Debunking The AI Reset: Alien Mind Fear, Chat GPT, Future of AI & Slow Productivity

00:00:00.000 | So there are a lot of concerns and excitements and confusions surrounding our current moment

00:00:07.280 | in artificial intelligence technology.

00:00:10.480 | Perhaps one of the most fundamental of these concerns is this idea that in our quest to

00:00:15.600 | train increasingly bigger and more capable AI systems that we might accidentally create

00:00:20.280 | something smarter than we expected.

00:00:23.760 | I want to address this particular concern from the many concerns surrounding AI in my

00:00:29.000 | capacity as a computer science professor and one of the founding members of Georgetown's

00:00:33.160 | Center for Digital Ethics.

00:00:34.720 | I have been thinking a lot from a sort of technological perspective about this idea

00:00:39.760 | of runaway or unexpected intelligence and AI systems.

00:00:42.840 | I have some new ideas I want to preview right here.

00:00:45.600 | These are in rough form, but I think they're interesting.

00:00:48.440 | And I hopefully will give you a new way and a more precise and hopefully comforting way

00:00:52.960 | of thinking about the possibility of AI getting smarter than we hope.

00:00:57.600 | All right, so where I want to start with here is the fear.

00:01:02.000 | Okay, so one way to think of the fear that I want to address is what I call the alien

00:01:06.640 | mind fear, that we are training these systems, most popularly as captured by large language

00:01:13.740 | model systems like the GPT family systems, and we don't know how they work.

00:01:18.400 | They're big.

00:01:19.400 | They sit in big data centers and do stuff for months and hundreds of millions of dollars

00:01:23.120 | of compute cycles.

00:01:24.120 | We get this thing afterwards and we engage with it and say, "What can this thing do now?"

00:01:29.160 | And so we are creating these minds.

00:01:30.680 | We don't understand how they're going to work.

00:01:33.360 | That's what sets up this fear of these minds getting too smart.

00:01:36.240 | I want to trace some of the origins of this particular fear.

00:01:39.200 | I'm going to load up on the screen here for people who are watching instead of just listening,

00:01:43.600 | an influential article along these lines.

00:01:46.220 | This comes from the New York Times in March of 2023.

00:01:49.120 | So this was pretty soon after CHAT-GPT made its big late 2022 debut.

00:01:54.560 | The title of this opinion piece is, "You can have the blue pill or the red pill, and we're

00:01:58.920 | out of blue pills."

00:02:00.520 | This is co-authored by Yuval Harari, who you know from his book, Sapiens, as well as Tristan

00:02:05.640 | Harris, who you know as the sort of whistleblower on social media who now runs a nonprofit dealing

00:02:12.840 | with the harms of technology, and Asa Raskin, who works with Tristan at his nonprofit.

00:02:18.600 | There's a particular quote.

00:02:19.660 | So this was essentially a call for we need to be worried about what we're building with

00:02:24.120 | these large language model systems like CHAT-GPT.

00:02:28.640 | There is a particular quote in here that I want to pull out, and I'll read this off of

00:02:33.760 | my page here.

00:02:34.960 | "We have summoned an alien intelligence.

00:02:38.360 | We don't know much about it, except that it is extremely powerful and offers us bedazzling

00:02:43.400 | gifts but could also hack the foundations of our civilization."

00:02:46.920 | So we get this alien terminology, this notion of we don't really know how this thing works,

00:02:52.480 | and so we don't really know what this thing might be capable.

00:02:56.240 | Let me give you another example of this thinking.

00:02:59.880 | This is an academic paper that was from this same time.

00:03:03.480 | This is April 2023.

00:03:06.120 | This is coming out of Microsoft Research.

00:03:08.600 | I wrote about this paper, actually, in a New Yorker piece that I wrote earlier this year

00:03:12.800 | about AI.

00:03:13.920 | But the title of this paper is important.

00:03:15.640 | This was very influential.

00:03:17.440 | The title of this influential paper is important.

00:03:20.320 | "Sparks of Artificial General Intelligence, Early Experiments with GPT-4."

00:03:27.280 | The whole structure of this paper is that these researchers, because they're at Microsoft,

00:03:31.440 | had early access to GPT-4.

00:03:33.240 | This was before it was publicly released, and they ran intelligence tests, sort of human

00:03:39.220 | intelligence tests that had been developed for humans.

00:03:41.040 | They were running these intelligence tests on GPT-4 and were really surprised by how

00:03:45.320 | well it did.

00:03:46.440 | So this sort of glimmers of AGI, this glimmers of artificial general intelligence, the sort

00:03:51.240 | of theme of this is, "My God, these things are smarter than we thought.

00:03:54.640 | They can do reasoning.

00:03:56.720 | These machines are becoming rapidly powerful."

00:03:59.000 | So it was sort of, "Hey, if you were worried about GPT-3.5, the original chat GPT language

00:04:05.080 | model that they were writing about in the New York Times op-ed, wait until you see what's

00:04:08.480 | coming next."

00:04:11.480 | There's a general rational extrapolation to make here.

00:04:16.600 | The original GPT worried people, like Yuval Harari.

00:04:21.440 | This new one, GPT-4, seemed even better.

00:04:25.520 | We keep extrapolating this curve, the GPT-5, GPT-6.

00:04:30.040 | It's going to keep getting more capable in ways that are unexpected and surprising.

00:04:35.360 | It's very rational to imagine this extrapolation bringing these alien minds to abilities where

00:04:40.920 | our safety is at stake.

00:04:43.000 | We're uncomfortable about how smart they are.

00:04:45.720 | They can do things we don't even really understand.

00:04:49.080 | This is this origin of this, "These things are going to get smarter than we hoped."

00:04:52.120 | I had a conversation with a friend of mine about this who's really interested in AI,

00:04:56.040 | has been reading a lot about it.

00:04:57.720 | The way he conceptualized this is he just said, "Look, we're going to keep building

00:05:00.680 | bigger models.

00:05:02.320 | One of these days, probably pretty soon, as we keep 10x-ing the size of these models,

00:05:06.000 | we are going to be very uncomfortable with what we build."

00:05:08.240 | All right, so that is our setup for the concern.

00:05:13.360 | To address this concern, the first thing I want to do is start with a strong but narrow

00:05:20.880 | observation.

00:05:23.000 | A large language model in isolation can never be understood to be a mind.

00:05:29.800 | All right, so let's be really clear.

00:05:31.520 | I'm being very precise about this, okay?

00:05:33.680 | And what I'm saying here is actually very narrow.

00:05:35.520 | If we actually take a specific large language model like GPT-4, that by itself, even if

00:05:42.480 | we make it bigger, even if we train it on many more things, cannot by itself be something

00:05:48.600 | that we imagine as an alien mind with which we have to contend like we might another mind.

00:05:56.480 | The reason is, in isolation, what a large language model does is it takes an input.

00:06:03.240 | This information moves forward through layers.

00:06:05.960 | It's fully feed-forward.

00:06:06.960 | And out of the other end comes a token, which is a part of a word.

00:06:11.120 | In reality, it's a probability distribution over tokens, but whatever, a part of a word

00:06:15.240 | comes out the other end.

00:06:17.440 | That's all a language model can do.

00:06:20.680 | Now how it generates what token to spit out next can have a huge amount of sophistication,

00:06:27.600 | right?

00:06:28.600 | It's difficult to come up with the proper analogies for describing this.

00:06:31.760 | But I think a somewhat reductive but useful way for understanding how these tokens are

00:06:36.920 | produced is the following analogy that I used in a New Yorker piece from a few months ago.

00:06:41.060 | You can imagine what happens is when you have your input, which is like the prompt or the

00:06:44.880 | prompt plus the part of the answer you've generated already, as this is going through

00:06:50.480 | the large language model, it can come up with candidates for like the next word or part

00:06:56.040 | of a word to output next, right?

00:06:57.600 | Like, okay, that's not too hard to do.

00:07:00.080 | This is known as Ngram prediction in some sense, except for here, it's a little bit

00:07:03.440 | more sophisticated because with self-attention, it can look at multiple parts of the input

00:07:07.260 | to figure out what to come next.

00:07:08.320 | But it's not too hard to be like, this is kind of the pool of grammatically correct,

00:07:13.320 | semantically correct next words that we could output.

00:07:16.280 | How do we figure out which of those things to output to actually match what's being asked

00:07:19.880 | or what's actually being discussed in the prompt?

00:07:21.760 | Well, this is where these models go through something like complex pattern recognition.

00:07:25.160 | I like to use the metaphor of a massive checklist, a checklist that has billions of possible

00:07:30.080 | properties on it.

00:07:31.840 | This is a discussion of chess.

00:07:34.760 | We're in the middle of producing moves for a chess game.

00:07:38.600 | This is like a middle of a chess game move that's being produced.

00:07:41.600 | This is a discussion of ancient history.

00:07:45.360 | This is a discussion of Rome.

00:07:46.840 | This is a discussion of buildings, right?

00:07:49.280 | Whatever.

00:07:50.280 | Huge checklist.

00:07:51.280 | We're sort of understanding as it goes to these recognizers, this is what we're trying,

00:07:55.800 | this is what we're in the middle of talking about.

00:07:57.640 | And then you can imagine, again, this is a rough analogy that you have these really complex

00:08:00.880 | rule books.

00:08:02.880 | It looks at the combination of different properties that describes what we're talking about.

00:08:06.360 | The rule books are combinatorial.

00:08:07.800 | They combine these properties to say, okay, given this combination of properties of what

00:08:11.600 | we're talking about, which of these possible correct, grammatically correct next word or

00:08:16.500 | tokens we could output, which of these makes the most sense, right?

00:08:20.400 | But then it's combining, okay, it's a chess game and here's the recent chess moves and

00:08:27.480 | we're supposed to be describing a middle game move.

00:08:29.720 | And the rules might say, these are legal moves given like this current situation.

00:08:35.800 | So of the different things we could output here that looks like the move in a chess game,

00:08:39.760 | these are actually legal moves and so let's choose one of these, right?

00:08:43.920 | So you have possible next words, you have checklist of properties, you have combinatorial

00:08:48.080 | combinations of those properties with rules that then help you influence which of these

00:08:52.080 | correct words to output next.

00:08:53.640 | And all of this sort of happens in this sort of feed forward manner.

00:08:56.160 | Those checklists and the rules in particular can be incredibly complicated.

00:09:00.840 | The rules can have novel combinations of properties.

00:09:05.280 | So combinations of properties that were never seen in the training data, but you have rules

00:09:09.180 | that just combine properties and that's how you can produce output with these models that

00:09:13.360 | don't directly match anything that ever solved before.

00:09:15.360 | So there's this nice generalization.

00:09:17.000 | This is all very sophisticated.

00:09:18.520 | This is all very impressive.

00:09:20.880 | But in the end, this is still, you can imagine it like a giant metal machine with dials and

00:09:26.440 | gears, and you're turning this big crank and hundreds of thousands of gears are all cranking

00:09:33.560 | and turning.

00:09:34.560 | And at the very end, at the far end of the machine, there's a dial of letters.

00:09:38.360 | These dials turn to spell out one word.

00:09:40.640 | Like in the end, that's what's happening.

00:09:41.760 | A word or a piece of the word is what comes out on the other side after you've turned

00:09:44.640 | these dials for a long time.

00:09:46.120 | It can be a very complicated apparatus, but in the end, what it does at the end is it

00:09:50.120 | can spit out a word or a piece of a word.

00:09:52.120 | All right.

00:09:53.120 | So it doesn't matter how big you make this thing.

00:09:56.160 | It can, it spits out parts of words, no matter how sophisticated its pattern recognizers

00:10:01.940 | and combinatorial rule generators, no matter how sophisticated these are, it's a word spitter

00:10:09.080 | router.

00:10:10.080 | Hey, it's Cal.

00:10:11.080 | I wanted to interrupt briefly to say that if you're enjoying this video, then you need

00:10:15.040 | to check out my new book, Slow Productivity, The Lost Art of Accomplishment Without Burnout.

00:10:22.500 | This is like the Bible for most of the ideas we talk about here in these videos.

00:10:27.920 | You can get a free excerpt at calnewport.com/slow.

00:10:33.280 | I know you're going to like it.

00:10:35.100 | Check it out.

00:10:36.100 | Now let's get back to the video.

00:10:37.480 | Okay.

00:10:38.480 | That's true.

00:10:39.480 | But where things get interesting, as people like to tell me when I talk to people, is

00:10:43.420 | when you begin to combine this really, really sophisticated word generator with control

00:10:52.320 | layers, something that sits outside of and works with the language model, that's really

00:10:58.040 | where everything interesting happens.

00:11:00.760 | Okay.

00:11:02.160 | This is what I want to better understand.

00:11:04.020 | It's better understanding the control logic that we place outside of the language models

00:11:09.320 | that we get a better understanding of the possible capabilities of artificial intelligence,

00:11:13.780 | because it's the combined system, language model plus control logic that becomes more

00:11:18.520 | interesting.

00:11:19.520 | Because what can control logic do?

00:11:21.520 | It can do two things.

00:11:22.520 | It chooses what to activate the model with, what input to give it, and it can then second

00:11:27.200 | actuate in the real world or the digital world based on what the model says.

00:11:31.440 | So it's the control logic that can put input into the model and then take the output of

00:11:35.640 | the model and actuate that, like take action.

00:11:38.680 | Do something on the internet, move a physical thing.

00:11:41.720 | So it's that control logic with its activation actuation capability that when combined with

00:11:45.700 | a language model, which again is just a word generator, that's when these systems begin

00:11:49.160 | to get interesting.

00:11:51.200 | So something I've been doing recently is sort of thinking about the evolution of control

00:11:58.040 | logic that can be appended to generative AI systems like large language models.

00:12:04.480 | And I want to go through like what we know right now.

00:12:07.400 | I'm going to draw this on the screen.

00:12:08.960 | For people who are watching instead of just listening, you can watch me draw this on the

00:12:12.920 | screen and see my beautiful handwriting.

00:12:15.240 | All right.

00:12:16.240 | So there's different layers to this.

00:12:18.240 | I'll actually draw this out.

00:12:19.340 | So we'll start with down here.

00:12:23.040 | I'm going to call this layer zero.

00:12:25.360 | Oh man, Jesse, my handwriting is only getting worse.

00:12:30.080 | People are like, "Oh my God, Cal's having a stroke."

00:12:33.280 | No, I just have really bad handwriting.

00:12:34.600 | All right.

00:12:35.600 | So layer zero control logic is actually what we got right away with the basic chatbots

00:12:40.980 | like ChatGPT.

00:12:42.660 | So I'm going to label this like, for example, ChatGPT.

00:12:51.640 | So level zero control logic basically just implements what's known as auto regression.

00:12:58.520 | So large language model spits out a single word or part of a word, but when you type

00:13:02.500 | a query into ChatGPT, you don't want just a one word answer.

00:13:06.280 | You want a whole response.

00:13:07.860 | So there's a basic what I'm calling layer zero control logic that takes your prompt,

00:13:14.280 | submits it to the underlying large language model, gets the answer of the language model,

00:13:18.800 | which is a single word or part of word that expands the input in a reasonable way.

00:13:23.480 | It appends it to the input.

00:13:25.320 | So now the input is your original prompt plus the first word of the answer.

00:13:29.880 | It then inputs fresh, fresh copy of the model, inputs a slightly longer input.

00:13:35.700 | It generates the next word of the answer.

00:13:37.480 | The control logic adds that and now submits the slightly longer input to the model.

00:13:43.180 | And it sort of keeps doing this until it judges this as a complete answer.

00:13:47.680 | And then it returns that answer to you, the user who are typing into the ChatGPT interface,

00:13:52.480 | right?

00:13:53.480 | That's called auto regression.

00:13:55.880 | That's how we just repeatedly keep using the same language model to get very long answers,

00:13:59.840 | right?

00:14:00.960 | So this is a control logic.

00:14:02.520 | The model by itself can just spit out one thing.

00:14:04.040 | We add some logic.

00:14:05.040 | Now we can spit out big answers.

00:14:07.140 | Another thing that we got in early versions and contemporary versions of chatbots is the

00:14:11.360 | other thing level layer zero control logic might do is append previous parts of your

00:14:18.440 | conversation to the prompt, right?

00:14:20.160 | So you know how when you're using ChatGPT or you're using Cloud or something like this

00:14:24.000 | or perplexity, you can sort of ask a follow-up question, right?

00:14:29.040 | So there's a little bit of control logic here where what it's really doing is it's not just

00:14:32.200 | submitting your follow-up question by itself to the language model.

00:14:35.400 | Remember, the language models have no memory.

00:14:36.880 | It's the exact same snapshot of this model, trained whenever it was trained, that's used

00:14:41.120 | for every word generated.

00:14:43.160 | What the control logic will do is take your follow-up question, but then also take all

00:14:47.440 | of the conversation before that and paste that whole thing into the input, right?

00:14:51.400 | So this is simple logic, but it makes the token generators useful.

00:14:55.440 | All right.

00:14:56.440 | So we already have some control logic and even the most basic generative AI tools.

00:15:01.520 | All right.

00:15:03.280 | Now let's go up to what I'm going to call layer one.

00:15:06.800 | All right.

00:15:08.800 | So with layer one, now we get two things we didn't have in layer zero.

00:15:14.240 | We're still taking input from a user, like you're typing some sort of prompt, but now

00:15:19.160 | we might get a substantial transformation of what you typed in for whatever is actually

00:15:27.920 | put into the language model.

00:15:28.920 | So what you type in might go through a substantial transformation by the control logic before

00:15:33.240 | it's passed on to the actual language model.

00:15:37.080 | The other key part of layer one is there's actuation.

00:15:42.640 | So it might also do some actions on behalf of you or the language model based on the

00:15:49.560 | output of language model, instead of just sending text back to the user, it might actually

00:15:53.160 | go and take some other action.

00:15:55.240 | All right.

00:15:56.880 | So an example of this, for example, would be the web enabled chatbots like Google's

00:16:03.760 | Gemini, right?

00:16:07.080 | So Google's Gemini, you can ask it a question where it's going to do a contemporary web

00:16:11.560 | search, like stuff that's on the internet now, not what it was trained with when they

00:16:15.040 | changed the original model, but it can actually look at stuff on the web and then give you

00:16:19.260 | an answer based on stuff that actually found contemporaneously on the web.

00:16:24.560 | This is layer one control logic.

00:16:25.880 | What's really happening here is when you ask something like Gemini or something like perplexity,

00:16:31.240 | a question about, you know, a current, a web search, an advanced web search, the control

00:16:35.680 | logic before the language model is ever involved, actually just goes and does a Google search

00:16:43.320 | and it finds like, these are relevant articles.

00:16:46.600 | It then takes the text of these articles and it puts it together into a really long prompt,

00:16:50.920 | which it then submits to the language model.

00:16:53.400 | I'm simplifying this, but this is basically what's going on.

00:16:56.040 | So the language model doesn't know about the specific articles necessarily in advance.

00:17:00.240 | It wasn't trained on them, but it gets a really long prompt that the prompt written by the

00:17:04.280 | control logic might say something like, please look at the following, you know, text that's

00:17:10.760 | pasted in this prompt and summarize from it, you know, an answer to the following question,

00:17:16.480 | which is then your original question.

00:17:17.840 | And then below it is, you know, 5,000 words of web results, right?

00:17:22.160 | So the prompt that's actually being submitted under the covers to the language model here

00:17:26.640 | is not what you typed in.

00:17:29.040 | It's a much bigger, substantially transformed prompt, right?

00:17:32.620 | We also see actuation.

00:17:34.660 | So if we consider like OpenAI's original plugin, you know, so these are these things you can

00:17:40.880 | turn on in GPT-4 that can do things, for example, like generate a picture for you or book airline

00:17:47.040 | flights or show you the schedules of airlines.

00:17:49.480 | You can talk to it about things.

00:17:51.820 | In the new Microsoft Copilot integrations, you can have the model take action on your

00:17:57.520 | behalf and tools like Microsoft Excel or in Microsoft Word.

00:18:01.620 | So there's actual action happening in the software world based on the model.

00:18:05.480 | This is also being done by the control logic, right?

00:18:09.300 | So you're saying like, help me find a flight to, you know, whatever, this place at this

00:18:15.020 | time.

00:18:16.020 | The control logic is going to, before we get to a language model, you know, it might make

00:18:20.780 | some queries of a flight booking service.

00:18:24.140 | Or what it might do is actually create a prompt to give to the language model and says, hey,

00:18:28.740 | please take this question about, you know, flight request and summarize it in the following

00:18:33.900 | format for me, which is like a very, you know, flight day destination.

00:18:38.280 | The language model then returns to the control logic a better, more consistently formatted

00:18:44.980 | version of the query you originally had.

00:18:47.800 | Now the control logic, which can understand this really well, format a request, talk over

00:18:52.460 | the internet to a flight booking service, get the results, and then it can pass those

00:18:57.020 | results to the language model and say, okay, take these flight results and please like

00:19:00.580 | write a summary of these in like a polite English.

00:19:03.660 | And then it returns that to you.

00:19:06.220 | And so what you see as the user is like, okay, I asked about flights and then I got back

00:19:10.300 | like a really nice response, like here's your various options for flights.

00:19:13.380 | And then maybe you say, hey, can you book this flight for me?

00:19:16.140 | The control logic takes that and say, hey, can you take this request from the user?

00:19:19.340 | And again, put it into this really precise format, you know, flight number, flight, whatever.

00:19:24.460 | And the language model does that.

00:19:25.780 | And now the control logic can take that and talk over the internet to the flight booking

00:19:29.340 | service and make the booking on your behalf.

00:19:31.840 | So this sort of actuation that happens in the sort of our current level of plugins.

00:19:36.180 | Same thing if you're asking co-pilot, Microsoft co-pilot to do something, build a table in

00:19:42.040 | Microsoft Word or something like this, it's taking your request.

00:19:45.500 | It's asking the language model to essentially reformat your request into something much

00:19:49.180 | more systematic and canonical, and then the control logic talks to Microsoft Word.

00:19:54.700 | These language models are just giant tables of numbers in a data warehouse somewhere being

00:19:58.820 | simulated on GPUs.

00:19:59.820 | They don't talk to Microsoft Word in your computer, the control logic does as well.

00:20:03.780 | So that's layer one control logic.

00:20:05.260 | So now we have substantial transformation of your prompts and some actuation on the

00:20:09.780 | responses.

00:20:10.780 | Okay.

00:20:11.780 | All right.

00:20:12.780 | So now we move up and things begin to get more interesting.

00:20:16.460 | Layer two is where the action is right now.

00:20:19.660 | I've been writing some about this for the New Yorker among other places.

00:20:23.860 | So in layer two, we now have the control logic able to keep state and make complex planning

00:20:32.620 | decisions.

00:20:33.700 | So it's going to be highly interactive with the language model, perhaps making many, many

00:20:37.820 | queries to the language model on route to trying to execute whatever the original request

00:20:42.460 | is.

00:20:43.460 | So this is where things start to get interesting.

00:20:46.140 | A less well-known, but illustrative example of this is that the meta put out this bot

00:20:55.700 | called Cicero, which I've talked about on the show before.

00:20:58.580 | Cicero can play the game diplomacy, the strategy war game diplomacy very well.

00:21:04.700 | The way Cicero works is we can actually think about it as a large language model combined

00:21:08.440 | with layer two control intelligence.

00:21:10.900 | So diplomacy is a board game, but it has lots of interpersonal negotiation with the other

00:21:15.060 | players.

00:21:16.620 | The way this diplomacy plane AI system works is the language model, the control logic will

00:21:22.320 | use the language model to take the conversations happening with the players and explain to

00:21:28.180 | the control program, the control logic in a very consistent systematic way, what's being

00:21:33.220 | proposed by the various players in a way that like the control program understands without

00:21:36.960 | having to be a natural language processor.

00:21:39.680 | Then the control program simulates lots of possible moves, but what if we did this, right?

00:21:44.260 | And what it's really doing here is simulating possibilities.

00:21:47.100 | If this person is lying, like they're trying to, but these two are honest and we do this,

00:21:51.660 | what would happen?

00:21:52.660 | Well, what if this person was lying, but they're being honest, which move would be best?

00:21:55.820 | What if they're all being honest?

00:21:56.820 | It kind of figures out all these possibilities for what's really happening to figure out

00:22:00.140 | what play gives it its best chance of being successful.

00:22:03.180 | And then it tells the language model, okay, here's what we want to do now.

00:22:06.740 | Please like talk to this player.

00:22:09.600 | Give me a message that's in this player that would be convincing to get them to do the

00:22:13.460 | action we want them to do.

00:22:14.460 | And the language model actually generates the text that then the control logic sends.

00:22:18.180 | So in Cicero, we have much more complicated control logic where now we're simulating moves,

00:22:23.020 | we're simulating the mind of other people.

00:22:25.440 | The logic might have multiple queries of the language model to actually implement a turn.

00:22:30.620 | We also see this in Devon.

00:22:33.200 | So Devon AI has been building these agent-based systems to do complicated computer programming

00:22:38.860 | tasks.

00:22:40.600 | And the way it works is you give a more complicated computer programming task to the Devon and

00:22:44.740 | it has control logic that's going to continually talk to a language model to generate code,

00:22:50.400 | but it can actually keep track of there's multiple steps to this task that we're trying

00:22:54.100 | to do.

00:22:55.140 | We're now on step two.

00:22:56.600 | We need code that does this.

00:22:58.060 | All right, let me get some code from the language model that we think does this.

00:23:01.960 | Let me test this code.

00:23:03.220 | Does it actually do this?

00:23:04.460 | Okay, great.

00:23:05.460 | Now we're on the step two of this task.

00:23:07.500 | Okay, we need code that integrates this into this system.

00:23:10.440 | Let me ask the language model for that code.

00:23:12.580 | So again, it's keeping track of a complex plan, the control logic, and using a language

00:23:17.100 | model as the actual production of a specific code that solves a specific request.

00:23:21.860 | A language model can't keep track of a long-term plan like this.

00:23:25.100 | It can't simulate novel futures because again, it's just a token generator.

00:23:29.220 | The control logic can't.

00:23:30.220 | So that's layer two.

00:23:31.220 | This is where a lot of the energy is in AI right now, is these sort of complex control

00:23:34.680 | layers.

00:23:35.680 | The layer that doesn't exist yet, but this is the layer that we speculate about, I call

00:23:39.620 | it layer three.

00:23:41.220 | And this is where we get closer to something like a general intelligence.

00:23:46.220 | So I'll put AGI here.

00:23:49.460 | And this is where, and I'm going to put a question mark, it's unclear exactly how close

00:23:52.380 | we can get to this.

00:23:53.380 | But now we have a very complicated, this is hypothetical, we'd have a very complicated

00:23:57.220 | control logic that keeps track of intention and state and understanding of the world.

00:24:02.300 | It might be interacting with many different generative models and recognizers.

00:24:06.480 | So it has a language model to help understand the world of language and produce texts, but

00:24:10.260 | it might have other types of models as well.

00:24:13.380 | If this was a fully actuated, like robotic artificial general intelligence, you would

00:24:17.140 | have something like visual recognizers that really can do a good job of saying, here's

00:24:21.940 | what we're seeing in the world around us.

00:24:24.420 | It might have some sort of like social intention recognizer where just trained to take recent

00:24:31.660 | conversations and try to understand what people's intent are.

00:24:35.180 | And then you have all of this being orchestrated by some master control logic that's trying

00:24:39.020 | to keep a sort of stateful existence and interaction in the world of some sort of simulated agent.

00:24:44.380 | So that's how you get to something like artificial general intelligence.

00:24:48.220 | So, here's the critical observation.

00:24:52.140 | In all of these layers, the control logic is not self-trained.

00:24:58.020 | The control logic, unlike a language model, is not something where we just turn on the

00:25:02.060 | switch and it looks at a lot of data and trains itself and then we have to say, how does this

00:25:06.020 | thing work?

00:25:07.020 | I don't know.

00:25:08.220 | At least in the layers that exist so far, layers two through layer zero, the control

00:25:13.460 | logics are hand-coded by humans.

00:25:16.680 | We know exactly what they do, right?

00:25:20.100 | Here's what's something interesting about Cicero.

00:25:22.260 | In the game Diplomacy, one of the big strategies that's common is lying, right?

00:25:28.260 | You make an alliance with another player, but you're backstabbing them and you have

00:25:31.140 | a secret alliance with another player.

00:25:33.540 | That is very common.

00:25:35.500 | The developers of Cicero were uncomfortable with having their computer program lie to

00:25:39.260 | real people.

00:25:40.260 | So they said, okay, though other people are doing that, our player, Cicero, will not lie.

00:25:46.420 | That's really easy to do because the control logic where that simulates moves, this is

00:25:49.900 | not some emergent thing they don't understand.

00:25:51.420 | They coded it themselves.

00:25:52.500 | It's a simulator that simulates moves.

00:25:54.420 | They just don't consider moves with lies.

00:25:57.980 | So we have this reality about the control plus generative AI combination.

00:26:04.300 | We have this reality that at least so far, the control is just hand-coded by people to

00:26:12.060 | do what we want it to do.

00:26:15.540 | There is no way for the intelligence in these cases of the language model, no matter how

00:26:19.780 | sophisticated its checklist and rules get at being able to produce tokens using very,

00:26:24.060 | very sophisticated digital contemplations, that cannot control the control logic.

00:26:31.140 | It can't break through and control the control logic.

00:26:33.420 | It can just generate tokens.

00:26:34.860 | The control logic we build.

00:26:38.260 | We don't want to lie.

00:26:39.260 | It doesn't want to lie.

00:26:40.260 | We don't want it to produce versions of us that are smarter.

00:26:41.820 | We just don't have that coded into the control logic.

00:26:44.660 | It's actually relatively straightforward.

00:26:46.540 | We have this with plugins.

00:26:47.860 | The plugins, there's a lot of control over these things of like, okay, we have gotten

00:26:52.740 | a request.

00:26:53.740 | We've asked for a formatted request, the book of flight from the LLM.

00:26:58.220 | Let's just look at this because we're not going to spend more than this much money and

00:27:01.340 | we're not going to fly to places that aren't on this list we think are appropriate places

00:27:05.580 | to fly or whatever it is.

00:27:07.740 | The control logic is just programmed right there.

00:27:09.560 | So I think we've extrapolated the emergent, hard to interpret reality of generative models

00:27:16.280 | to these full systems.

00:27:18.280 | But the control logic in these systems right now is not at all difficult to understand

00:27:24.080 | because we're creating them.

00:27:26.960 | All right.

00:27:29.240 | There's a couple of caveats here.

00:27:31.240 | One, this doesn't mean that we have nothing to be practically concerned about, but the

00:27:35.840 | biggest practical concern, especially about layer two or below artificial intelligence

00:27:40.440 | systems of this architecture is exceptions, right?

00:27:44.760 | Our control logic didn't think to worry about a particular opportunity.

00:27:49.720 | We didn't put the right checks and something that is like practically damaging happens.

00:27:56.400 | What do I mean by that?

00:27:57.680 | Well, for example, we're doing flight booking and our control logic doesn't have a check

00:28:02.160 | that says make sure the flight doesn't cost more than X and don't book it if it costs

00:28:06.800 | more than that.

00:28:07.800 | We forgot to put that check in and the LLM gives us a first class flight on Emirates

00:28:13.160 | that costs $20,000 or something.

00:28:14.880 | It's like, whoops, we spent a lot of money, right?

00:28:17.160 | Or we have a Devon type setup where it's giving us a program to run and we don't have a check

00:28:23.920 | that says make sure that it doesn't use more than its computational resources and that

00:28:28.320 | program actually is like a giant resource consuming infinite loop and it uses a hundred

00:28:32.800 | thousand dollars of Amazon cloud time before anyone realizes like what's going on here,

00:28:37.160 | right?

00:28:38.160 | So that's certainly a problem.

00:28:39.160 | Like your control logic doesn't check for the right things.

00:28:41.320 | You can have excessive behaviors, sure, but that's a very different thing than the system

00:28:46.480 | itself is somehow smarter than we expected or in taking intentional actions that we don't

00:28:51.760 | expect.

00:28:53.160 | So that we need to be worried about.

00:28:56.760 | Yeah, too, in theory, when we get to layer two, these really complicated control layers,

00:29:02.760 | in theory, one could imagine hand coding control logic that we completely understand that is

00:29:09.760 | working with LLMs to produce computer code for a better control logic.

00:29:16.460 | And it may be then you could get this sort of runaway superintelligence scenario of Nick

00:29:20.340 | Bostrom.

00:29:21.340 | But here's the thing, A, we're nowhere close to being knowing how to do that, how to write

00:29:25.740 | a control program that can talk to a coding machine like LLMs and get a better version

00:29:30.040 | of the control program.

00:29:31.040 | There's a lot of CS to be done there that quite frankly, no one's really working on.

00:29:34.740 | And two, there's no reason to do that.

00:29:36.460 | That won't accidentally happen.

00:29:39.020 | You would have to build a system to do that and then to start executing the new program.

00:29:44.980 | And so maybe we just don't build those types of systems.

00:29:47.240 | I call this whole way of thinking about things and I'll write this on here.

00:29:51.260 | I call this whole way of thinking about things, I'll use a different color text here, IAI,

00:29:58.220 | right, lowercase I capital AI for intentional artificial intelligence.

00:30:03.300 | The idea being that there can be tons of intention in the control logic, even if we can't interpret

00:30:07.580 | very well the generative models like the language models that these control logics use.

00:30:12.540 | And we should really lean into the control we have in the control logics to make sure

00:30:17.460 | that this is how we keep sort of predictability on what these systems actually do.

00:30:23.300 | There might actually be a legislative implication here, one way or the other, making sure that

00:30:28.340 | we do not develop a legal doctrine that says AI systems are unpredictable.

00:30:32.340 | So it's not your fault as the developer of an AI system for what it does once actuated.

00:30:36.300 | We say it is, you're liable.

00:30:39.100 | That would put a lot of emphasis on this control layers.

00:30:41.780 | We really want to be careful here.

00:30:43.860 | And exactly what we put in these control layers matter, especially once there's actuation,

00:30:48.700 | this is on us.

00:30:50.380 | And so we got to be really careful.

00:30:52.020 | The language model can be as smart as we want, but we're gonna be very careful on the actions

00:30:56.380 | that our control logic is willing to take on its behalf.

00:30:58.900 | Anyways, this is super nerdy.

00:31:01.140 | But I do think it is interesting.

00:31:02.540 | And I do think it is important that we separate the emergent, hard to predict, uninterpretable

00:31:07.420 | intelligence of self trained generative models, we separate that from the control logics that

00:31:11.540 | use them.

00:31:12.540 | The control logics aren't that complicated.

00:31:14.180 | We are building them.

00:31:16.200 | This is where the actuation happens.

00:31:17.900 | This is where the activation happens.

00:31:19.460 | If we go back to our analogy of the giant machine, the Babbage style machine of meshing

00:31:23.620 | gears and dials, that when you turn it, great sophistication happens inside the machine.

00:31:28.740 | And at the very end, a word comes out on these dials on the other end of this massive city

00:31:33.600 | block size machine.

00:31:35.820 | We're not afraid of a machine like that in that analogy.

00:31:39.080 | We do worry about what the people who are running the machine do with it.

00:31:42.320 | So that's where we should keep our focus is the people who are actually running the machine,

00:31:47.160 | you know, what they do should be constrained.

00:31:50.160 | Don't let them spend money without constraint.

00:31:53.160 | Don't let them fire missiles without constraint.

00:31:56.240 | Don't let the control logic have full access to all computational resources.

00:32:01.120 | Don't let the control logic be able to install an improved version automatically of its own

00:32:06.080 | control logic.

00:32:07.080 | We code the control logic, we can tell it what to do and what not to do.

00:32:10.040 | And let's just make it clear, whatever people do with this big system, like you are a liable,

00:32:15.980 | the whole systems you build, you're liable for it.

00:32:17.640 | So you'll be very careful about who you let in in this metaphor to actually turn those

00:32:20.800 | cranks and take action on the other end.

00:32:22.440 | So that's IAI, that's intentional AI.

00:32:24.160 | This is early thinking, just putting it out there for comment, but hopefully it diffuses

00:32:28.420 | a little bit of the sort of incipient idea that GPT-6 or 7 is going to become HAL.

00:32:33.580 | That's not the way this actually works.

00:32:34.840 | What do you think, Jesse, is that sufficiently nerdy?

00:32:37.560 | That was solid.

00:32:38.840 | For our return to highly technical topics?

00:32:41.840 | What do you think the comments will be for those that think the other way, that don't

00:32:45.640 | necessarily agree with you?

00:32:47.800 | It's interesting, you know, when I first pointed out in my article last year, the language

00:32:52.600 | model is just a feed-forward network.

00:32:54.640 | It has no state, it has no recursion, it has no interactivity.

00:32:57.680 | All it can do is generate a token.

00:32:59.200 | So this is not a mind in any sort of self-aware way.

00:33:01.920 | What a lot of people came back to me with is like, yeah, yeah, but it's, they were talking

00:33:05.800 | about, back then they were talking about auto-GPT, which was one of these very early, very early

00:33:10.080 | layer-2 control logics.

00:33:11.280 | Yeah, but people are writing programs that sort of keep track of things outside of the

00:33:16.120 | language model, and they talk back to the language model, and that's where the sophistication

00:33:19.560 | is going to come out.

00:33:20.600 | So in some sense, I'm reacting.

00:33:23.280 | Look at this.

00:33:24.280 | By the way, I'm looking at our screen here.

00:33:25.280 | Let's correct this.

00:33:26.280 | Look what I did.

00:33:27.280 | That should be a three.

00:33:29.560 | I wrote layer-2 twice.

00:33:31.680 | Sorry, for those who are listening, I realized that for all the precision of my speech,

00:33:35.600 | I wrote three, I wrote two instead of three.

00:33:40.720 | So I, you know, I think that diffuses that.

00:33:44.480 | I think some of the more just philosophical thinkers who just sort of tackle these issues

00:33:47.920 | of like superintelligence from an abstract perspective, like an abstract logical perspective,

00:33:52.640 | I think their main response will be like, yeah, but all it takes is one person to write

00:33:56.160 | layer-3 control logic that says write control logic program and then install it, replace

00:34:02.760 | myself with that program, and like that's what could allow sort of like runaway whatever.

00:34:06.880 | But I think that's a very hard problem.

00:34:08.560 | We don't know how to write a control program.

00:34:10.160 | If we think of the language model like a coder, we can tell it to write code that does something.

00:34:14.400 | Very constrained, but we can write this function, write that function.

00:34:18.280 | That's a very hard problem to sort of work with a language model to produce a different

00:34:24.460 | type of control program, right?

00:34:27.440 | It's a hard problem, and there's no reason to write that program, and I think it's not

00:34:30.620 | just one, you could, again, it's just a very hard problem.

00:34:35.360 | We don't even know if it's possible to write a significantly smarter control program, or,

00:34:42.560 | you know, the control program is limited by the intelligence of what the language model

00:34:47.040 | can produce.

00:34:48.040 | We don't have any great reason to believe that a language model trained on a bunch of

00:34:53.440 | existing code, and what it does is predict code that matches the type of things it can

00:34:58.440 | see, can produce code that is somehow better than any code a human has ever produced.

00:35:04.440 | We don't know that a language model can do that.

00:35:06.440 | What it does is it's been trained to try to expand text based on the structures it's seen

00:35:11.000 | in text it's already seen.

00:35:12.160 | So do we know that even with the right control program?

00:35:14.920 | So I think that whole thing is more messy than people think, and we're nowhere near

00:35:18.960 | there.

00:35:19.960 | No one's working on it.

00:35:20.960 | What I care about mainly is layer zero through two, and layer zero through two, we're in

00:35:23.000 | control here.

00:35:24.000 | Nothing gets out of control.

00:35:25.120 | I think it's very hypothetical to think about like a control layer that's trying to write

00:35:28.600 | a better control layer.

00:35:32.080 | It's just unclear what you can even do.

00:35:33.840 | Eventually the control layer's value is stuck on like what the language model can do, and

00:35:38.360 | the language model can only do so much, and, you know, there's a lot of interesting debates

00:35:42.040 | at layer three, but they're also very speculative right now.

00:35:45.720 | They're not things we're going to stumble into the next six months or so.

00:35:48.640 | And you went to the OpenAI headquarters like a year ago, right?

00:35:51.080 | Yeah, I've been there.

00:35:52.080 | Yeah, in the mission district.

00:35:53.080 | Did you guys talk about any of this stuff?

00:35:54.080 | No, they're not worried about this stuff.

00:35:55.920 | They're worried about just the practicality of how do you actually have a product that

00:35:58.880 | a hundred million people use around the world.

00:36:00.400 | That's just like a very complicated software problem.

00:36:04.960 | And just figuring out all the different things they have to worry about, like there's copyright

00:36:09.640 | law in this country that like affects this in a way, and it's just, you know, it's just

00:36:13.200 | a practical problem.

00:36:14.560 | Like OpenAI, it's not based on my visit, but based on just listening to interviews with

00:36:18.360 | Sam Altman recently, they care more right now I think about, for example, getting smaller

00:36:22.960 | models that can fit on a phone and can be much more responsive.

00:36:26.200 | I think they see a future in which their models can be a very effective voice interface to

00:36:31.160 | software.

00:36:32.160 | Like that's a really effective future.

00:36:33.160 | Like it's very practical what the companies are thinking about.

00:36:35.680 | This is more the philosophers and the P. Doomers in San Francisco that are thinking about mad

00:36:43.200 | scientists like recursive self-improvement.

00:36:47.200 | But anyways, it's just important.

00:36:48.200 | The control is not emergent, the control we code.

00:36:51.560 | And that's why I think the core tenet of IAI is if you produce a piece of software, you're

00:36:57.760 | responsible for its actuation.

00:37:00.000 | And that's what's going to keep you very careful about your control layers, like what you allow

00:37:04.280 | them to do or not do, no matter how smart the language model is that they're talking

00:37:08.360 | to.

00:37:09.360 | And again, I keep coming back to the language model is inert.

00:37:12.760 | The control logic can autoregressively keep calling it to get tokens out of it, but it

00:37:17.040 | is inert.

00:37:18.040 | The language model is not an intelligence that can sort of take over.

00:37:22.360 | It's just the giant collection of gears and dials that if you turn long enough, a word

00:37:26.560 | comes out the other side.

00:37:28.000 | I like your IAI.

00:37:29.000 | IAI.

00:37:30.000 | IAI.

00:37:31.000 | Easy to say, right?

00:37:32.000 | Yeah.

00:37:33.000 | IAI.

00:37:34.000 | It's like zoktok.com.

00:37:35.000 | Hopefully zoktok.com gets in some IAI.

00:37:38.560 | Oh man, I keep things difficult.

00:37:40.920 | All right.

00:37:41.920 | We got some good questions.

00:37:42.920 | A lot of them are very techie, so we'll kind of keep this nerd thing going.

00:37:45.160 | But first I want to briefly talk about one of our sponsors.

00:37:47.720 | I thought it was appropriate after a discussion of AI to talk about one of our very first

00:37:52.760 | sponsors who has integrated language model-based AI in a very interesting way into its product.

00:37:58.480 | And that is our friends at Grammarly, right?

00:38:03.000 | Grammarly, quite simply, is an AI writing partner that helps you not only get your work

00:38:07.520 | done faster, but communicate more clearly.

00:38:11.880 | 96% of Grammarly users report that Grammarly helps them craft more impactful writing.

00:38:18.440 | It works across over 500,000 apps and websites, right?

00:38:22.160 | So when you subscribe and use Grammarly, it's there for wherever you're already doing your

00:38:26.680 | writing, in your word processor, in your email client.

00:38:30.760 | Grammarly is there to help you make that writing better.

00:38:34.400 | The ways it can do this now continue to expand.

00:38:37.520 | So it's not just, hey, you said this wrong, or here's a more grammatically correct way

00:38:41.160 | to say it.

00:38:42.200 | It can now do sophisticated things, for example, like tone detection.

00:38:45.360 | Hey, what's the tone of this?

00:38:47.680 | Can you rewrite what I just wrote to sound more professional, to sound more friendly,

00:38:52.000 | to sound less friendly, right?

00:38:53.880 | It can help you get the tone just right.

00:38:56.760 | It can help you now with suggestions.

00:38:59.640 | Can you give me some ideas for this?

00:39:01.880 | Can you take this outline and write like a draft of a summary of these points, right?

00:39:05.820 | So it can generate, not just correct or rewrite, but generate in ways that as you get more

00:39:11.180 | used to it, helps you.

00:39:12.380 | And again, the key thing with Grammarly is that it integrates into all these other tools.

00:39:15.660 | You're not just over at some separate website typing into a chat interface, like Grammarly

00:39:20.380 | is where you're already doing your writing.

00:39:25.140 | It's the gold standard of responsible AI in the sense that they have for 15 years, have

00:39:29.900 | best in class communication trusted by tens of millions of professionals in IT development.

00:39:34.800 | It is a secure AI writing partner that's going to make you or your team make their point

00:39:39.020 | better, write more clearly and get your work done more faster.

00:39:42.780 | So get AI writing support that works where you work.

00:39:46.660 | Sign up and download for free at Grammarly.com/podcast, that's G-R-A-M-M-A-R-L-Y.com/podcast, easier

00:39:57.780 | said done, all right?

00:40:01.100 | Now wherever it is you happen to be doing this work, you want to feel comfortable while

00:40:04.100 | you do it.

00:40:05.100 | So I want to talk about our friends at Roan and in particular their commuter collection,

00:40:11.060 | the most comfortable, breathable and flexible set of products known to man.

00:40:14.220 | All right, here's the thing.

00:40:15.500 | The commuter collection has clothing for any situation from the world's most comfortable

00:40:21.100 | pants to dress shirts to quarter zips and polos so you never have to worry about what

00:40:24.340 | to wear when you have the commuter collection.

00:40:27.680 | It's four way stretch fabric makes it very flexible, it's lightweight, it's very breathable.

00:40:32.860 | It even has gold fusion anti-odor technology and wrinkle release technology so you can

00:40:37.500 | travel with this as you wear it, the wrinkles go away.

00:40:40.500 | It's very useful if you're on like a, you have an active day or you're on a trip, you

00:40:44.820 | can just throw this on and be speaking at conferences or teaching classes or on let's

00:40:50.060 | say a four month book tour, I know well.

00:40:54.620 | And having the commuter collection means you're going to look good but it's going to be breathable,

00:40:57.620 | you're not going to overheat, it's not going to look wrinkled, it's going to whisk away

00:41:01.820 | the odor.

00:41:02.820 | It's going to look fantastic if you are active.

00:41:05.700 | Just a great clothes.

00:41:06.700 | So the commuter collection can get you through any workday and straight into whatever comes

00:41:10.340 | next.

00:41:11.340 | Head to rhone.com/cow and use that promo code cow to save 20% off your entire order.

00:41:18.900 | That's 20% off your entire order when you head to rhone.com/cow and use that code cow

00:41:26.580 | when you check out, it's time to find your corner offer office comfort.

00:41:31.460 | All right Jesse, let's do some questions.

00:41:34.620 | Hi, first question is from Bernie.

00:41:38.100 | You often give advice on methods to consume news.

00:41:40.940 | With the advent of chat GPT and other tools, should I be worried about the spread of disinformation

00:41:46.060 | on a grand scale?

00:41:47.500 | If so, how should I manage this?

00:41:49.340 | Yeah, this is a common concern.

00:41:50.980 | So when people are trying to say what are we worried about with these large language

00:41:54.160 | models that are good at generating text, one of the big concerns is you could use it to

00:41:57.500 | generate misinformation, right?

00:42:00.740 | Generate text that's false, but people might believe, and of course it could then therefore

00:42:06.160 | be used equally for disinformation where you're doing that for particular purposes.

00:42:09.340 | I want to influence the way people think about it.

00:42:13.200 | I have two takes on this.

00:42:15.660 | I think in the general sense, I'm not as worried and let me explain why.

00:42:21.500 | What do you need for, let's just call it negative, high impact negative information.

00:42:27.220 | What do you need for these type of high impact negative information events?

00:42:30.020 | Well, you need a combination of two things, a tool that is really good at engendering

00:42:35.100 | viral spread of information that hits just like the right combination of stickiness.

00:42:41.580 | And you need a pool of this sort of available negative information that's potentially viral.

00:42:46.740 | So you have this big pool and then a selection algorithm on that pool that can find the thing

00:42:50.460 | that clicks and then let that really spread.

00:42:52.580 | That's what allows us to be in our current age of sort of widespread mis or disinformation

00:42:57.060 | is that there's a lot of information out there.

00:42:59.500 | And because in particular of social media curation algorithms, which are engagement

00:43:03.740 | focused, this tool exists that's basically surveying this pool of potential viral spreading

00:43:08.860 | information that can take this negative information and expand it everywhere, right?

00:43:15.340 | That's what makes our current moment different than, say, like 25 years ago, where the viral

00:43:19.740 | spread of information is hard.

00:43:21.540 | So it could be a lot of people with either malintended or just wrong and they don't realize

00:43:26.660 | it thoughts, you know, hey, I think the earth is flat.

00:43:29.100 | It's hard to spread it.

00:43:30.460 | Right.

00:43:31.460 | But when we added the viral spread potential of recommendation algorithms in the social

00:43:35.840 | media world, we got this current moment where mis or disinformation has the potential spreading

00:43:40.280 | really wide.

00:43:41.280 | All right.

00:43:42.340 | So what is generative AI change in this equation?

00:43:45.220 | It makes the pool of available bad information bigger.

00:43:48.340 | It is easier to generate information about whatever you want.

00:43:53.420 | For most topics we care about, that doesn't matter, right?

00:43:58.300 | Because what matters only is if AI can create content in this pool that is stickier than

00:44:03.700 | the stickiest stuff that's already there.

00:44:05.620 | There's only so many things that can spread and have a big impact.

00:44:08.140 | Right.

00:44:09.140 | And it's going to be the stickiest, the perfectly calibrated things that get identified by these

00:44:12.940 | recommendation algorithms.

00:44:14.620 | If large language models are just generating a lot of mediocre, bad information, that doesn't

00:44:18.620 | really change the equation much.

00:44:20.260 | Probably the stickiest stuff, the stuff that's going to spread best in the small number of

00:44:23.780 | slots that each of our attention has to be impacted, it's going to be like very carefully

00:44:27.140 | crafted by people.

00:44:28.220 | Like I really have a sense of like, this is going to work and we already have enough of

00:44:32.580 | that.

00:44:33.580 | And most of our slots of ideas that can impact us are filled.

00:44:38.580 | The exception to this would be very niche topics for that pool of potential bad information

00:44:43.740 | is empty because it's so niche.

00:44:45.300 | It's just nothing that there's no information about it.

00:44:47.900 | That's the case where language models could come into play because if that pool is empty,

00:44:51.860 | because it's a very specific topic, like this election in this like county, you know, it's

00:44:59.140 | not something that people are writing a lot about.

00:45:01.900 | Now someone can come in who otherwise maybe before, because they didn't have like the

00:45:05.500 | language skills, wouldn't be able to produce any text that could get virally spread here,

00:45:11.540 | could use a language model to produce it.

00:45:13.460 | The stickiest things spread, but if the pool is empty, almost anything reasonable you produce

00:45:18.460 | has the capability of being sticky.

00:45:20.000 | So that's the impact I see most immediately of missing disinformation in large language

00:45:24.620 | models is hyper-targeted mis- or disinformation.

00:45:28.920 | When it comes to big things like a national election or the way we're thinking about a

00:45:33.460 | pandemic or conspiracies about major figures or something like this, there's already a

00:45:39.540 | bunch of information adding more mediocre, bad information is not gonna change the equation.

00:45:44.780 | But in these narrow instances, that's where we have to be more wary about it.

00:45:48.420 | Unfortunately, like the right solution here is probably the same solution that we've been

00:45:52.980 | promoting for the last 15 years, which is increased internet literacy.

00:45:57.600 | Just we keep having to update what by default we trust or don't trust.

00:46:01.900 | We have to keep updating that sort of sophisticated understanding of information.

00:46:05.420 | But again, it's not changing significantly what's possible.

00:46:10.400 | It's just, it's allowing, it's simplifying the act of producing this sort of bad information

00:46:14.860 | of which there's already a lot of it that already exists.

00:46:18.740 | All right, what do we got next?

00:46:20.260 | - Next question is from Alyssa.

00:46:22.020 | Is the lack of good measurement and evaluation for AI systems a major problem?

00:46:26.740 | Many AI companies use vague phrases like improved capabilities to describe how their models

00:46:31.980 | differ from one version to the next.

00:46:34.020 | As most tech companies don't publish detailed release notes, how do I know what changes?

00:46:38.700 | - Yeah, I mean, it's a problem right now in this current age of what is happening is like

00:46:46.060 | an arms race for these mega Oracle models.

00:46:49.900 | This is not however the long-term business model of these AI companies.

00:46:52.700 | So the mega Oracle models, think of this as the chat GPT model.

00:46:56.460 | Think about this as the clod model, where you have a Oracle that you talk to through

00:47:02.500 | a chat bot about anything and you ask it to do anything and it can do whatever you ask

00:47:06.420 | it.

00:47:07.500 | And so we build these huge models, GPT-3 went to GPT-35, which went to GPT-4, which

00:47:12.860 | went to GPT, whatever it is, 4S or 5S or whatever they're calling it.

00:47:16.620 | And you are absolutely right, Alyssa, it's not always really clear what's different,

00:47:20.940 | what can this do that the other model can't.

00:47:23.740 | Often that's discovered, like, I don't know, we trained this on 5X more parameters.

00:47:28.540 | Now let's just go mess around with it and see what it does better than the last one.

00:47:31.820 | So the sort of the release notes are emergently created in a distributed fashion over time.

00:47:37.380 | But it's not the future of these companies because it's not very profitable to have these

00:47:41.780 | massive right now, the biggest models, like a trillion parameter, Sam Altman's talking

00:47:46.020 | about a potential 10 trillion parameter model.

00:47:49.580 | This is something that's going to cost on the orders of multiple hundreds of millions

00:47:52.420 | of dollars to train.

00:47:54.680 | These models are not profitable.

00:47:56.140 | They're computationally very expensive to train.

00:47:58.340 | They're computationally very expensive to run, right?

00:48:02.040 | It's like having a Bugatti supercar to drop your kids off at school five blocks away,

00:48:07.880 | you know, to be using a trillion or 10 trillion parameter model to, you know, do a summary

00:48:14.120 | of this page that you got on a Google search is just way over provisioned and it's costing

00:48:18.960 | like a lot of money.

00:48:19.960 | It's a lot of computational resources, it's expensive.

00:48:22.680 | What they want, of course, is smaller customized models to do specific things.

00:48:26.760 | We're seeing this move.

00:48:28.560 | GitHub Copilot's a great example.

00:48:30.520 | Computer programmers have an interface to a language model built right into their integrated

00:48:35.520 | development environments.

00:48:36.520 | So they can just right there where they're coding, ask for a code to be finished or another

00:48:42.480 | function to be added or ask it, what is the library that does this?

00:48:47.440 | And it will come back like, this is the library you mean, and here's the description.

00:48:50.660 | It's integrated right there.

00:48:52.680 | Microsoft Copilot, which again is confusingly named in an overlapping way, is trying to

00:48:58.040 | do something similar with Microsoft Office tools.

00:49:01.020 | You kind of have this universal chat interface to ask for actuated help with their Microsoft

00:49:07.240 | Office tools.

00:49:08.240 | Can you create a table for this?

00:49:09.440 | Can you reformat this?

00:49:10.560 | And it's going to work back and forth using layer one control with those products.

00:49:15.840 | So it's gonna be more of this.

00:49:17.080 | Again, OpenAI has this dream of having a better, like a voice interface to lots of different

00:49:21.840 | things.

00:49:22.840 | Apple Intelligence, which they've just added to their products is, you know, they're using

00:49:27.360 | chat GPT as a backend to sort of more directly deal with specific things you're doing on

00:49:32.280 | your phone.

00:49:33.280 | Like, can you take a recording of this phone conversation I just had and get a transcript

00:49:39.280 | of it and summarize that transcript of it and email it to me?

00:49:41.960 | So this is where these tools are going to get more interesting when they're doing specific,

00:49:45.860 | what I call actuated behavior.

00:49:47.840 | So they're actually like taking action on your behalf, you know, in typically the digital

00:49:52.240 | world.

00:49:53.520 | Now release notes will be more relevant.

00:49:56.440 | What can this now do?

00:49:57.440 | Okay, it can summarize phone calls, it can produce computer code, it can help me do formatting

00:50:02.520 | queries on my Microsoft Word documents.

00:50:05.000 | So I think as these models get more specialized and actuated and integrated into specific

00:50:09.400 | things we're already doing in our digital lives, the capabilities will be much more

00:50:12.880 | clearly enumerated.

00:50:14.560 | This current era of just, we all go to chat.openai.com and like, what can this thing do now?

00:50:21.040 | This is really just about, it's the equivalent of the car company having the Formula One

00:50:27.240 | racer.

00:50:28.240 | They're not planning to sell Formula One racers to a lot of people.

00:50:32.220 | But if they have a really good Formula One race car, people think about them as being

00:50:35.680 | a really good car company and so then they buy the car that's actually meant for their

00:50:39.840 | daily life.

00:50:40.840 | And so I think that's what these big models are right now.

00:50:43.280 | The bespoke models, their capabilities I think will be more clearly enumerated.

00:50:47.800 | And that's where we're going to begin to see more disruptions.

00:50:50.600 | I mean, notice we're at the year and a half mark of the chat GPT breakthrough, hasn't

00:50:56.720 | been a lot of major disruptions.

00:50:57.880 | The chat interface to a large language model, it's really cool what they can do, but right

00:51:03.360 | away they were talking about imminent disruptions to major industries.

00:51:06.440 | And we're still playing this game of like, well, I heard about this company over here

00:51:11.320 | who their neighbor's cousin replaced six of their customer service representatives.

00:51:16.240 | Like we're sort of still in that sort of passing along like a small number of examples.

00:51:22.560 | Because I don't think these models are in the final form in which they're going to have

00:51:25.120 | their key disruption.

00:51:26.120 | They haven't found their, if we can use a biological metaphor, the viral vector that's

00:51:31.880 | actually able to propagate really effectively.

00:51:34.160 | So stay tuned.

00:51:35.880 | But that's the future of these models.

00:51:37.480 | And I think their capabilities will be much more clearly enumerated when we're actually

00:51:41.280 | using them much more integrated into our daily workflow.

00:51:43.640 | I didn't know there was two co-pilots.

00:51:46.280 | Yeah.

00:51:47.280 | So Microsoft is calling their Microsoft office integration co-pilot as well.

00:51:51.240 | So it's very confusing.

00:51:52.240 | It is confusing.

00:51:53.240 | Yeah.

00:51:54.240 | All right.

00:51:55.240 | Next question is from Frank.

00:51:56.440 | Is the development of AI the biggest thing that happened in technology since the internet?

00:52:00.320 | Maybe.

00:52:01.320 | And we'll see.

00:52:02.320 | We'll see.

00:52:03.320 | I mean, what are the disruptions of the last 40 years?

00:52:06.960 | Personal computing, number one, because that's what actually made computing capable of being

00:52:11.400 | integrated into our daily lives.

00:52:14.400 | Next was the internet, democratized information and information flows, made that basically

00:52:19.280 | free.

00:52:20.280 | That's a really big deal.

00:52:21.280 | After that came mobile computing slash the rise of a mobile computing assisted digital

00:52:28.640 | attention economy.

00:52:29.640 | So this idea that the computing was portable and that the main use, the main economic engine

00:52:36.120 | of these portable computing devices would be monetizing attention, hugely disruptive

00:52:40.900 | on just like the day-to-day pattern of what our life is like.

00:52:43.800 | AI is the next big one.

00:52:46.760 | The other big one that's lurking, of course, I think is augmented reality and the rise

00:52:50.300 | of virtual screens over actual physical screens that you hold in real life.

00:52:55.520 | That's going to be less disruptive for our everyday life because that's going to be simulating

00:52:58.960 | something we're doing now in a way that's better for the companies.

00:53:01.640 | But the whole goal will be just to kind of take what we're doing now and make it virtual.

00:53:06.000 | But that's going to be hugely economically disruptive because so much of the hardware

00:53:09.640 | technology market is based on building very sleek individual physical devices.

00:53:14.280 | So I think that and AI are vying to be like, what's going to be the next biggest disruption.

00:53:19.880 | How big will it be compared to those prior disruptions?

00:53:23.080 | There's a huge spectrum here, right?

00:53:26.160 | On one end of the spectrum, it's going to be, you know, there's places where it has

00:53:33.120 | a part of our daily life where it wasn't there before.

00:53:35.880 | Like basically, maybe like email, right?

00:53:38.560 | Email really changed the patterns of work, but didn't really change what work was.

00:53:42.320 | On the other end of the spectrum, it could be much more comprehensive, maybe something

00:53:45.740 | like personal computing, which just sort of changed how the economy operated.

00:53:51.480 | You know, pre-computers, after computers fundamentally just changed the way that we interact with

00:53:56.000 | like the world and the world of information.

00:53:57.440 | It could be anywhere on the spectrum.

00:53:59.600 | Of course, there's the off-spectrum options as well as like, no, no, it like comes alive

00:54:04.040 | and completely, it's so smart that it either takes over the world or it just takes over

00:54:09.300 | all work and we all just live on UBI.

00:54:12.140 | I tend to call those off-spectrum because of what I talked about in the deep dive.

00:54:15.720 | Like we just, I don't see us having the control logic to do that yet.

00:54:20.400 | So I think really the spectrum is like personal computer on one end, email on the other.

00:54:24.480 | I don't really know where it's going to fall, but I do go back to saying the current form

00:54:28.560 | factor, I think we have to admit this, the current form factor of generative AI talking

00:54:34.040 | to a chat interface through a web or phone app has been largely a failure to cause the

00:54:38.880 | disruption that people predicted.

00:54:40.760 | It has not changed most people's lives.

00:54:42.880 | There's heavy users of it who like it, but it really has a novelty feel.

00:54:46.600 | They'll really get into detail about these really specific ways that they're, I'm getting

00:54:50.760 | ideas for my articles and having these interactions with it, but it really does have that sort

00:54:54.660 | of early internet novelty feel where you had the mosaic browser and you're like, this is

00:54:59.240 | really cool, but most people aren't using it yet.

00:55:01.680 | It's going to have to be another form factor before we see its full disruptive potential.

00:55:05.320 | And I think we do have to admit most things have not been changed.

00:55:09.960 | We're very impressed by it, but we're not impressed by its footprint on our daily life

00:55:14.040 | yet.

00:55:15.640 | So I guess this is like a dot, dot, dot, stay tuned.

00:55:19.680 | Unless your students just use it to put pass in papers, right?

00:55:23.680 | Maybe.

00:55:24.680 | So look, I have a New Yorker article I'm writing on that topic that's still in editing.

00:55:28.640 | So stay tuned for that.

00:55:30.400 | But the picture about what's happening with students and paper ride in AI, that's also

00:55:35.000 | more complicated than people think.

00:55:37.080 | What's going on there might not be what you really think, but I'll hold that discussion

00:55:40.800 | until my next New Yorker piece on this comes out.

00:55:43.560 | All right.

00:55:45.400 | Next question is from Dipta.

00:55:47.200 | How do I balance a 30 day declutter with my overall technology use?

00:55:51.080 | I'm a freelance remote worker that uses Slack, online search, stuff like that.

00:55:55.800 | All right.

00:55:56.800 | So Dipta, when talking about the 30 day declutter, is referencing an idea from my book, Digital

00:56:02.560 | Minimalism, where I suggest spending 30 days not using personal, optional personal technologies,

00:56:09.400 | get reacquainted with what you care about and other activities that are valuable.

00:56:12.500 | And then in the end, only add back things that you have a really clear case of value.

00:56:16.360 | But Dipta is mentioning here, work stuff, right?

00:56:21.040 | She's a freelance worker, use Slack, use online search, et cetera.

00:56:25.080 | My book, Digital Minimalism, which has the declutter is a book about technology in your

00:56:29.600 | personal life.

00:56:30.600 | It's not about technology at work, deep work, a world without email and slow productivity.

00:56:36.760 | Those books really tackle the impact of technology on the workplace and what to do about it.

00:56:41.480 | So digital knowledge work is one of the main topics that I'm known for.

00:56:45.480 | It's why I'm often cast, I think, somewhat incorrectly as a productivity expert.

00:56:49.840 | I'm much more of a like, how do we actually do work and not drown and hate our jobs in

00:56:55.320 | a world of digital technology?

00:56:57.520 | And it looks like productivity advice, but it's really like survival advice.

00:57:00.880 | How do we do work in an age of email and Slack without going insane?

00:57:04.840 | Digital minimalism is not about that.

00:57:06.160 | That was my book where I said, hey, I acknowledged there's this other thing going on, which is

00:57:11.440 | like, we're looking at our phones all the time in work, outside of work, unrelated to

00:57:14.800 | our work.

00:57:15.800 | We're on social media all the time.

00:57:17.280 | We're watching videos all the time.

00:57:18.760 | Why are we doing this?

00:57:19.760 | What do we do about it?

00:57:21.440 | So digital declutter is what to do with the technology in your personal life.

00:57:26.360 | When it comes to the communication technologies in your work life, read a world without email,

00:57:31.400 | read slow productivity, read deep work.

00:57:33.200 | That's sort of a separate issue.

00:57:35.240 | So I'll just use that as a roadmap for people who are struggling with the promises and peril

00:57:39.360 | of technology.

00:57:41.280 | Use my minimalism book for like the phone, the stuff you're doing on your phone that's

00:57:44.680 | unrelated to your work.

00:57:45.960 | My other books will be more useful for what's happening in your professional life.

00:57:52.040 | That often gets mixed up, Jesse, actually.

00:57:54.280 | I think in part because the symptoms are similar.

00:57:59.280 | I look at social media on my phone all the time more than I want to.

00:58:03.040 | I look at email on my computer at work all the time more than I want to.

00:58:07.320 | These feel like similar problems and the symptoms are similar.

00:58:10.780 | I am distracted in some sort of abstract way from things that are more important, but the

00:58:15.160 | causes and responses are different.

00:58:17.300 | But you're looking at your phone too much and social media too much because these massive,

00:58:21.280 | massive attention economy conglomerates are producing apps to try to generate exactly

00:58:25.240 | that response from you to monetize your attention.

00:58:27.880 | You're looking at your email so much, not because someone makes money if you look at

00:58:30.560 | your email more often, but because we have evolved this hyperactive hive mind style of

00:58:36.400 | on-demand digital aided collaboration in the workplace, which is very convenient in the

00:58:40.840 | moment, but just fries our brain in the long term.

00:58:43.280 | We have to keep checking our email because we have 15 ongoing back and forth timely conversations

00:58:47.680 | and the only way to keep those balls flying in the air is to make sure I see each response

00:58:51.540 | in time to respond in time so that things can keep unfolding in a timely fashion.

00:58:55.400 | It's a completely different cause and therefore the responses are different.

00:58:59.380 | So if you want to not be so caught up in the attention economy in your phone and in your

00:59:02.800 | personal life, well, the responses there have a lot to do with like personal autonomy, figuring

00:59:08.040 | out what's valuable, making decisions about what you use and don't use.

00:59:10.960 | In the workplace, it's all about replacing this collaboration style with other collaboration

00:59:14.940 | styles that are less communication dependent.

00:59:16.840 | So it's similar causes, but very different, I mean, similar symptoms, but very different

00:59:21.800 | causes and responses.

00:59:23.620 | Little known fact, Jesse.

00:59:25.040 | So I sold digital minimalism and a world without email together.

00:59:29.680 | It was a two book deal, like I'm going to write one and then the other.

00:59:34.980 | One of the, and it went to auctions.

00:59:36.480 | We talked to a bunch of editors about it.

00:59:39.360 | One of the editors was like, this is the, which was an interesting point, but I think

00:59:45.400 | gets to this issue.

00:59:46.400 | He's like, these are the, this is the same thing.

00:59:47.960 | We're just like looking at stuff too much in our, in our, uh, digital lives.

00:59:52.240 | This should be one book.

00:59:54.020 | These two things should be combined.

00:59:55.020 | And I was really clear, like, no, they shouldn't because actually it confuses the matter because

01:00:00.160 | they already seem so similar, but it's so different.

01:00:02.480 | Yeah.

01:00:03.480 | A world without email and slow productivity are such different books than digital minimalism.

01:00:08.400 | The causes are so different and the responses are so different that they can't be one book.

01:00:14.640 | It's, it's, it's like two fully separate issues.

01:00:16.880 | The only thing to commonality is they involve screens and they involve looking at the screens

01:00:19.820 | too much.

01:00:20.820 | Yeah.

01:00:21.820 | And so I was like, I think you're wrong about that.

01:00:22.820 | And we kept those books separate.

01:00:25.800 | Other little known fact about that, it was originally supposed to be the other order.

01:00:29.640 | The slope, uh, a world without email was the direct followup to deep work was the idea,

01:00:35.360 | but the issues in digital minimalism became so pressing so quickly that I say, no, no,

01:00:40.600 | no.

01:00:41.600 | We got, I got to write that book first.

01:00:42.600 | And so that's why slope, um, a world without email did not directly follow deep work is

01:00:47.920 | because in 2017 and 18, these issues surrounding our phone and social media, mobile, like that's

01:00:54.360 | when that really took off.

01:00:55.980 | When you were writing deep work, did you know you were going to write a world without email

01:00:59.000 | or it kind of happened?

01:01:00.000 | No, I just wrote, I just wrote deep work.

01:01:02.920 | Yeah.

01:01:04.080 | And then after I wrote deep work, I was thinking about what to write next.

01:01:06.560 | And the very next idea I had was ruled without email.

01:01:08.960 | And it was basically a response to the question of like, well, why is it so hard to do deep

01:01:11.960 | work?

01:01:12.960 | Yeah.

01:01:13.960 | Right.

01:01:14.960 | In the book, deep work, I don't get too much into it.

01:01:15.960 | I was like, we know it's technology.

01:01:16.960 | We know we're distracted all the time.

01:01:17.960 | Um, I'm not going to get into why we're in this place.

01:01:22.120 | I just want to emphasize focus is diminishing, but it's important and here's how you can

01:01:25.960 | train it.

01:01:27.400 | And then I got more into it after that book was written.

01:01:30.200 | Why did we get here?

01:01:31.200 | And it was a pretty complicated question, right?

01:01:33.240 | Like why did we get to this place where, uh, we check email 150 times a day?

01:01:37.800 | Yeah.

01:01:38.800 | It's a long book.

01:01:39.800 | Who thought this was a good idea?

01:01:40.800 | Right.

01:01:41.800 | So it was its own sort of like Epic investigation.

01:01:42.800 | Yeah.

01:01:43.800 | Yeah.

01:01:44.800 | I really liked that book.

01:01:45.800 | Um, yeah, it didn't sell the same as like digital minimalism or deep work because it's

01:01:49.840 | less just let me give you this image of a lifestyle that you can shift to right now.

01:01:55.800 | It's much more critical.

01:01:56.800 | It's much more, how did we end up in this place?

01:01:59.880 | Is this really a problem?

01:02:00.960 | It's much more of like social professional commentary.

01:02:02.760 | I mean, it has solutions, but they're more systemic.

01:02:05.600 | There's no easy thing you can do as an individual.

01:02:07.920 | I think intellectually it's a very important book and it's had influence that way, but

01:02:11.800 | it's hard to make a book like that be like a million copy seller.

01:02:15.680 | Atomic habits.

01:02:16.680 | It's not atomic habits.

01:02:17.680 | Atomic habits is easier to read than a world without email.

01:02:22.440 | I will.

01:02:23.440 | I will say that with confidence.

01:02:24.440 | Let's see what we got here.

01:02:27.560 | Uh, we've got another question.

01:02:28.960 | Ooh, it's just a slow productivity corner.

01:02:31.120 | It is.

01:02:32.120 | Do we play the music before we asked a question or do we play the music after?

01:02:34.960 | I forgot.

01:02:35.960 | Usually we play it twice.

01:02:36.960 | All right.

01:02:37.960 | Before and after.

01:02:38.960 | Let's get the before.

01:02:39.960 | All right.

01:02:40.960 | What do we got?

01:02:44.960 | Hi, this question is from Hanzo.

01:02:52.800 | I work at a large tech company as a software engineer and I'm starting to feel really overwhelmed

01:02:57.280 | by the number of projects getting thrown at us.

01:02:59.720 | How do I convince my team that we should say no to more progress projects when everyone

01:03:05.000 | has their own agenda?

01:03:06.040 | For example, pushing their next promotion?

01:03:08.040 | Well, okay.

01:03:09.600 | So this is a great question for the corner because the whole point of the slow productivity

01:03:13.240 | corner segment is that we ask a question that's relevant to my book, slow productivity, which

01:03:19.120 | as we announced the beginning of the show, the number one business book of 2024 so far

01:03:23.120 | is chosen by the Amazon editors.

01:03:25.120 | Uh, is this appropriate?

01:03:26.780 | Because I have an answer that comes straight from the book.

01:03:29.560 | So in chapter three of slow productivity, where I talk about the principle of doing

01:03:36.360 | fewer things, I have a case study that I think is very relevant to what you should, your

01:03:41.040 | team should consider Hanzo.

01:03:42.760 | So this case study comes from the technology group at the Brood Institute, a joint Harvard

01:03:48.720 | and MIT Brood Institute in Cambridge, Massachusetts.

01:03:52.480 | This is like a large sort of interdisciplinary genomics research Institute that has all these

01:03:57.480 | sequencing machines.

01:03:59.720 | But I give a profile of a team that worked at this Institute.

01:04:04.640 | These were not biologists.

01:04:05.720 | It was basically, it's not the IT team, but it's a team that like what they do is they

01:04:09.380 | build tech stuff that other scientists and people in the Institute need.

01:04:13.200 | So you come to this team and they're like, Hey, could you build us a tool to do this?

01:04:16.360 | It's a bunch of programmers and they'll let's do this, let's build that.

01:04:20.200 | They had a very similar problem as what you're describing, Hanzo.

01:04:24.800 | They, all these ideas would come up.

01:04:26.920 | Some of them would be their own.

01:04:28.040 | Some of them would be suggested by other stakeholders, you know, other scientists or teams in the

01:04:32.920 | Institute.

01:04:33.920 | And they'd be like, okay, let's work on this.

01:04:36.000 | You do this.

01:04:37.000 | I'll do this.

01:04:38.000 | Well, can you do this as well?

01:04:39.000 | And people are getting overloaded with all these projects and just things were getting

01:04:42.000 | gummed up, right?

01:04:43.000 | I mean, it's the classic idea from this chapter of the book is that if you're working on too

01:04:46.720 | many things at the same time, nothing makes progress.

01:04:50.520 | You put too many logs down the river, you get a log jam.

01:04:53.960 | None of them make it to the mill.

01:04:55.800 | So they were having this problem.

01:04:57.780 | So what they did is they went to a relatively simple pull based agile inspired project management

01:05:03.680 | workload system where whenever an idea came up, here's a project we should do.

01:05:09.900 | They put it on an index card and they put it on the wall and they had a whole section

01:05:13.560 | of the wall for like things we should, or at least consider working on.

01:05:18.080 | Then they had a column on the wall for each of the programmers.

01:05:22.440 | The things that each programmer were working on were put under their name.

01:05:26.240 | So now you have like a really clear workload management thing happening.

01:05:29.560 | If you had four or five cards under your name, they're like, this is crazy.

01:05:32.640 | We don't want you doing four or five things.

01:05:34.340 | That's impossible.

01:05:35.340 | You're going to log jam.

01:05:36.340 | You should just do one or two things at a time.

01:05:37.560 | And when you're done, we can decide as a team, okay, there's now space here for us to pull

01:05:43.080 | something new onto this person's column.

01:05:45.600 | And as a team, you could look at this big collection on the wall of stuff that you've

01:05:48.800 | identified or has been proposed to you and say, which of these things is most important.

01:05:53.440 | Equally important here as well is during this process of selecting what you're going to

01:05:57.920 | work on next, because everyone is here, it's a good time to say, well, what do I need to

01:06:02.680 | get this done?

01:06:03.680 | And you can talk to the people right there.

01:06:05.280 | I'm going to need this from you.

01:06:06.280 | I'm going to need that from you.

01:06:07.280 | When are we going to do this?

01:06:08.280 | You sort of build your contract for execution.

01:06:11.100 | So one of the things they did here is, okay, so first of all, this prevented overload.

01:06:15.220 | Each individual person can only have a couple of things in their column.

01:06:17.400 | So you didn't have people working on too many things at once.

01:06:19.640 | So you got rid of the logjam problem.

01:06:21.780 | But number two, this reminds me of your question, Hanzo.

01:06:25.560 | They noted that this also made it easier for them to, over time, weed out projects that

01:06:32.200 | they might've at some point been excited about, but are no longer excited about to weed those

01:06:40.000 | out.

01:06:41.000 | And the way they did it was they would say, this thing has been sitting over here in this

01:06:44.280 | pile of things we could work on.

01:06:46.400 | This has been sitting over there for months, and we're consistently not pulling it onto

01:06:50.100 | someone's plate.

01:06:51.280 | Let's take it off the wall.

01:06:53.600 | And so this allowed them to get past that trap of momentary enthusiasm.

01:06:57.480 | Like, this sounds awesome.

01:06:58.660 | We got to do this.

01:07:00.340 | We have those enthusiasms all the time, because here, that would just put something on the

01:07:03.560 | wall.

01:07:04.560 | But if it didn't get pulled over after a month or so, they would take it off the wall.

01:07:07.560 | So they had a way of sort of filtering through which projects should we actually work on.

01:07:11.280 | Anyways, this prevented overload.

01:07:13.400 | This is almost always the answer here.

01:07:15.040 | We need transparent workload management.

01:07:16.720 | We can't just push things on people's plates in an obfuscated way and just sort of try

01:07:20.020 | to get as much done as possible.

01:07:21.780 | We need to know what needs to be done.

01:07:23.800 | Things need to exist separate from individuals' obligations.

01:07:28.160 | And then we need to be very clear about how many things each individual should work on

01:07:30.880 | at the same time.

01:07:31.880 | So, Hanzo, you need some version of this sort of vaguely Kanban Agile-style workload management

01:07:38.680 | pull-based system.

01:07:39.680 | It could be very simple, like I talk about.

01:07:41.520 | Read the case study in chapter three of Slow Predictivity to get details.

01:07:44.960 | That will point you towards a paper from the Harvard Business Review that does an even

01:07:47.960 | more detailed case study on this team.

01:07:50.320 | Read that in detail.

01:07:51.760 | Send that around to your team or send my chapter around to your team.

01:07:56.440 | Advocate for that.

01:07:58.120 | And I think your team's going to work much better.

01:08:00.600 | All right.

01:08:01.600 | Let's get that music.

01:08:04.600 | All right.

01:08:07.600 | Do we have a call this week?

01:08:13.480 | We do.

01:08:14.480 | Let's hear it.

01:08:15.480 | Hey, Cal.

01:08:16.480 | Jason from Texas.

01:08:17.480 | Long-time listener and reader.

01:08:18.480 | First-time caller.

01:08:19.480 | For the last couple of episodes, you've been talking about applying the distributed trust

01:08:26.000 | model to social media.

01:08:27.000 | There's a lot that I like about that, but I'd like to hear you evaluate that thought

01:08:32.600 | in light of Fogg's behavioral model, which says that for an action to take place, motivation,

01:08:40.360 | prompt, and ability have to converge.

01:08:42.160 | I don't see a problem with ability, but I'm wondering about the other two.

01:08:46.320 | So if someone wants to follow, say, five creators, they're going to need significant motivation

01:08:55.600 | to be checking those sources when they're not curated in one place.

01:09:00.560 | Secondly, what is going to prompt them to go look at those five sources?

01:09:06.800 | I think if those two things can be solved, this has a real chance.

01:09:10.640 | One last unrelated note, somebody was asking about reading news articles.

01:09:15.360 | I use Send to Kindle, and I send them my Kindle and read them later.

01:09:18.720 | Works for me.

01:09:19.720 | Thanks.

01:09:20.720 | Have a great day.

01:09:21.720 | All right.

01:09:22.720 | So it's a good question.

01:09:23.720 | So I think what's key here is separating discovery from consumption.

01:09:28.320 | So the consumption problem is once I've discovered, let's say, a creator that I'm interested in,

01:09:34.560 | you know, how do I then consume that person's information in a way that's not going to be

01:09:40.620 | insurmountably high friction, right?

01:09:43.400 | So if there's a bunch of different people I've discovered one way or the other, put

01:09:46.460 | aside how I do that, how do I consume their information?

01:09:49.440 | That's the consumption problem, and that's fine.

01:09:51.680 | We've had solutions to that before.

01:09:53.160 | I mean, this is what RSS readers were.

01:09:55.800 | If I discovered a syndicated blog that I enjoyed, I would subscribe to it.

01:10:04.480 | Then that person's content is added to this sort of common list of content in my RSS reader.

01:10:10.600 | This is what, for example, we currently do with podcast.

01:10:14.320 | Podcast players are RSS readers.

01:10:16.080 | The RSS feeds now are describing podcast episodes and not blog posts, but it's the exact same

01:10:22.880 | technology, right?

01:10:25.040 | So when you have a podcast, you host your MP3 files on whatever server you want to.

01:10:29.720 | This is what I love about podcasts.

01:10:31.640 | It's not a centralized model like Facebook or like Instagram, where everything is stored

01:10:39.880 | on the servers of a single company that makes sense of all of it and helps you discover

01:10:43.280 | it.

01:10:44.280 | No, we have our servers on, our podcast are on Buzzsprout server somewhere, right?

01:10:48.520 | It's just a company that does nothing but host podcast.

01:10:50.880 | We could have our podcast, like in the old days of podcast, on a Mac studio in our HQ.

01:10:57.360 | It doesn't matter, right?

01:10:58.720 | But what you do is you have an RSS feed that every time you put out a new episode, you

01:11:02.640 | update that feed to say, here's the new episode.

01:11:05.720 | Here's the location of the MP3 file.

01:11:07.400 | Here's the title of the episode.

01:11:08.480 | Here's the description of the episode.

01:11:10.760 | All a podcast listener is, like a podcast app, is an RSS reader.

01:11:14.600 | You subscribe to a feed.

01:11:16.360 | It checks these feeds.

01:11:18.020 | When it sees there's a new episode of a show because that RSS feed was updated, it can

01:11:22.280 | put that information in your app.

01:11:24.980 | It can go and retrieve the MP3 file from whatever server you happen to be serving it on, and

01:11:29.200 | then it can play it on your local device.

01:11:30.580 | So we still use something like RSS.

01:11:32.800 | So consumption's fine.

01:11:33.800 | We get very nice interfaces for where do I pull together and read in a very nice way

01:11:39.360 | or listen in a very nice way or watch in a very nice way.

01:11:43.520 | Because by the way, I think video RSS is going to be a big thing that's coming.

01:11:47.600 | You make really nice readers.

01:11:49.840 | Now we go over to the discovery problem.

01:11:51.000 | Okay, well, how do I find the things that subscribe to in the first place?

01:11:54.000 | This is where distributed trust comes into play.

01:11:55.860 | It's the way we used to do this pre-major social media platforms.

01:12:00.040 | How did I discover a new blog to read?

01:12:02.520 | Well, typically it would be through these distributed webs of trust.

01:12:05.640 | I know this person.

01:12:06.820 | I've been reading their stuff.

01:12:07.940 | I like their stuff.

01:12:09.440 | They link to this other person.

01:12:10.840 | I trust them, so I followed that link.

01:12:13.120 | I liked what I saw over there, and so now I'm going to subscribe to that person.

01:12:18.260 | Or three or four people that I trust are in my existing web of trust have mentioned this

01:12:24.240 | other person over here.

01:12:26.080 | That now builds up this human to human curation, this human to human capital of this is a person

01:12:31.500 | who is worthy of attention.

01:12:33.720 | So now I will go and discover them, and I like what I see, and then I subscribe, and

01:12:36.880 | the consumption happens in like a reader.

01:12:38.760 | So we've got to break apart discovery and consumption.

01:12:42.620 | It's the moving discovery away from algorithms and back towards distributed webs of trust.

01:12:48.660 | That's where things are interesting.

01:12:50.260 | That's where things get interesting.

01:12:51.460 | That's where we get rid of this feedback cycle of production, recommendation algorithm, feedback

01:12:59.940 | to producers about how popular something was, which changes how they produce things into

01:13:05.100 | the feedback algorithm, feedback.

01:13:07.820 | That cycle is what creates this sort of hyper palatable, lowest common denominator, amygdala

01:13:14.620 | plucking highly distractible content.

01:13:17.940 | You get rid of the recommendation algorithm piece of that, that goes away.

01:13:21.940 | It also solves problems about disinformation and misinformation.

01:13:24.940 | I mean, I argued this early in the COVID pandemic.

01:13:29.180 | I wrote this op-ed for Wired, where I said like the biggest thing we could do for both

01:13:34.620 | the physical and mental health of the country right now would be to shut down Twitter.

01:13:38.220 | I said what we should do instead is go back to an older web two model, where information

01:13:43.860 | was posted on websites, like blogs and articles posted on websites, and yeah, it's going to

01:13:48.060 | be higher friction to sort of discover which of these sites you trust.

01:13:52.740 | But this distributed web of trust is going to make it much easier for people to curate

01:13:57.300 | the quality of information, right?

01:13:59.420 | Like this blog here is being hosted by a center of a major university.

01:14:06.340 | I have all of this capital in me trusting that more than trusting johnnybananas.com/covidconspiracies.

01:14:16.780 | I don't trust that as much, right?

01:14:18.820 | Or I'm going to have to follow old-fashioned webs of trust to find my way to sort of like

01:14:23.180 | a new commentator on something like this.

01:14:25.500 | And this is not really an argument for, yeah, but you're going to fall back to unquestioning

01:14:28.780 | authority.

01:14:29.780 | Webs of trust work very well for independent voices.

01:14:33.540 | They work very well.

01:14:35.100 | They're very useful for critiques of major voices.

01:14:38.100 | It is slower for people to find independence or critical voices.

01:14:43.040 | But if you find them through a web of trust, they're much more powerful and it filters

01:14:48.000 | out the crank stuff, which is really bad for independent and critical voices because it

01:14:52.420 | can get pushed in.

01:14:54.420 | That's the same.

01:14:55.420 | This person here critiquing this policy, that's the same as this other person over here who

01:14:59.340 | says it's the lizard people.

01:15:01.700 | Webs of trust, I think, are a very effective way to navigate information in a low friction

01:15:05.860 | information environment like the internet.

01:15:08.340 | So I think distributed webs of trust, I really love that model.

01:15:11.580 | It's what we're doing with podcasts.

01:15:13.680 | It's also what we're doing with newsletters.

01:15:15.940 | So this is not like a model that is retroactive or reactionary.

01:15:20.540 | It's not regressive.

01:15:22.140 | It's not let's go back to some simpler technological age to try to get some...

01:15:27.500 | We're doing it right now in some sectors of online content and it's working great.

01:15:33.780 | Podcast or digital trust.

01:15:36.700 | Algorithms don't show us what podcast to listen to.

01:15:39.220 | They don't spread virally and then we're just shown it and it catches our attention.

01:15:42.820 | We have to hear about it.

01:15:43.820 | We probably have to hear about it multiple times from people we trust before we go over

01:15:47.420 | and we sample it, right?

01:15:48.940 | That's distributed webs of trust.

01:15:51.260 | Email newsletters are the same thing.

01:15:53.460 | It's a vibrant online content community right now.

01:15:56.140 | How do people discover new email newsletters?

01:15:58.860 | People they know forward them individual email newsletters like, "You might like this."

01:16:05.780 | And they read it and they say, "I do and I trust you and so now I'm going to consider

01:16:10.540 | subscribing to this," right?

01:16:14.140 | That's webs of trust.

01:16:15.140 | It's not an algorithm as much as Substack is trying to get into the game of algorithmic

01:16:19.380 | recommendation or be like the Netflix of text.

01:16:21.900 | Right now that model works.

01:16:22.900 | So anyways, that's where I think we go.

01:16:25.900 | I like to think of the giant monopoly platform social media age as this aberration, this

01:16:31.220 | weird divergence of the ultimate trajectory of the internet as a source of good.

01:16:37.220 | And the right way to move forward on that trajectory is to continually move away from

01:16:40.700 | the age of recommendation algorithms in the user-generated content space and return more

01:16:45.620 | to distributed webs of trust.

01:16:48.080 | Recommendation algorithms themselves, these are useful, but I think they're more useful

01:16:52.500 | when we put them in an environment where we don't have the user-generated content and

01:16:56.500 | feedback bit of that loop.

01:16:59.020 | They're very useful on like Netflix.

01:17:00.980 | "Hey, you might like this show if you like that other show."

01:17:05.220 | That's fine.

01:17:06.960 | They're very useful Amazon to say, "This book is something you might like if you like that

01:17:10.460 | book."

01:17:11.460 | That's fine.

01:17:12.460 | I'm happy for you to have recommendation algorithms in those contexts.

01:17:14.940 | But if you hook them up with user-generated content and then feedback to the users about

01:17:18.740 | popularity, that's what in a Marshall McLuhan way sort of evolves the content itself in

01:17:24.860 | the ways that are, I think, undesirable and as we see have really negative externalities.

01:17:28.700 | So anyways, we've gone from geeking out on AI to geeking out on my other major topic,

01:17:34.020 | which is distributed webs of trust.

01:17:35.380 | But I think that is the way to discover information.

01:17:38.380 | Hopefully that's the future of the internet as well.

01:17:40.460 | And I love your idea, by the way, of the Send a Kindle cool app.

01:17:44.740 | You send articles to your Kindle, and then you can go take that Kindle somewhere outside

01:17:49.140 | under a tree to read news articles.

01:17:51.300 | No ads, no links, no rabbit holes, no social media.

01:17:53.940 | It's a beautiful application.

01:17:55.940 | Send a Kindle.

01:17:57.220 | I highly recommend.

01:17:58.220 | All right.

01:17:59.220 | I think we have a case study.

01:18:00.820 | This is where people send in a description of using some of my ideas out there in the

01:18:05.620 | real world.

01:18:07.820 | Are we been asking people to send these to you, Jesse?

01:18:10.260 | Yeah.

01:18:11.260 | Yeah.

01:18:12.260 | Jesse@CalNewport.com.

01:18:13.260 | Yeah.

01:18:14.260 | So if you have a case study of putting any of these ideas into action, send those to

01:18:15.260 | Jesse@CalNewport.com.

01:18:16.260 | If you want to submit questions or calls, just go to thedeeplife.com/listen.

01:18:19.980 | Yeah.

01:18:20.980 | And there's also a section in there if they go to that website where they can put in a

01:18:24.200 | case study.

01:18:25.200 | Yeah.

01:18:26.200 | Okay.

01:18:27.200 | And we have links there for submitting questions.

01:18:28.200 | We have a link there where you can record a call straight from your phone or browser.

01:18:29.700 | It's real easy.

01:18:30.700 | All right.

01:18:31.700 | Our next case study comes from Salim who says, "I work at a large healthcare IT software

01:18:37.380 | company in our technical solutions division.

01:18:40.380 | Our work is client-based, so we'll always work with the same analyst teams as our assigned

01:18:44.820 | clients.

01:18:45.820 | While I enjoy the core work, which is problem-solving based, I was struggling with a large client

01:18:50.640 | load and specifically with one organization that did not align well with my communication

01:18:55.740 | style and work values.

01:18:57.540 | This was a constant problem in my quarterly feedback, and I was struggling with convincing

01:19:01.700 | the staffing team to make reassignment.

01:19:04.620 | Around this time, our division had recently rolled out a work plan site for employees

01:19:09.180 | to plan out their weekly hours in advance.

01:19:11.220 | The issue here was that it was communicated as a requirement.

01:19:14.620 | So most of us saw this as upper micromanagement.

01:19:18.020 | The site itself is also unstructured, so we didn't see the utility in doing this since

01:19:21.820 | we already log our time retroactively anyways.

01:19:24.980 | At this point, I had already read Deep Work and was using the time block planner, but

01:19:29.300 | was lacking a system for planning at a weekly timescale.

01:19:33.740 | This is where I started leveraging our work plan site and structured it in terms of what

01:19:37.700 | I was working on during any given week.

01:19:40.540 | This included itemizing my recurring calls, office hours with clients, and a general estimate

01:19:45.620 | of how much time I would spend on client work per client.

01:19:48.820 | I incorporated sections for a top priority list and a pull list backlog so I could quickly

01:19:52.980 | go in and reprioritize as new ideas came in or as I had some free time.

01:19:57.980 | I also added a section to track my completed tasks so that I could visually get a sense

01:20:01.580 | of my progress as the week went by.

01:20:04.540 | After I made this weekly planning a habit, my team lead highlighted my approach at a

01:20:07.980 | monthly team meeting, and we presented on how I leveraged the tool into something useful

01:20:12.100 | for managing my work.

01:20:13.500 | I spoke to how this helped me organize me week to week so that I can take a proactive

01:20:17.820 | approach and slow down versus being at the mercy of a hive mind mentality, constantly

01:20:23.900 | reacting to incoming emails and team messages, and he goes on to mention some good stuff

01:20:30.580 | that happened after that.

01:20:31.580 | All right, it's a great case study, Salim.

01:20:34.060 | What I like about it is that it emphasizes there are alternatives to what I call the

01:20:39.020 | list reactive method.

01:20:41.340 | The list reactive method says you kind of just take each day as it comes, reacting the

01:20:45.540 | stuff that's coming in over the transom through email and Slack, trying to make progress on

01:20:49.860 | some sort of large to-do list as well.

01:20:51.660 | Like, okay, what should I work on next?

01:20:53.060 | I'll react to things and try to make some progress on my to-do list.

01:20:56.420 | It is not a very effective way to make use of your time and resources.

01:21:01.980 | You get caught up in things that are lower value.

01:21:04.700 | You lose the ability to give things the focus work required to get them done well and fast.

01:21:09.380 | You fall behind on high priorities and get stuck on low priorities, so you have to be

01:21:13.180 | more proactive about controlling your time.

01:21:15.740 | Control, control, control is a big theme about how I talk about thriving in digital age knowledge

01:21:19.900 | work.

01:21:20.900 | So I love this idea that the weekly plan discipline I talk about could be a big part of that answer.

01:21:25.220 | Look at your week as a whole and say, what do I want to do with this week?

01:21:29.220 | Where are my calls?

01:21:30.220 | Where's my client office hours?

01:21:31.220 | When am I working on this client?

01:21:32.220 | Why don't I consolidate all this time into this time over here surrounding this call

01:21:35.100 | we're already going to have?

01:21:36.100 | Why don't I cancel these two things because they're really making the rest of the week

01:21:39.020 | not work?

01:21:40.700 | When you plan your week in advance, it really helps you have a better week than if you just

01:21:46.780 | stay at the scale of what am I doing today or even worse, the scale of just what am I

01:21:51.420 | doing next?

01:21:53.300 | So multi-scale planning is critical for this control, control, control rhythm that I preach.

01:21:59.060 | That's the only way really to survive in digital area knowledge work.

01:22:02.620 | So what a cool example of weekly planning, helping you feel like you actually had some

01:22:07.860 | autonomy once again over your schedule.

01:22:10.620 | All right.

01:22:11.620 | So we've got a cool final segment.

01:22:13.020 | I want to react to an article in the news, but first let's hear from another sponsor.

01:22:17.660 | Look, the older I get and trust me, my birthday's in a few days.

01:22:21.740 | So I'm thinking about this.

01:22:23.700 | The more I find myself wanting to be more intentional about the way I live.

01:22:28.380 | And we talk about this all the time, my month, my birthday life planning, but I also want

01:22:34.740 | to make sure I'm being intentional about how I eat and take care of my body.

01:22:44.920 | This is why I really like our sponsor Mosh, M-O-S-H, Mosh bars.

01:22:52.820 | I love these things.

01:22:54.260 | It's one of my sort of go-to when I have these around, it's really like a go-to snack that

01:23:01.300 | I like to turn to.

01:23:02.900 | I love my Mosh bars.

01:23:06.260 | Let me tell you a little bit more why Mosh, which was started by actually interestingly

01:23:11.420 | Maria Shriver and her son, Patrick Schwarzenegger, with the mission of creating a conversation

01:23:16.780 | about brain health because Maria's father suffered from Alzheimer's.

01:23:21.420 | So they care a lot about this.

01:23:22.940 | They developed Mosh bars by joining forces with the world's top scientists and functional

01:23:28.220 | nutritionist that give you a protein bar that goes beyond what you normally get in a protein

01:23:32.920 | bar product.

01:23:33.920 | It has 10 delicious flavors, including three that are plant-based.

01:23:37.340 | It is made with ingredients that support brain health, like ashwagandha, lion's mane, collagen,

01:23:43.800 | and omega-3.

01:23:46.500 | Mosh bars come with a new look and a new formulation featuring a game-changing brain-boosting ingredient

01:23:50.580 | you won't find at any other bar.

01:23:53.100 | It is the first and only food brand boosted with Cognizant, a premium nootropic that supplies

01:23:58.900 | the brain with a patented form of cytokoline or cytokoline.

01:24:04.740 | Here's the thing about the Mosh bars, they taste good.

01:24:06.860 | I like them because they are soft with a little bit of crunch inside of them.

01:24:10.780 | So you really, really crave eating these things.

01:24:13.420 | A lot of protein sort of gives you what you need, and it has these sort of brain-boosting

01:24:16.780 | ingredients as well, which really comes from Maria and her son, Patrick's real concern

01:24:21.940 | about brain and brain's health.

01:24:25.780 | That concern is also why Mosh donates a portion of all proceeds to fund gender-based brain

01:24:30.500 | health research through the Women's Alzheimer's Movement.

01:24:34.160 | Why gender-based?

01:24:35.260 | Because Maria and Patrick noted that two-thirds of all Alzheimer's patients are women.

01:24:39.800 | So Mosh is working closely to close the gap between women's and men's health research,

01:24:43.540 | especially on this particular topic.

01:24:45.300 | All right, so great tasting protein bars that have all this really cool stuff in them, built

01:24:51.740 | by cool people pushing a good cause.

01:24:55.480 | So if you want to find ways to give back to others and fuel your body and your brain,

01:24:59.420 | Mosh bars are the perfect choice for you.

01:25:01.860 | Head to moshlife.com/deep to save 20% off plus free shipping on either the Best Sellers

01:25:07.540 | Trial Pack or the new Plant-Based Trial Pack.

01:25:10.740 | That's 20% off plus free shipping on either the Best Sellers or Plant-Based Trial Pack

01:25:16.060 | at moshlife.com/deep.

01:25:23.440 | Thank you, Mosh, for sponsoring this episode.

01:25:26.020 | I also want to talk about our friends at Shopify.

01:25:30.140 | Whether you're selling a little or a lot, Shopify helps you do your things however you

01:25:36.420 | cha-ching.

01:25:37.860 | If you sell things, you need to know about Shopify.

01:25:41.380 | It's the global commerce platform that helps you sell at every stage of your business.

01:25:46.120 | From the launch your online shop stage to the first real life store stage, all the way

01:25:51.660 | to the "did we just hit a million order" stage, Shopify is there to help you grow.

01:25:58.700 | They have an all-in-one e-commerce platform, which makes checking out online an absolute

01:26:04.100 | breeze with super high conversion.

01:26:06.220 | They also have in-person POS systems right there in the store that what the people use

01:26:10.600 | to actually do their credit card and do their transactions.

01:26:14.020 | However you're selling, Shopify is really best in class at what they're doing.

01:26:18.900 | They've even added, matching the theme of today's program, an AI feature called Shopify

01:26:25.400 | Magic that helps you sell even more to your customers, be more successful at conversions.

01:26:33.420 | It's a no-brainer.

01:26:34.780 | You're selling something, you do need to check out Shopify.

01:26:38.740 | The good news is I can help you do that with a good deal.

01:26:42.740 | You can sign up for a $1 per month trial period at shopify.com/deep.

01:26:48.180 | You got to type that all lowercase, but if you go to shopify.com/deep now, you can grow

01:26:56.640 | your business no matter what stage you're in.

01:26:58.940 | That's shopify.com/deep.

01:27:00.380 | All right, Jesse, let's do our final segment.

01:27:06.640 | All right, this article was sent to me a lot, and I guess it's because I'm mentioned in

01:27:12.680 | it or because it feels like it's really important.

01:27:14.780 | I brought it up here on the screen for people who are watching instead of just listening.

01:27:20.020 | The article that most people sent me on this issue came from Axios.

01:27:25.020 | Emily Peck wrote it.

01:27:26.020 | The title of the Axios articles is "Why Employers Wind Up With Mouse-Jiggling Workers."

01:27:33.540 | All right, so they're talking about mouse jigglers, which I had to look up.

01:27:40.820 | But it is software you can run on your computer that basically moves your mouse pointer around.

01:27:45.260 | So it simulates like if you're actually there jiggling your formal mouse.

01:27:51.740 | Well, it turns out a bunch of mouse jigglers got fired at Wells Fargo.

01:27:59.880 | They discovered that they were using the mouse jigglers, and they fired workers from their

01:28:05.780 | wealth and investment management unit.

01:28:10.060 | So we're kind of looking into this.

01:28:11.100 | There's a couple of reasons why the mouse jiggling is useful for remote workers.

01:28:15.620 | One of them is the fact that common instant message agents like Slack and Microsoft Teams

01:28:22.520 | puts this little status circle next to your name.

01:28:25.620 | So if I'm looking at you in Slack or Teams, there's a status circle that says whether

01:28:29.500 | you're active or not.

01:28:31.360 | The idea being like, "Hey, if you're not active, then I won't text you.

01:28:35.420 | I won't send a message.

01:28:36.420 | And if you are, like, if I know you're there working on your computer, I will."

01:28:39.260 | Well, if your computer goes to sleep, your circle turns to inactive.

01:28:44.060 | So the mouse jigglers keeps your circles active.

01:28:47.420 | So if your boss is just like, "Hey, what's going on with Cal over here?"

01:28:50.940 | They just sort of see like, "Oh, he must be working all very hard because his circle is

01:28:54.660 | always green.

01:28:55.660 | So he's there on your computer."

01:28:56.820 | When in reality, you could be away from your computer, but the mouse jiggler is making

01:28:59.620 | it seem active.

01:29:00.620 | All right.

01:29:01.620 | So there's been a kind of a lot of outrage about the mouse jigglers and about this type

01:29:07.060 | of surveillance.

01:29:09.780 | So what do I feel about it?

01:29:11.100 | Well, I'm cited in this Axios article, so we can see what they think I feel about it.

01:29:16.220 | Let's see here.

01:29:17.220 | All right.

01:29:18.300 | Here is how my take is described by Axios, and I'll see if I agree with this.

01:29:23.140 | Remote surveillance is just the latest version of a boss looking out at the office floor

01:29:27.460 | to check that there are butts in seats.

01:29:30.040 | These kinds of crude measures are part of a culture of pseudoproductivity that kicked

01:29:33.260 | off in the 1950s with the advent of office work, as Cal Newport writes in his latest

01:29:37.860 | book with a link to slow productivity.

01:29:42.260 | With technology-enabled 24-hour connection to the workplace, pseudoproductivity evolved

01:29:45.980 | in ways that wound up driving worker burnout, like replying to emails at all hours or chiming

01:29:49.900 | in on every Slack message.

01:29:51.300 | And with the rise of remote work, this push for employees to look busy and for managers

01:29:55.180 | to understand who's actually working got even worse, Newport told me in a recent interview.

01:29:59.660 | It just spiraled completely out of control.

01:30:01.620 | Well, you know what?

01:30:02.620 | I agree with this Cal Newport character.

01:30:05.980 | This is the way I see this, and I think this is the right way to see this.

01:30:10.540 | There's a smaller argument, which I think is too narrow, which is the argument of bosses

01:30:17.420 | are using remote surveillance, we should tell bosses to stop using remote surveillance.

01:30:22.460 | I think this is the narrower thing here.

01:30:25.420 | Digital tools are giving us ways to do this privacy-violating surveillance, and we should

01:30:30.940 | push back on that.

01:30:32.620 | Fair enough.

01:30:33.620 | It's not the bigger issue.

01:30:35.200 | The bigger issue is what's mentioned here, this bigger trend.

01:30:38.300 | This is what I outline in chapter one of my book, Slow Productivity.

01:30:43.700 | It's what explicitly puts this book in the tradition of my technology writings, why this

01:30:48.140 | book is really a technology book, even though it's talking about knowledge work.

01:30:51.900 | And here's the argument.

01:30:54.280 | For 70 years, knowledge work has depended on what I call pseudo-productivity, this heuristic

01:30:59.540 | that says visible activity will be our proxy for useful effort.

01:31:04.300 | We do this not because our bosses are mustache twirlers or because they're trying to exploit

01:31:09.020 | us, but because we didn't have a better way of measuring productivity in this new world

01:31:13.540 | of cognitive work.

01:31:14.540 | There's no widgets I can point to.

01:31:16.880 | There's no pile of Model Ts lined up in the parking lot that I can count.

01:31:21.420 | So what we do is like, well, to see you in the office is better than not.

01:31:24.420 | So come to the office, do factory shifts, be here for eight hours, don't spend too much

01:31:27.900 | time at the coffee machine.

01:31:30.180 | So we had this sort of crude heuristic because we didn't know how else to manage knowledge

01:31:36.060 | workers.

01:31:37.060 | And as pointed out in this article, that way of crudely managing productivity didn't play

01:31:44.660 | nicely with the front office IT revolution.

01:31:47.740 | And this mouse jiggler is just the latest example of this reality.

01:31:52.260 | When we added 24-hour remote internet-based connectivity through mobile computing that's

01:31:57.940 | with us at all times to the workplace, pseudo-productivity became a problem.

01:32:02.500 | When pseudo-productivity meant, okay, I guess I have to come to an office for eight hours

01:32:05.540 | like I'm putting steering wheels on a Model T, that's kind of dumb, but I'll do it.

01:32:09.460 | And that's what pseudo-productivity meant.

01:32:11.300 | And also like, if I'm reading a magazine at my desk, keep it below where my boss can see

01:32:16.500 | it.

01:32:17.500 | Fair enough.

01:32:18.580 | But once we got laptops and then we got smartphones and we got the mobile computing revolution,

01:32:23.700 | now pseudo-productivity meant every email I reply to is a demonstration of effort.

01:32:29.700 | Every Slack message I reply to is a demonstration of effort.

01:32:32.100 | I could be doing more effort at any point.

01:32:34.500 | In the evening, I could be doing it.

01:32:36.100 | At my kid's soccer game, I could be showing more effort.

01:32:39.340 | This was impossible in 1973, completely possible in 2024.

01:32:43.580 | This is what leads us to things like, I'm going to have a piece of software that artificially

01:32:47.000 | shakes my mouse because that circle being green next to my name in Slack longer is showing

01:32:52.840 | more pseudo-productivity.

01:32:55.340 | So the inanity of pseudo-productivity becomes pronounced and almost absurdist in its implications

01:33:00.940 | once we get to the digital age.

01:33:03.620 | That's why I wrote Slow Productivity Now.

01:33:05.420 | That's why we need slow productivity now, because we have to replace pseudo-productivity

01:33:09.700 | with something that's more results oriented and that plays nicer with the digital revolution.

01:33:13.780 | So this is just like one of many, many symptoms of the diseased state of modern knowledge

01:33:19.480 | work that's caused by us relying on this super vague and crude heuristic of just like doing

01:33:25.020 | stuff is better than not doing stuff.

01:33:26.740 | We have to get more specific.

01:33:28.880 | Slow productivity gives you a whole philosophical and tactical roadmap to something more specific.

01:33:34.300 | It's based on results.

01:33:35.300 | It's not based on activity.

01:33:37.320 | It's based on production over time, not on busyness in the moment.

01:33:42.300 | It's based on sequential focus and not on concurrent overload.

01:33:47.060 | It's based on quality and not activity, right?

01:33:50.900 | So it's an alternative to the pseudo-productivity that's causing problems like this mouse jiggler

01:33:55.440 | problem.

01:33:56.800 | So that's the bigger problem.

01:33:59.120 | New technologies requires us to finally do the work of really updating what we think

01:34:02.740 | about knowledge work.

01:34:03.740 | That's why I wrote that most recent book about it.

01:34:07.020 | It's also why I hate that status light in Slack or Microsoft Teams.

01:34:12.100 | Of course that's going to be a problem.

01:34:13.100 | Of course that's going to be a problem.

01:34:15.900 | And even the underlying mentality of that status light, which is like, if you're at

01:34:19.380 | your computer, it's fine for someone to send you a message.

01:34:23.780 | Why is that fine?

01:34:24.780 | What if I'm at my computer?

01:34:25.780 | What if I'm doing something cognitively demanding?

01:34:27.460 | It's a huge issue for me to have to turn over to your message.

01:34:30.640 | So it also underlines the degree to which the specific tools we use completely disregard

01:34:36.580 | the psychological realities of how people actually do cognitive effort.

01:34:40.360 | So we have such a mess in knowledge work right now.

01:34:42.860 | It's why, whatever, three of my books are about digital knowledge work.

01:34:46.340 | It's why we talk about digital knowledge work so much on this technology show is because

01:34:50.780 | digital age knowledge work is a complete mess.

01:34:53.840 | The good news is that gives us a lot of low hanging fruit to pick.

01:34:56.740 | That's going to cause advantages, delicious advantages.

01:35:00.300 | So there's a lot of good work to do.

01:35:02.100 | There's a lot of easy changes we could make, but anyways, I'm glad people sent me this

01:35:05.660 | article.

01:35:06.660 | I'm glad I'm appropriately quoted here.

01:35:08.100 | This is accurate.

01:35:09.100 | This is the way I think about it.

01:35:10.420 | And this is the big issue.

01:35:12.980 | Not narrow surveillance, but broad pseudo productivity plus technology is an unsustainable

01:35:19.220 | combination.

01:35:20.220 | All right, well, I think that's all the time we have for today.

01:35:23.740 | Thank you everyone who sent in their questions, case studies and calls.

01:35:27.500 | Be back next week with another episode, though it will probably be an episode filmed from

01:35:32.560 | an undisclosed location.

01:35:34.340 | I'm doing my sort of annual retreat into the mountains for the summer.

01:35:37.640 | No worries.

01:35:38.640 | The show will still come out on its regular basis, but just like last year, we'll be recording

01:35:42.420 | some of these episodes with Jesse and I in different locations and I'll be in my undisclosed

01:35:46.780 | mountain location.

01:35:47.780 | I think next week might be the first week that is the case, but the shows will be otherwise

01:35:50.980 | normal and I'll give you a report from what it's like from wherever I end up.

01:35:55.660 | I'll tell you about my sort of deep endeavors, whatever deep undisclosed location I find,

01:36:01.900 | but otherwise we'll be back and I'll see you next week.

01:36:03.860 | And until then, as always, stay deep.

01:36:06.660 | Hey, if you liked today's discussion about diffusing AI panic, you might also like episode

01:36:12.700 | 244 where I gave some of my more contemporaneous thoughts on chat GPT right around the time

01:36:19.140 | that it first launched.

01:36:20.580 | Check it out.

01:36:21.580 | That is the deep question I want to address today.

01:36:25.340 | How does chat GPT work and how worried should we be about it?

Debunking The AI Reset: Alien Mind Fear, Chat GPT, Future of AI & Slow Productivity | Cal Newport

Chapters