François Chollet: Scientific Progress is Not Exponential

00:00:00.000 | - What is your intuition why an intelligence explosion

00:00:05.000 | is not possible?

00:00:06.800 | Like taking the scientific,

00:00:08.400 | all the scientific revolutions,

00:00:11.500 | why can't we slightly accelerate that process?

00:00:16.360 | - So you can absolutely accelerate

00:00:19.480 | any problem-solving process.

00:00:21.420 | So recursive search improvement

00:00:25.000 | is absolutely a real thing.

00:00:26.920 | But what happens with recursively self-improving system

00:00:30.160 | is typically not explosion

00:00:31.980 | because no system exists in isolation.

00:00:34.800 | And so tweaking one part of the system

00:00:36.920 | means that suddenly another part of the system

00:00:39.160 | becomes a bottleneck.

00:00:40.480 | And if you look at science, for instance,

00:00:42.080 | which is clearly recursively self-improving,

00:00:45.100 | clearly a problem-solving system,

00:00:47.340 | scientific progress is not actually exploding.

00:00:50.280 | If you look at science,

00:00:51.800 | what you see is the picture of a system

00:00:54.760 | that is consuming an exponentially

00:00:57.080 | increasing amount of resources.

00:00:58.800 | But it's having a linear output

00:01:02.240 | in terms of scientific progress.

00:01:04.280 | And maybe that will seem like a very strong claim.

00:01:07.240 | Many people are actually saying that,

00:01:09.440 | scientific progress is exponential.

00:01:12.820 | But when they're claiming this,

00:01:14.400 | they're actually looking at indicators

00:01:16.680 | of resource consumptions,

00:01:19.600 | resource consumption by science.

00:01:21.360 | For instance, the number of papers being published,

00:01:24.920 | the number of patents being filed and so on,

00:01:28.220 | which are just completely correlated

00:01:31.840 | with how many people are working on science today.

00:01:36.720 | So it's actually an indicator of resource consumption.

00:01:38.920 | But what you should look at is the output,

00:01:41.440 | is progress in terms of the knowledge that science generates

00:01:46.320 | in terms of the scope and significance

00:01:48.920 | of the problems that we solve.

00:01:50.800 | And some people have actually been trying to measure that.

00:01:55.000 | Like Michael Nielsen, for instance.

00:01:58.400 | He had a very nice paper,

00:02:00.200 | I think that was last year about it.

00:02:02.000 | So his approach to measure scientific progress

00:02:06.600 | was to look at the timeline of scientific discoveries

00:02:11.600 | over the past, you know, 100, 150 years.

00:02:15.400 | And for each major discovery,

00:02:19.640 | ask a panel of experts to rate

00:02:22.640 | the significance of the discovery.

00:02:25.040 | And if the output of science as an institution

00:02:27.880 | were exponential, you would expect the temporal density

00:02:32.880 | of significance to go up exponentially,

00:02:36.400 | maybe because there's a faster rate of discoveries,

00:02:39.240 | maybe because the discoveries are, you know,

00:02:41.240 | increasingly more important.

00:02:43.200 | And what actually happens if you plot

00:02:46.120 | this temporal density of significance

00:02:48.320 | measured in this way,

00:02:49.600 | is that you see very much a flat graph.

00:02:52.800 | You see a flat graph across all disciplines,

00:02:54.880 | across physics, biology, medicine, and so on.

00:02:58.000 | And it actually makes a lot of sense if you think about it,

00:03:01.560 | because think about the progress of physics

00:03:04.280 | 110 years ago, right?

00:03:06.240 | It was a time of crazy change.

00:03:08.320 | Think about the progress of technology,

00:03:10.240 | you know, 170 years ago,

00:03:12.640 | when we started having, you know,

00:03:13.640 | replacing horses with cars,

00:03:15.840 | when we started having electricity and so on.

00:03:18.280 | It was a time of incredible change.

00:03:19.800 | And today is also a time of very, very fast change,

00:03:22.920 | but it would be an unfair characterization

00:03:26.360 | to say that today technology and science

00:03:28.880 | are moving way faster than they did 50 years ago,

00:03:31.240 | 100 years ago.

00:03:32.680 | And if you do try to

00:03:35.520 | rigorously plot the temporal density of

00:03:40.440 | the significance, yeah, of significance,

00:03:44.360 | of significance, sorry,

00:03:45.680 | you do see very flat curves.

00:03:48.040 | - That's fascinating.

00:03:48.880 | - And you can check out the paper that Michael Nielsen

00:03:52.080 | had about this idea.

00:03:54.280 | And so the way I interpret it is,

00:03:58.280 | as you make progress,

00:03:59.720 | you know, in a given field

00:04:02.440 | or in a given subfield of science,

00:04:04.360 | it becomes exponentially more difficult

00:04:06.920 | to make further progress.

00:04:08.680 | Like the very first person to work on information theory,

00:04:13.240 | if you enter a new field,

00:04:14.720 | and it's still the very early years,

00:04:16.160 | there's a lot of low hanging fruit you can pick.

00:04:19.400 | - That's right, yeah.

00:04:20.240 | - But the next generation of researchers

00:04:22.200 | is gonna have to dig much harder actually

00:04:25.560 | to make smaller discoveries,

00:04:28.440 | a probably larger number of smaller discoveries.

00:04:30.880 | And to achieve the same amount of impact,

00:04:32.880 | you're gonna need a much greater headcount.

00:04:35.760 | And that's exactly the picture you're seeing with science,

00:04:38.280 | is that the number of scientists and engineers

00:04:42.000 | is in fact increasing exponentially.

00:04:44.800 | The amount of computational resources

00:04:46.640 | that are available to science

00:04:48.280 | is increasing exponentially and so on.

00:04:50.120 | So the resource consumption of science is exponential,

00:04:53.840 | but the output in terms of progress,

00:04:56.440 | in terms of significance is linear.

00:04:59.240 | And the reason why is because,

00:05:01.320 | and even though science is recursively self-improving,

00:05:04.240 | meaning that scientific progress

00:05:06.680 | turns into technological progress,

00:05:08.440 | which in turn helps science.

00:05:11.160 | If you look at computers, for instance,

00:05:13.520 | our products of science and computers

00:05:16.720 | are tremendously useful in speeding up science.

00:05:19.760 | The internet, same thing.

00:05:20.960 | The internet is a technology that's made possible

00:05:22.920 | by very recent scientific advances.

00:05:25.680 | And itself, because it enables scientists to network,

00:05:30.640 | to communicate, to exchange papers and ideas much faster,

00:05:33.760 | it is a way to speed up scientific progress.

00:05:35.680 | So even though you're looking at

00:05:36.800 | a recursively self-improving system,

00:05:39.680 | it is consuming exponentially more resources

00:05:42.320 | to produce the same amounts of problem solving very much.

00:05:47.320 | - So that's a fascinating way to paint it.

00:05:49.360 | And certainly that holds for the deep learning community.

00:05:52.720 | Right?

00:05:53.560 | If you look at the temporal, what did you call it?

00:05:56.320 | The temporal density of significant ideas.

00:05:59.480 | If you look at in deep learning,

00:06:02.160 | I think, I'd have to think about that,

00:06:05.200 | but if you really look at significant ideas

00:06:07.280 | in deep learning, they might even be decreasing.

00:06:10.640 | - So I do believe the per paper significance is decreasing,

00:06:15.640 | but the amount of papers is still today

00:06:20.600 | exponentially increasing.

00:06:21.680 | So I think if you look at an aggregate,

00:06:24.120 | my guess is that you would see a linear progress.

00:06:27.760 | If you were to sum the significance of all papers,

00:06:32.080 | you would see a roughly linear progress.

00:06:36.880 | And in my opinion, it is not a coincidence

00:06:41.880 | that you're seeing linear progress in science,

00:06:44.080 | despite exponential resource consumption.

00:06:45.960 | I think the resource consumption

00:06:48.560 | is dynamically adjusting itself to maintain linear progress

00:06:53.560 | because we as a community expect linear progress,

00:06:56.800 | meaning that if we start investing less

00:06:59.520 | and seeing less progress,

00:07:00.600 | it means that suddenly there are some lower hanging fruits

00:07:04.000 | that become available

00:07:05.040 | and someone's gonna step up and pick them.

00:07:08.720 | So it's very much like a market for discoveries and ideas.

00:07:14.560 | - But there's another fundamental part

00:07:17.000 | which you're highlighting,

00:07:18.080 | which is the hypothesis that science

00:07:20.880 | or like the space of ideas,

00:07:23.400 | any one path you travel down,

00:07:26.440 | it gets exponentially more difficult

00:07:29.360 | to develop new ideas.

00:07:33.000 | In your sense, is that's gonna hold

00:07:35.880 | across our mysterious universe?

00:07:39.760 | - Yes, well, exponential progress

00:07:41.560 | triggers exponential friction.

00:07:43.720 | So that if you tweak one part of the system,

00:07:45.680 | suddenly some other part becomes a bottleneck.

00:07:48.040 | For instance, let's say you develop some device

00:07:53.120 | that measures its own acceleration

00:07:55.400 | and then it has some engine

00:07:56.920 | and it outputs even more acceleration

00:07:59.040 | in proportion of its own acceleration

00:08:00.600 | and you drop it somewhere.

00:08:01.560 | It's not gonna reach infinite speed

00:08:03.440 | because it exists in a certain context.

00:08:06.120 | So the air around it is gonna generate friction

00:08:09.240 | and it's gonna block it at some top speed.

00:08:12.520 | And even if you were to consider the broader context

00:08:15.680 | and lift the bottleneck there,

00:08:18.040 | like the bottleneck of friction,

00:08:20.480 | then some other part of the system

00:08:23.360 | would start stepping in

00:08:24.600 | and creating exponential friction,

00:08:26.360 | maybe the speed of flight or whatever.

00:08:28.160 | And this definitely holds true

00:08:30.160 | when you look at the problem solving algorithm

00:08:33.200 | that is being run by science as an institution,

00:08:36.400 | science as a system.

00:08:37.960 | As you make more and more progress,

00:08:40.280 | despite having this recursive self-improvement component,

00:08:44.080 | you are encountering exponential friction.

00:08:47.640 | Like the more researchers you have

00:08:49.800 | working on different ideas,

00:08:51.800 | the more overhead you have

00:08:53.200 | in terms of communication across researchers.

00:08:56.360 | If you look at, you were mentioning quantum mechanics,

00:09:00.360 | right?

00:09:01.200 | Well, if you want to start making significant discoveries

00:09:05.120 | today, significant progress in quantum mechanics,

00:09:07.920 | there is an amount of knowledge you have to ingest,

00:09:11.280 | which is huge.

00:09:12.360 | So there's a very large overhead

00:09:14.800 | to even start to contribute.

00:09:17.480 | There's a large amount of overhead

00:09:18.960 | to synchronize across researchers and so on.

00:09:22.320 | And of course, the significant practical experiments

00:09:26.560 | are going to require exponentially expensive equipment

00:09:30.120 | because the easier ones have already been run, right?

00:09:34.440 | - So in your senses, there's no way escaping,

00:09:39.760 | there's no way of escaping this kind of friction

00:09:43.640 | with artificial intelligence systems.

00:09:47.720 | - Yeah, no, I think science is a very good way

00:09:50.720 | to model what would happen with a superhuman

00:09:53.760 | or a recursive release of improving AI.

00:09:55.960 | - That's your sense, I mean, the-

00:09:58.000 | - That's my intuition.

00:09:59.160 | It's not like a mathematical proof of anything.

00:10:02.880 | That's not my point.

00:10:03.920 | Like I'm not trying to prove anything.

00:10:05.600 | I'm just trying to make an argument

00:10:07.040 | to question the narrative of intelligence explosion,

00:10:10.320 | which is quite a dominant narrative.

00:10:12.000 | And you do get a lot of pushback if you go against it.

00:10:14.960 | Because, so for many people, right,

00:10:18.400 | AI is not just a subfield of computer science.

00:10:21.320 | It's more like a belief system.

00:10:23.120 | Like this belief that the world is headed towards an event,

00:10:27.760 | the singularity, past which, you know,

00:10:30.960 | AI will become, will go exponential very much

00:10:36.240 | and the world will be transformed

00:10:37.720 | and humans will become obsolete.

00:10:40.080 | And if you go against this narrative,

00:10:43.000 | because it is not really a scientific argument,

00:10:46.000 | but more of a belief system,

00:10:48.120 | it is part of the identity of many people.

00:10:50.400 | If you go against this narrative,

00:10:51.720 | it's like you're attacking the identity

00:10:53.520 | of people who believe in it.

00:10:54.680 | It's almost like saying God doesn't exist or something.

00:10:57.360 | - Right.

00:10:58.480 | - So you do get a lot of pushback

00:11:01.160 | if you try to question these ideas.

00:11:02.920 | (upbeat music)

00:11:05.520 | (upbeat music)

00:11:08.120 | (upbeat music)

00:11:10.720 | (upbeat music)

00:11:13.320 | (upbeat music)

00:11:15.920 | (upbeat music)

00:11:18.520 | [BLANK_AUDIO]

François Chollet: Scientific Progress is Not Exponential | AI Podcast Clips

Chapters