does that mean in the entire lifetime of the galaxy

02:43:40.700 | before it, so for in our case, before we run into Andromeda,

02:43:45.700 | I think it's highly probable, I shouldn't say I think,

02:43:52.940 | it's tempting to believe that it's highly probable

02:43:56.740 | that in that entire lifetime of your galaxy,

02:43:59.980 | you're gonna get at least one intelligent species,

02:44:02.540 | maybe thousands or something like that.

02:44:05.140 | But it's also, I think, a little bit naive

02:44:09.980 | to think that they're going to coincide in time

02:44:13.220 | and we'll be able to observe them.

02:44:14.980 | - And also, if you look at the span of life on Earth,

02:44:19.980 | the Earth history, it was surprising to me

02:44:24.460 | to kind of look at the amount of time,

02:44:27.860 | first of all, the short amount of time

02:44:29.700 | there's no life is surprising.

02:44:31.520 | Life spring up pretty quickly.

02:44:33.460 | - Cellular, single cell.

02:44:34.980 | - But that's the point I'm trying to make

02:44:36.980 | is so much of life on Earth

02:44:41.980 | was just single cell organisms.

02:44:44.780 | Most of it, most of it was boring bacteria type of stuff.

02:44:48.660 | - Well, bacteria are fascinating, but I take your point.

02:44:50.660 | - No, I get it.

02:44:51.500 | I mean, no offense to them.

02:44:52.900 | (laughing)

02:44:54.420 | This kind of speaking from the perspective of your paper

02:44:57.660 | of something that's able to generate technology

02:44:59.940 | as we kind of understand it,

02:45:01.440 | that's a very short moment in time

02:45:03.400 | relative to that full history of life on Earth.

02:45:08.400 | And maybe our universe is just saturated

02:45:12.220 | with bacteria like humans,

02:45:15.920 | but not the special extra AGI superhumans,

02:45:22.500 | that those are very rare.

02:45:25.560 | And once those spring up,

02:45:28.420 | everything just goes to like, it accelerates very quickly.

02:45:33.360 | - Yeah, we just don't have enough data to really say,

02:45:36.520 | but I find this whole subject extremely engaging.

02:45:40.000 | I mean, there's this concept,

02:45:41.680 | I think it's called the rare Earth hypothesis,

02:45:43.880 | which is that basically stating that,

02:45:46.920 | okay, microbes were here right away

02:45:49.040 | after the Hadean era where we were being bombarded.

02:45:52.140 | Well, after, yeah, bombarded by comets, asteroids,

02:45:54.920 | things like that, and also after the moon formed.

02:45:57.100 | So once things settled down a little bit,

02:45:59.560 | in a few hundred million years,

02:46:02.260 | you have microbes everywhere.

02:46:03.660 | And it could have been, we don't know exactly when,

02:46:05.320 | it could have been remarkably brief that that took.

02:46:08.040 | So it does indicate that,

02:46:09.880 | okay, life forms relatively easily.

02:46:12.120 | I think that alone is sort of a checker on the scale

02:46:15.880 | for the argument that the parameters

02:46:20.480 | that allow even microbial life to form

02:46:22.880 | are not just a fluke.

02:46:24.320 | But anyway, that aside, yes,

02:46:27.500 | then there was this long dormant period,

02:46:29.860 | not dormant, things were happening,

02:46:31.560 | but important things were happening

02:46:34.060 | for some two and a half billion years or something

02:46:37.260 | after the metabolic process that releases oxygen

02:46:41.620 | was developed.

02:46:42.780 | Then basically the planet's just sitting there

02:46:46.060 | getting more and more oxygenated,

02:46:47.500 | more and more oxygenated until it's enough

02:46:50.140 | that you can build these large, complex organisms.

02:46:54.140 | And so the rare earth hypothesis would argue

02:46:56.440 | that the microbes are common in everywhere,

02:47:00.960 | in any planet that's roughly in the habitable zone

02:47:04.120 | and has some water on it's probably gonna have those.

02:47:06.620 | But then getting to this Cambrian explosion

02:47:09.280 | that happened some between five and 600 million years ago,

02:47:13.720 | that's rare.

02:47:15.420 | And I buy that, I think that is rare.

02:47:19.040 | So if you say how much life is in our galaxy,

02:47:21.840 | I think that's probably the right answer

02:47:24.080 | is that microbes are everywhere.

02:47:26.380 | Cambrian explosion is extremely rare.

02:47:29.280 | And then, but the Cambrian explosion kind of went like that

02:47:32.700 | where within a couple tens or a hundred million years,

02:47:37.700 | all of these body plans came into existence.

02:47:40.920 | And basically all of the body plans

02:47:43.040 | that are now in existence on the planet

02:47:46.040 | were formed in that brief window.

02:47:48.680 | And we've just been shuffling around since then.

02:47:51.600 | So then what caused humans to pop out of that?

02:47:54.800 | I mean, that could be another extremely rare threshold

02:47:59.800 | that a planet roughly in the habitable zone with water

02:48:06.160 | is not guaranteed to cross.

02:48:08.360 | - To me, it's fascinating for being humble,

02:48:10.200 | like the humans cannot possibly be the most amazing thing

02:48:13.040 | that such, if you look at the entirety of the system,

02:48:15.840 | that at least small and you paint,

02:48:17.780 | that cannot possibly be the most amazing thing

02:48:20.040 | that process generates.

02:48:21.440 | So like, if you look at the evolution,

02:48:23.680 | what's the equivalent in the cosmological evolution

02:48:27.020 | and its selection for technology,

02:48:28.960 | the equivalent of the human eye or the human brain.

02:48:32.440 | Universes that are able to do some like,

02:48:35.580 | they don't need the damn stars.

02:48:37.760 | They're able to just do some incredible generation

02:48:42.080 | of complexity fast on scale, like much more than,

02:48:46.640 | if you think about it,

02:48:47.480 | it's like most of our universe is pretty freaking boring.

02:48:50.880 | There's not much going on.

02:48:51.920 | There's a few rocks flying around

02:48:53.360 | and there's some like apes that are just like doing podcasts

02:48:58.360 | on some weird planet.

02:49:00.480 | It just seems very inefficient.

02:49:02.840 | If you think about like the amazing thing in the human eye,

02:49:05.960 | the visual cortex can do, the brain, the nervous,

02:49:09.440 | everything that makes us more powerful

02:49:12.840 | than single cell organisms.

02:49:15.480 | Like if there's an equivalent of that for universes,

02:49:19.320 | like the richness of physics that could be expressed

02:49:24.320 | through a particular set of parameters.

02:49:26.840 | Like, I mean, like for me,

02:49:31.040 | I'm a, so from a computer science perspective,

02:49:33.760 | huge fan of cellular automata,

02:49:35.600 | which is a nice sort of pretty visual way to illustrate

02:49:40.400 | how different laws can result

02:49:42.780 | in drastically different levels of complexity.

02:49:46.400 | So like, it's like, yeah, okay.

02:49:48.960 | So we're all like celebrating,

02:49:50.280 | look, our little cellular automata

02:49:52.080 | is able to generate pretty triangles and squares

02:49:54.560 | and therefore we achieve general intelligence.

02:49:57.640 | And then there'll be like some bad-ass Chuck Norris type,

02:50:01.820 | like universal Turing machine type of cellular automata.

02:50:06.520 | They're able to generate other cellular automata

02:50:08.920 | and that does any arbitrary level of computation off the bat.

02:50:14.120 | Like those have to then exist.

02:50:16.440 | And then we're just like, we'll be forgotten is the story.

02:50:20.840 | This podcast just entertains a few other apes

02:50:24.880 | for a few months.

02:50:26.880 | - Well, I'm kind of surprised to hear your cynicism.

02:50:30.240 | - No, I'm very, I--

02:50:32.080 | - I usually think of you as like one who celebrates humanity

02:50:36.120 | in all its forms and things like that.

02:50:37.600 | And I guess I just, I don't,

02:50:39.240 | I see it the way you just described.

02:50:41.000 | I mean, okay, we've been here for 13.7 billion years

02:50:44.520 | and you're saying, gosh, that's a long time.

02:50:47.240 | Let's get on with the show already.

02:50:48.480 | Some other universe could have kicked our butt by now,

02:50:51.500 | but that's putting a characteristic time.

02:50:55.320 | I mean, why is 13.7 billion a long time?

02:50:58.360 | I mean, compared to what?

02:51:00.400 | I guess, so when I look at our universe,

02:51:02.280 | I see this extraordinary hierarchy

02:51:05.380 | that has developed over that time.

02:51:08.040 | So at the beginning, it was a chaotic mess

02:51:10.520 | of some plasma and nothing interesting going on there.

02:51:15.520 | And even for the first stars to form,

02:51:18.880 | that a lot of really interesting evolutionary processes

02:51:23.880 | had to occur by evolutionary in that sense,

02:51:26.240 | I just mean taking place over extended periods of time

02:51:30.520 | and structures are forming then.

02:51:32.020 | And then it took that first generation of stars

02:51:34.640 | in order to produce the metals

02:51:38.760 | that then can more efficiently produce

02:51:41.280 | another generation of stars.

02:51:42.440 | We're only the third generation of stars.

02:51:44.720 | So we might still be pretty quick to the game here.

02:51:48.040 | But I don't think, I don't, okay,

02:51:50.960 | so then you have these stars,

02:51:52.400 | now you have solar systems.

02:51:53.520 | On those solar systems, you have rocky worlds,

02:51:57.760 | you have gas giants, like all this complexity.

02:51:59.760 | And then you start getting life

02:52:01.080 | and the complexity that's evolved

02:52:03.760 | through the evolutionary process in life forms

02:52:06.400 | is just, it's not a letdown to me.

02:52:09.840 | Just some of it. - No, no, no.

02:52:10.680 | And there's some of it is like,

02:52:11.600 | some of the planets is like icy.

02:52:14.600 | It's like different flavors of ice cream.

02:52:16.080 | They're icy, but there might be water underneath.

02:52:18.560 | All kinds of life forms, some volcanoes.

02:52:21.120 | All kinds of weird stuff.

02:52:22.320 | No, no, no, I don't, I think it's beautiful.

02:52:24.600 | I think our life is beautiful.

02:52:25.960 | And I think it was designed that,

02:52:28.280 | by design, the scarcity of the whole thing.

02:52:31.220 | I think mortality, as terrifying as it is,

02:52:33.880 | is fundamental to the whole reason we enjoy everything.

02:52:37.280 | No, I think it's beautiful.

02:52:38.200 | I just think that all of us conscious beings,

02:52:42.400 | in the grand scheme of basically,

02:52:44.560 | at every scale, will be completely forgotten.

02:52:46.960 | - Well, that's true.

02:52:47.800 | I think everything is transient.

02:52:49.320 | And that would go back to maybe something more like Lao Tzu,

02:52:52.520 | the Tao Te Ching or something,

02:52:53.880 | where it's like, yes, there is nothing but change.

02:52:57.740 | There is nothing but emergence and dissolve,

02:52:59.680 | and that's it.

02:53:00.720 | But I just, in this picture of this hierarchy

02:53:03.760 | that's developed, I don't mean to say

02:53:05.520 | that now it gets to us and that's the pinnacle.

02:53:07.440 | In fact, I think, at a high level,

02:53:10.560 | the story I'm trying to tease out in my research

02:53:13.780 | is about, okay, well, so then what's the next level

02:53:16.280 | of hierarchy?

02:53:17.120 | And if it's, okay, we're kind of pretty smart.

02:53:21.560 | I mean, talking about people like Lee Small

02:53:23.880 | and Alan Guth, Max Tegmark, okay, we're really smart.

02:53:26.280 | Talking about me, okay, we're kind of,

02:53:28.520 | we can find our way to the grocery store or whatever.

02:53:30.760 | But-- - Sometimes.

02:53:31.960 | - But what's next?

02:53:32.800 | You know, I mean, what if there's another level

02:53:35.440 | of hierarchy that grows on top of us

02:53:37.780 | that is even more profoundly capable?

02:53:40.280 | And I mean, we've talked a lot

02:53:42.080 | about superconducting sensors.

02:53:43.560 | Imagine these cognitive systems far more capable than us

02:53:48.560 | residing somewhere else in the solar system

02:53:52.320 | off of the surface of the Earth,

02:53:53.640 | where it's much darker, much colder,

02:53:55.400 | much more naturally suited to them.

02:53:57.080 | And they have these sensors that can detect

02:53:59.360 | single photons of light from radio waves

02:54:02.640 | out to all across the spectrum to gamma rays

02:54:05.200 | and just see the whole universe.

02:54:07.380 | And they just live in space with these massive

02:54:09.960 | collection optics so that they, what do they do?

02:54:13.660 | They just look out and experience that vast array

02:54:18.660 | of what's being developed.

02:54:22.520 | - And if you're such a system,

02:54:24.240 | presumably you would do some things for fun.

02:54:28.920 | And the kind of fun thing I would do,

02:54:31.720 | as somebody who likes video games,

02:54:33.840 | is I would create and maintain

02:54:37.040 | and observe something like Earth.

02:54:39.320 | - Mm-hmm.

02:54:40.160 | - So in some sense, we're like all what players on a stage

02:54:47.320 | for this superconducting cold computing system out there.

02:54:52.320 | I mean, all of this is fascinating to think.

02:54:56.680 | The fact that you're actually designing systems

02:54:59.320 | here on Earth that are trying to push this technological,

02:55:01.560 | at the very cutting edge, and also thinking about

02:55:04.800 | how does the evolution of physical laws

02:55:09.760 | lead us to the way we are, is fascinating.

02:55:14.200 | That coupling is fascinating.

02:55:15.920 | It's like the ultimate rigorous application of philosophy

02:55:20.780 | to the rigorous application of engineering.

02:55:23.680 | So Jeff, you're one of the most fascinating,

02:55:26.360 | I'm so glad, I did not know much about you

02:55:29.000 | except through your work, and I'm so glad

02:55:31.480 | we got this chance to talk here.

02:55:35.000 | You're one of the best explainers

02:55:37.760 | of exceptionally difficult concepts,

02:55:39.800 | and you're also, speaking of like fractal,

02:55:44.560 | you're able to function intellectually

02:55:46.640 | at all levels of the stack, which I deeply appreciate.

02:55:50.200 | This was really fun.

02:55:51.600 | You're a great educator, a great scientist.

02:55:53.560 | It's an honor that you spend your valuable time with me.

02:55:57.240 | - It's an honor that you would spend your time with me.

02:56:00.120 | - Thanks, Jeff.

02:56:01.760 | Thanks for listening to this conversation

02:56:03.560 | with Jeff Shainlein.

02:56:05.240 | To support this podcast, please check out our sponsors

02:56:08.040 | in the description.

02:56:09.600 | And now, let me leave you with some words

02:56:12.060 | from the great John Carmack, who surely will be a guest

02:56:15.960 | on this podcast soon.

02:56:17.300 | Because of the nature of Moore's law,

02:56:20.100 | anything that an extremely clever graphics programmer

02:56:22.800 | can do at one point can be replicated

02:56:26.040 | by a merely competent programmer some number of years later.

02:56:29.380 | Thank you for listening, and hope to see you next time.

02:56:33.360 | (upbeat music)

02:56:35.940 | (upbeat music)

02:56:38.520 | [BLANK_AUDIO]

Jeffrey Shainline: Neuromorphic Computing and Optoelectronic Intelligence | Lex Fridman Podcast #225

Chapters