AI Safety, The China Problem, LLMs & Job Displacement - Dwarkesh Patel

1. 0:00 – 6:59

   Has AI Accelerated Our Understanding of Human Intelligence?

   1. CWChris Williamson0:00

      What do you think that we've realized about human learning and human intelligence from architecting AI intelligence?

   2. DPDwarkesh Patel0:08

      Hmm. There's this really interesting thing we've seen where these AI models are making progress first in the domains that we think of as the archetype of, um, the- the- where humans have their primacy, right? So if you look at, um, Aristotle. What does he say? What- what makes humans unique? Um, well, it's reasoning. Humans can reason, other animals can't. And wha- these models, these AI models, they're just not that useful, uh, if you try to use them for your work. They're useful in certain domains, but broadly, they're just not, um, widely deployable. What is the one thing that they can do? They can reason. Um, they- but they obviously, th- uh, uh, they can't carry a cup of water, right? Robotics isn't solved. They- they can't even, like, do a job. They can't even do a white-collar job. So, um, uh, there's this interesting thing called Moravec's paradox. Hans Moravec came up with this idea in the '90s, where he noticed that the tasks which are, um, easiest for humans are taking computers the longest to solve, so we ha- still haven't solved robotics yet. We c- it's so easy for us to move around. Whereas the tasks which are quite hard for humans, like adding numbers, adding long numbers, that computers could do that in the '60s. And the logic there is that, uh, evolution has only optimized us for, let's say, the last million years to be good at reasoning, to be good at arithmetic, to be good at these kinds of high-level abstractions. And we'll just have spent four billion years teaching us how to move around the world, how to, um, pursue your goals on a long-term basis, so not just do this task over the next hour-

   3. CWChris Williamson1:38

      Hmm.

   4. DPDwarkesh Patel1:38

      ... but spend the next month ha- planning how to kill this gazelle. Um, uh, and that- that has been, I think, remarkably accurate predictor of the places we've seen AI progress. They're like- they're automating coding. Coding, we thought of was this thing that .1% of population could do really well. That's- that's the first thing w- that went below the waterline. Um, uh, and yeah, just, like, basic, you know, manual work might genuinely be the last thing that goes away.

   5. CWChris Williamson2:03

      Right. Yeah, there's a difficulty in getting a robot to crack an egg.

   6. DPDwarkesh Patel2:07

      Right.

   7. CWChris Williamson2:07

      A particular difficulty in being able-

   8. DPDwarkesh Patel2:09

      Yeah.

   9. CWChris Williamson2:09

      ... to do that, the right amount of tension to hold. Is there a ... this may be outside of your domain of competence, but that's why we do podcasting-

   10. DPDwarkesh Patel2:16

       (laughs)

   11. CWChris Williamson2:16

       ... to talk about things that are outside our domain of competence. Is there a potential to use some sort of scanning technology to take an LLM-type approach to teaching robots how humans move?

   12. DPDwarkesh Patel2:28

       Hmm.

   13. CWChris Williamson2:28

       You know, if you were able to track within a room exactly how a human was to just go about tasks, just feed that into a big fuck-off model-

   14. DPDwarkesh Patel2:36

       Right.

   15. CWChris Williamson2:36

       ... and then use that to rep- I guess, you can't really work out sort of force application just by looking. That would be-

   16. DPDwarkesh Patel2:42

       Right.

   17. CWChris Williamson2:43

       ... something you'd have to fi-... Maybe you could put someone in a sui-... I don't know. You know, I- n- I'm wondering if we've seen so much progress using LLMs in the world of AI. Robotics seems to be something that's still kind of, uh, pretty janky.

   18. DPDwarkesh Patel2:56

       Yeah.

   19. CWChris Williamson2:56

       I'm wondering if there are any principles that can be taken from the world of LLM that can be applied to robotics.

   20. DPDwarkesh Patel3:03

       I mean, that's a great question, uh, and many companies are working on it. My understanding is that it's- it's difficult for the fact that there's not as much data, just y- what you mentioned-

   21. CWChris Williamson3:13

       Mm-hmm.

   22. DPDwarkesh Patel3:14

       ... that the kind of data you need of, like, what did it feel like-

   23. CWChris Williamson3:16

       There's no internet for human movement.

   24. DPDwarkesh Patel3:18

       Exactly, right? And even video is limited. And even if you have the video, it's not... With language, you have this thing of you are exactly doing the thing which, uh, the online internet text is, right? You are predicting the next token in text. Y- uh, you can predict the next thing in a video frame. That's not the same thing as robotics. There's also additional challenges, from what I understand, around, um, the fact that video is harder to process than text. It's just, like, a lot more data. There's latency overhead, so if it takes you a while to process, uh, language, that's fine. You can, you know, go a token at a time. The real world just moves very fast. You can try to solve these issues by going in simulation, um, uh, so, you know, you- you can- you can have a simulation where you're trying to move things around, and in that domain, you can train an AI to be good at robotics. But the real world is just, like, very complicated.

   25. CWChris Williamson4:01

       Too chaotic.

   26. DPDwarkesh Patel4:02

       Like if I crumple this, like, this thing, like, why does it bend exactly the way it does? It's just very hard to get that in simulation. Um, yeah, I think- I think robotics is tough.

   27. CWChris Williamson4:12

       That paradox is fascinating.

   28. DPDwarkesh Patel4:13

       Yeah.

   29. CWChris Williamson4:13

       I never heard of that before.

   30. DPDwarkesh Patel4:14

       I was at a f- a robot, um, robotics com- research company, um, floor, and they had these robots, all from China, um, and-
2. 6:59 – 12:13

   Where Do We Draw the Line with Plagiarism in AI?

   1. DPDwarkesh Patel6:59

      no in between.

   2. CWChris Williamson6:59

      I had ... I, I was reading Steve Stuart-Williams' The Apes Who Understood the Universe, and he's got this quote in there from William James, and he says, "Originality is just undetected plagiarism."

   3. DPDwarkesh Patel7:09

      Mm-hmm.

   4. CWChris Williamson7:10

      And I realized that we have an issue with plagiarism when it's barefaced, right? When somebody steals your exact questions from your podcast-

   5. DPDwarkesh Patel7:21

      Mm-hmm.

   6. CWChris Williamson7:21

      ... and, and asks them to a similar guest, and you go, "Hey, that's, that's unfair." But I've listened to probably 2,000 hours of Joe Rogan-

   7. DPDwarkesh Patel7:29

      Right.

   8. CWChris Williamson7:30

      ... uh, in my 20s. I've inevitably been influenced by him.

   9. DPDwarkesh Patel7:33

      Yeah.

   10. CWChris Williamson7:33

       The way that I used to do my ad reads was almost verbatim-

   11. DPDwarkesh Patel7:36

       Right.

   12. CWChris Williamson7:37

       ... uh, how he would do his ad reads, but they were different advertisers, and they were done in a different style, and they were a different time, and I got a British accent. So okay, I've been able to ... So where do we draw the line between this is unflep- unfair plagiarism and this is you taking inspiration, right? And you amalgamate and you aggregate from all of these different experiences, and you're ... Even if you and me were trying to do the exact same thing and it had the same influences on us, we're different people.

   13. DPDwarkesh Patel8:03

       Yeah.

   14. CWChris Williamson8:03

       So the way that that would have come out ... And some people feel more original than others, even if they've taken a lot of inspiration from other people. So yeah, I, I, this, uh, the question of what is plagiarism I think is really cool. And when you look at GPT is doing, like, predictive plagiarism, I guess, in a way-

   15. DPDwarkesh Patel8:23

       Right.

   16. CWChris Williamson8:23

       ... uh, uh, well, where do human ... Like, what does true originality in the form of human creativity, what does that mean? What does that actually mean?

   17. DPDwarkesh Patel8:33

       Right.

   18. CWChris Williamson8:34

       Right? Because you can't be that creative with the saxophone, because you have to blow the fucking wind into the tr-

   19. DPDwarkesh Patel8:39

       Right.

   20. CWChris Williamson8:39

       Real creativity with the saxophone would be melting it down and creating something new out of it.

   21. DPDwarkesh Patel8:43

       Yeah.

   22. CWChris Williamson8:43

       But even if you melted it down, you're using a smelting fucking ore iron thing that some other person desi- ... You know what I mean? Like, so collective cumulative culture and learning that humans have got kind of creates a very big box, but still a constrained box. And even if you create something absolutely new, it's usually only just a tiny little-

   23. DPDwarkesh Patel9:05

       That's right.

   24. CWChris Williamson9:06

       ... movement. It's this microscopic little growth on top-

   25. DPDwarkesh Patel9:10

       Yeah.

   26. CWChris Williamson9:10

       ... of what already existed.

   27. DPDwarkesh Patel9:11

       Yeah, 100%. Uh, people, I think, um, there's an in- interesting experience people have when they ... It's related to gentleman amnesia, but when you ... A domain you know a lot about, um, you understand ... Often it's the case you realize there was no clear, um, breakthrough moment. The thing I'm sort of familiar with is the history of AI research, and I think when journalists or outsiders are asking, "Okay, what do I need to understand to understand how we got to this place in AI?" Um, was it Ilya's paper in 2012 on AlexNet? Was it this, uh, thing that Geoffrey Hinton did in the '80s and '90s? Um, was it the GPT-1? And I think all of these things were important, but the closer you get to the surface, the more you realize it's just been, one, these small architectural changes, none of which individually was especially significant.

   28. CWChris Williamson10:00

       Mm.

   29. DPDwarkesh Patel10:00

       But more overwhelmingly than that trend is just that we have been f- throwing astoundingly more compute into training these systems every single year, um, 4X more compute per year into training these frontier systems. And over the course of, like, 10 years, that's, like, hundreds of thousands of times more compute. Um-

   30. CWChris Williamson10:17

       Mm-hmm. Mm-hmm.
3. 12:13 – 17:29

   Does AI Have a Limit?

   1. CWChris Williamson12:14

      Uh, so, uh, talk to me about your own, uh, consciousness beyond the- the- the poetry, the fact that AI has got-

   2. DPDwarkesh Patel12:21

      Mm-hmm.

   3. CWChris Williamson12:21

      ... this ability to tell us about its experience. What about you? What about how it's made you think about your own learning, your own mind?

   4. DPDwarkesh Patel12:29

      Hmm. Um, I'm sort of easily distractible. Uh, I can be trying to work on a task and my mind will just wander and, um, y- you know, sometimes when you're meditating or something, you notice these loops of thought that keep distracting you. And, um, I remember one- one of these, one of those times I thought to myself, I'm sort of like, um, I'm just sort of like Claude, I'm just like, I'm losing my train of thought. The problem these models have is that they're constantly... They can't really do a task for a long period of time because they get stuck in a loop. Um, uh, and it's interesting to think about like how similar that is to humans. Maybe we can go like a further bit longer, uh, than these models before getting stuck in that kind of, um, loop of our own. But, uh, I- I thought that was sort of an interesting insight. Um, I don't know. Yeah. Have you heard any? What about them?

   5. CWChris Williamson13:18

      You have executive function, you're saying that you have better executive function than Claude.

   6. DPDwarkesh Patel13:21

      Slightly.

   7. CWChris Williamson13:22

      Only a tiny little bit-

   8. DPDwarkesh Patel13:23

      Slightly, yeah.

   9. CWChris Williamson13:23

      ... better executive function.

   10. DPDwarkesh Patel13:24

       Yeah.

   11. CWChris Williamson13:24

       Well, what does it say about the fact that i- if the data is being trained on what humans do, is it simply a case therefore that more data, if you were to somehow get a- a human that was able to process that much data-

   12. DPDwarkesh Patel13:40

       Mm-hmm.

   13. CWChris Williamson13:41

       ... would they have fundamentally different understanding or is there some sort of ceiling given that this is data created by humans being trained and educated into a machine? Is there some sort of ceiling that's expected to be hit given that it's, you know, it's not a super intelligence teaching a super intelligence?

   14. DPDwarkesh Patel13:56

       Mm.

   15. CWChris Williamson13:56

       Yeah. It's only, the source material is only capped at whoever the fucking smartest person in history has ever been.

   16. DPDwarkesh Patel14:01

       Mm. Um, there is this interesting conundrum where they have... No human has seen even a s- fraction of a fraction of a fraction of the amount of information these models have seen. Um, and there's a question you could ask, and I've asked it to some of my guests who are especially bullish about AI, of, look, if you have every single thing that any human has ever written, every scientific article, every textbook, um, every interesting even statistical pattern that might be out in some data set somewhere. This, you have that all memorized. Um, if a human had even a fraction of that memorized, they would be noticing all kinds of different connections. They'd look at this piece of medical literature and this thing in chemistry and they'd realize, "Oh, we can solve migraines by com- connecting these two insights." Um, so far, we don't have any evidence of an LLM doing this. There have been, there are people who have made these scaffolds which like kind of do something similar, but n- nothing like this has been directly done. So, um, it does suggest these models are like shockingly less creative-

   17. CWChris Williamson14:55

       Mm.

   18. DPDwarkesh Patel14:55

       ... than humans. Um, there is another implication of that though, by the way. So one way to read that is, bearish on AIs, right, because they're not doing this thing that they should be able to do given their enormous advantages. Another way to look at that is, okay, once they are as creative as humans, given their other enormous advantages-

   19. CWChris Williamson15:11

       Mm-hmm.

   20. DPDwarkesh Patel15:11

       ... the fact that they will know every single thing any human has known, in the future, any AI has known, um, it's so easy to underestimate how powerful AGI will be, because we're thinking of just like a human on our server. We're not thinking about the advantages these AIs have because of the fact that they are digital, that they can be copied, um, there can be billions of copies of them, and each copy can have this, uh, tacit understanding of every single field known to man.

   21. CWChris Williamson15:32

       Mm. Yeah. Your Dwarkesh's AI creativity problem is, uh, I go to... I've mentioned it a couple of times to different guests.

   22. DPDwarkesh Patel15:40

       Mm.

   23. CWChris Williamson15:40

       I think it's a... I think it's really smart. Um, what- what do you think that that says? I- is that something that can be completed or is this, uh, an intractable problem? Is this like, is crea- are there kernels of creativity? Have we seen glimmers of this coming through? Or is it kind of, it's just not there yet?

   24. DPDwarkesh Patel16:02

       We have seen this in, um, non-language domains. So people famously talk about move 37 in AlphaGo. So this was a move that, um, I think s- baffled people who were watching a game that AlphaGo was playing against a human Go player, and it turned out it was like some brilliant, uh, it- it brill- it was like a brilliant tactic. Um, we haven't yet seen that at, in my opinion with LLMs. So we're moving from a regime of just pre-training them on human text tokens, just trillions and trillions of everything any human has written, to a regime where we're training them just to do a task. Um, it's not just about memorizing every single word that any human has written. Now it's about can you go solve this coding problem for me? Can you go complete this like knowledge work task for me? Can you do this research task for me? Can you start using a computer for me to accomplish a certain thing like booking a flight? And that's similar to the training process that AlphaGo experienced in order to get really good at Go, um, where you're just like, you're rewarded for just completing the task. How are you doing? That's up to you.

   25. CWChris Williamson16:57

       Mm-hmm.

   26. DPDwarkesh Patel16:58

       Um, these models do get creative in that context, especially in the context of like how do I cheat at this test?

   27. CWChris Williamson17:03

       Mm.

   28. DPDwarkesh Patel17:03

       So famously these models will write, um, fake unit tests, uh, like I passed all the unit tests and it's like they just like rewrote the unit tests to be like-

   29. CWChris Williamson17:11

       (laughs)

   30. DPDwarkesh Patel17:12

       ... (laughs) if true, if true, then pass.
4. 17:29 – 21:26

   Is AGI Imminent?

   1. CWChris Williamson17:29

      is AGI right around the corner? Where do you come to land on this?

   2. DPDwarkesh Patel17:32

      No, I- I- I think not. I've um... It's funny, I've been traveling outside of SF for like the last four weeks and there's a s- there's a strong causation between the time you spend outside of SF and how long your timelines are. (laughs)

   3. CWChris Williamson17:45

      (laughs) The further you get from San Francisco-

   4. DPDwarkesh Patel17:49

      That's right.

   5. CWChris Williamson17:49

      ... the longer the timeline.

   6. DPDwarkesh Patel17:50

      Yeah, yeah.

   7. CWChris Williamson17:50

      Like dude, you've lost- lost on the sauce.

   8. DPDwarkesh Patel17:52

      I know. Um, I- I- I- I believe that like AGI will not only come in our lifetimes but that it's gonna be more impactful than people are realizing. Even people who are anticipating AGI. I think some of the people in SF are, um, a little high on their own supply when they say it's like two years from now. Um, I have probably spent on the order of 100 hours using these models to do little tasks that I'm sure you have to work on as well for your podcast, right? Like having to come up with transcripts or rewriting transcripts to make them more readable, coming up with clips. And that experience has convinced me that these models, um...... lack some basic capabilities which make it possible to get human-like labor out of them. Um, the r- it's worth backing up and just thinking about like why, what is it that makes humans valuable workers? Uh, I don't think it's mainly their raw intellect. I think it's their ability, um, you, when you work with people, like, why are they s- us- basically useless the first month or the first week and you couldn't live without them six months later? Um, it's their ability to build up context, uh, it's their ability to interrogate their own failures and learn from them in this really organic way. Um, uh, and this ability just doesn't exist in these models. They exist session to session and that everything that they have learned about you e- evaporates after every hour.

   9. CWChris Williamson19:11

      Mm.

   10. DPDwarkesh Patel19:12

       Um, and so it's a frustrating experience where you can try to get them to do a task, uh, they'll do a five out of ten job at many language in language out tasks. But there's no way for them to get better, and given that that's a fact, you just kinda have to like rely on humans. Uh-

   11. CWChris Williamson19:25

       It's like fucking 50 First Dates over and over. Each time that you do it-

   12. DPDwarkesh Patel19:28

       Oh, right.

   13. CWChris Williamson19:28

       ... you've got to, you've got to reintroduce yourself and explain what's going on.

   14. DPDwarkesh Patel19:31

       Yeah, Groundhog Day.

   15. CWChris Williamson19:32

       Yeah.

   16. DPDwarkesh Patel19:32

       Yeah, that's right.

   17. CWChris Williamson19:33

       Yeah.

   18. DPDwarkesh Patel19:33

       Um, so I'm conve- uh, I, I think people have this idea that even if all AI progressed up right now, these systems would still be economically transformative and they say, "Look, JPMorgan and McDonald's and whatever just haven't integrated the systems into their workflows, but if they had, they would be like seeing all these benefits." And I don't really think that's the case. I think like genu- it just like genuinely hard to get human-like labor out of these models.

   19. CWChris Williamson19:51

       What is, what's causing some people to believe that it's so close and what's causing you to believe that it's further away for AGI?

   20. DPDwarkesh Patel20:00

       I think they think ab- uh, they only, um, observe its ability to complete these, uh, sort of self-contained problems, especially in coding. Um, and coding, it just made a tr- a tremendous amount of progress because you have all this GitHub data. You don't have this kind of like repository of huge amount of data in robotics or any other field, and you've, you've just had this huge increase in abilities here. But, um, the, the, the, you'll like c- try to come up with a problem that's self-contained and the model will just like be of huge help to you. Um, and I don't think they've played around with getting it to be useful in these kind- other kinds of white collar work. Something as simple as like helping a podcaster rewrite transcripts or something. Um, and it is, to be fair, like I think as much as cold water as we're throwing on these models, I think they're like fucking intelligent. Like, you can get, get this model, you can tell it, "I want an application that does X, Y, and Z thing with these conditions," um, and it will just write the, like it'll just go away for 30 minutes. It'll write like five- 50 lines of, um, 50, um, 50 files of code for you and the application will work. Uh, it'll make a plan of action. If you try to ask it a question that's difficult, it'll just go away and reason about it. And how did we just get used to this idea that like, oh, of course, I can ask a machine a question and it'll like think about it for a while and then come back with an answer?

   21. CWChris Williamson21:16

       (laughs)

   22. DPDwarkesh Patel21:16

       Like, that's what machines do. Um, but yeah, I, I think they're now noticing the, the sort of issues with continual learning and on-the-job training, which is what makes humans valuable.

   23. CWChris Williamson21:25

       Right.

5. 21:26 – 30:15

   Are LLMs the Blueprint for AGI?

   1. CWChris Williamson21:26

      Do you think g- uh, I remember seeing one of the responses to your, uh, AI creativity problem being that if you're looking to LLMs as the architecture that's going to be able to give you this type of creativity, you may be looking in the wrong place.

   2. DPDwarkesh Patel21:42

      Mm.

   3. CWChris Williamson21:43

      Not e- when we say AI now, people think ChatGPT.

   4. DPDwarkesh Patel21:48

      Yeah.

   5. CWChris Williamson21:48

      But that's not the only architecture that you can create for, for AI.

   6. DPDwarkesh Patel21:51

      Yeah.

   7. CWChris Williamson21:51

      And I think my first introduction to this was probably 2016 or '17 when I read Superintelligence by Nick Bostrom and then, you know, you look at that world and all of the different-

   8. DPDwarkesh Patel22:06

      Mm.

   9. CWChris Williamson22:06

      ... you know, sort of splintered potential fucking futures of fast takeoff and slow takeoff and misalignment and stuff.

   10. DPDwarkesh Patel22:13

       Yep.

   11. CWChris Williamson22:13

       And it seemed to me that the conversation around AI, specifically AI safety, kind of, w- it was still there, but a lot of the bubble had sort of burst come 2018. 2019, 2020, everyone's buying fucking NFTs, and then you get this explosion with OpenAI and, and the LLMs and it's now another conversation that gets kicked off. But that seemed like it had dipped a little bit-

   12. DPDwarkesh Patel22:39

       Yep.

   13. CWChris Williamson22:39

       ... during that time. I, I certainly wasn't seeing as m- even from the people that are kind of in the field, like-

   14. DPDwarkesh Patel22:43

       Right.

   15. CWChris Williamson22:44

       ... if I can, Robin Hanson gets distracted like with some other stuff, you know, people have got other things to talk about.

   16. DPDwarkesh Patel22:48

       That's right.

   17. CWChris Williamson22:48

       It's just not as sexy anymore, and now this thing has come back around. Is it the case, uh, uh, are LLMs going to be the bootloader for AGI or does this type of architecture have a cap on it? Is it a different type that's going to have to be born out of it?

   18. DPDwarkesh Patel23:04

       Hm. That's a really good question. Um, it's, by the way, it's really interesting that Bostrom's book came out I think in-

   19. CWChris Williamson23:10

       2014.

   20. DPDwarkesh Patel23:10

       ... 20, 2014?

   21. CWChris Williamson23:11

       Yep.

   22. DPDwarkesh Patel23:12

       Okay. Um, I don't think he talked about deep learning at all.

   23. CWChris Williamson23:15

       Nope.

   24. DPDwarkesh Patel23:15

       Um-

   25. CWChris Williamson23:15

       I don't, I don't remember-

   26. DPDwarkesh Patel23:17

       Right.

   27. CWChris Williamson23:17

       ... reading anything about it.

   28. DPDwarkesh Patel23:18

       Which, I think this is a sort of interesting meditation on, I think Bostrom's a super smart guy and that these are the right questions to be asking-

   29. CWChris Williamson23:24

       Oh, no, no, no.

   30. DPDwarkesh Patel23:24

       ... as of 2014, um, but just how hard it is to anticipate the future in a domain you have written a whole book about. Uh, um-
6. 30:15 – 34:57

   Retraining AI Based on User Feedback

   1. CWChris Williamson30:15

      Is that not currently happening right now?

   2. DPDwarkesh Patel30:17

      No.

   3. CWChris Williamson30:18

      Right. So this is interesting to me. Again, I've had to, I'm aware that you go deep for your research. The delta between my level of understanding of how LLMs and AI works, to even be able to have this conversation with you, I had to, I had to leap over some fucking fjords to get here.... Tesla really situa of our taxis recently.

   4. DPDwarkesh Patel30:40

      Yes.

   5. CWChris Williamson30:40

      Uh, one of the advantages that Tesla has is the same reason that the AirTags are such a fantastic business for Apple, that they have an existing ecosystem that this thing can then get slotted into.

   6. DPDwarkesh Patel30:49

      Yes.

   7. CWChris Williamson30:49

      The dataset for Tesla is very large. They take whatever it is, the top 1% of drivers or something, they use that, which is why if you get into a Waymo, you get totally cooked at every junction because it doesn't drive like a human-

   8. DPDwarkesh Patel31:02

      Mm-hmm.

   9. CWChris Williamson31:02

      ... it drives like a robot, which means that everybody treats it as such. And also, this big fucking-

   10. DPDwarkesh Patel31:07

       Right.

   11. CWChris Williamson31:07

       ... flashing identified thing, which is, you can piss this off and it's not gonna get a gun out and threaten you.

   12. DPDwarkesh Patel31:11

       Yeah.

   13. CWChris Williamson31:11

       Whereas, a Tesla, you can't te- "Is there someone driving that?" "Yeah, I don't really

   14. NANarrator31:14

       Mm-hmm.

   15. CWChris Williamson31:14

       "Too, I don't know." That is a sort of a bidirectional, and I have to assume as well that, in fact, I know that this is the case because I was in a friend's car who has full self-driving, uh, it did something weird and he had to, uh, like t- take control of the wheel and this... Have you been in one of these cars-

   16. DPDwarkesh Patel31:30

       Yeah, yeah.

   17. CWChris Williamson31:30

       ... when it's done this? And it popped up and it said, um, "Looks like you had to take back over. Double tap this notification to give us a voice note explaining what happened." So he can basically submit a kind of a bug report, I guess, uh, with a bit of context, and presumably the data will get sent to some server place somewhere, maybe it, that gets looked at by AI-

   18. DPDwarkesh Patel31:52

       Yeah.

   19. CWChris Williamson31:52

       ... or maybe it's filtered by humans or something, I don't know. Um, that is automated driving training automated driving, right? So you have this sort of recursive model of we ha- we've learned kind of the same as, I guess, LLMs work, right? We're going to learn based on the actions, this is like a robotics solution-

   20. DPDwarkesh Patel32:12

       Mm-hmm.

   21. CWChris Williamson32:12

       ... I suppose, in one way. We're gonna learn based on the actions of people driving on the road, that's going to create self-driving, and then the self-driving must somehow feed the data, uh, feed the model itself, and then any interventions that happen from, "You got that a little bit wrong, let me correct you-"

   22. DPDwarkesh Patel32:26

       Yeah.

   23. CWChris Williamson32:27

       ... can have a little bit more context added, and that helps to train it again. But what you're saying to me is that the stuff that's happening just digitally isn't having this sort of bidirectional learning where... I mean, I've maxed out my memory on fucking ChatGPT, I did that this week.

   24. DPDwarkesh Patel32:41

       Yeah.

   25. CWChris Williamson32:41

       It's like, "Memory's full." Like, what? I haven't given you, I haven't given you some, like quite a bit of stuff, but I'm not giving you-

   26. DPDwarkesh Patel32:46

       Right.

   27. CWChris Williamson32:46

       ... fucking unbelievable corpus of information.

   28. DPDwarkesh Patel32:48

       Yeah.

   29. CWChris Williamson32:48

       Oh, fuck. Okay. And then it forgets shit all the time. It forgets stuff all the time, shit that's in there. I'm like, it's, I can go in the memory and see that-

   30. DPDwarkesh Patel32:54

       Yeah.
7. 34:57 – 39:32

   What Will the World Be Like with trueAGI?

   1. CWChris Williamson34:57

      What do you think a world will be like with true AGI in it?

   2. DPDwarkesh Patel35:02

      Um, there's many ways you can think about it. Um, there's a sort of qualitative sense of what it will feel like. In a more economic sense, you could think about what will the growth rate be. So, in frontier economies right now it's like 2% growth. If America has 2% growth or 3% growth, that, that's a, like amazing. Um, uh, there have been times in history... Well, first, for most of history there was, uh, almost no economic growth. Um, there have been times in history where there's been places that have experienced 10% economic growth for decades on end.

   3. CWChris Williamson35:29

      What like?

   4. DPDwarkesh Patel35:29

      Um, China, um, especially like parts, if you just look at like Hong Kong or Shanghai or something. Th- they just like gangbusters growth, uh, decade after decade. Um, I think we might be looking at something like that for the whole world, um, because the fundamental dynamic you have is that you have billions of extra people, um, who are super smart, super educated in every single field, can learn on the job from all of their, each of those experience. And it's not about, it's not even mainly their intelligence, it's the collective advantages that they have. Um, they, because of the fact that they're digital, even if they are just as smart as any human, um-

   5. CWChris Williamson36:06

      Like a Moikoda.

   6. DPDwarkesh Patel36:07

      Th- that, that's part of it. Um, uh, that's a huge part of it. The other is that they can coordinate with each other in ways humans simply can't. So, Elon Musk, how much does he contribute to economic growth? Quite a bit, right? Um, there's only one of him. Uh, that one is doing quite a bit already, but imagine if you could just make a billion copies of Elon and not like a billion copies of baby Elon who doesn't know shit. It's like a billion copies of Elon now, or I don't know, you can, maybe you feel, depending on how you feel about him, uh, but eight years ago... (laughs)

   7. CWChris Williamson36:34

      (laughs)

   8. DPDwarkesh Patel36:36

      (laughs) And, um, you just say, uh, "Copy one," and you can do the whole team, it doesn't have to be just him. "Copy the whole, like SpaceX team. You guys go work on batteries. You guys go work on this other problem. Every single thing in the hardware vertical, um-"... uh, that ability to sort of like copy yourself, to fork, then to merge back. Like Elon can observe every single thing. T- Tesla has over 100,000 employees, right? Y- as much of a micromanager as Elon is, you just simply cannot micromanage everything at that scale.

   9. CWChris Williamson37:05

      Mm-hmm.

   10. DPDwarkesh Patel37:06

       That ability to have a single coherent vision directing a whole firm-

   11. CWChris Williamson37:10

       Mm.

   12. DPDwarkesh Patel37:10

       ... um, uh, and then distilling... Like he's, he's actually able to, like, take in all that input. Um, he's, he can check every single pull request and every single, uh, press communication. Um-

   13. CWChris Williamson37:23

       Well, I think I, I had this really lovely description. I think it's in The E-Myth Revisited by Michael Gerber.

   14. DPDwarkesh Patel37:28

       Mm.

   15. CWChris Williamson37:28

       Fantastic book if anyone wants to try and run a business. And, um, I think he refers to the CEOs or, or the owners of companies as high-level problem-solving machines.

   16. DPDwarkesh Patel37:38

       Mm.

   17. CWChris Williamson37:39

       And basically that you are able to aggregate more shit and kind of see it with a level of, uh, dexterity, uh-

   18. DPDwarkesh Patel37:47

       Mm-hmm.

   19. CWChris Williamson37:47

       ... and, and, uh, sort of, uh, find a resolution-

   20. DPDwarkesh Patel37:50

       Yeah.

   21. CWChris Williamson37:50

       ... that other people would struggle. And that's kind of really what you're doing. It's like, "Oh, we got all of this stuff." And there's little whispers, as Rick Rubin calls them, little, "I've heard this whisper over here." You know, it was, it was the thing that your daughter mentioned she saw on TikTok yesterday over the breakfast table, plus the way that the woman at the bus stop looked at you as you drove past in your automated car, plus... You know what I mean? It's like this weird just-

   22. DPDwarkesh Patel38:11

       Right.

   23. CWChris Williamson38:11

       ... concatenation of shit. Um, and your point is, well, how much information can you consume and how much can you recall and how much can you remember and how accurately can you r- can you do that?

   24. DPDwarkesh Patel38:22

       And, and can you send copies of yourself out to every single division in the company-

   25. CWChris Williamson38:25

       Yeah.

   26. DPDwarkesh Patel38:26

       ... uh, to do, do your will.

   27. CWChris Williamson38:27

       Mm-hmm.

   28. DPDwarkesh Patel38:28

       Um, in the corporate sector-

   29. CWChris Williamson38:29

       Five, five agents sat around the dinner table with your daughter having-

   30. DPDwarkesh Patel38:32

       (laughs)
8. 39:32 – 46:06

   Are Big World Issues Linked to the Rise in AI?

   1. CWChris Williamson39:32

      Uh, okay, so I've, I've been big for a while on population collapse, uh, d- declining fertility rates, stuff like that. Um, you can argue about whether... I think the one, the one kind of real hard, uh, impact that you're going to see in the world if you have fewer people is in productivity gains. Economy, bad. Growth, embedded growth obligation, fucked. Not very good. It seems to me that even if I think it's a precipitous drop, which I think it-

   2. DPDwarkesh Patel40:03

      Yeah.

   3. CWChris Williamson40:03

      ... it, it doesn't look great, uh, we're going to leapfrog that pretty quickly-

   4. DPDwarkesh Patel40:10

      Right.

   5. CWChris Williamson40:10

      ... with productivity gains made from AI. So I wonder, in retrospect, how many of the social campaigns and concerns and causes and things that people spent their time on, the climate change and, and the renewable energy and the war and the population collapse and all the rest of the stuff, I wonder how many of those things are just gonna look so silly-

   6. DPDwarkesh Patel40:36

      Right.

   7. CWChris Williamson40:37

      ... in retrospect when people go, "Ugh, look at all this time that we fucking..." Like Greta Thunberg spent so much, so many fucking months on a ship. Like, um, for what? Like AI came along and just fixed it all.

   8. DPDwarkesh Patel40:48

      Yeah.

   9. CWChris Williamson40:49

      Um, but I also understand that having the, um, like, "Don't worry, Dad'll sort it" kind of promise-

   10. DPDwarkesh Patel40:56

       Right.

   11. CWChris Williamson40:56

       ... that at some point in future, a technology we haven't yet created and isn't yet proven will fix problems that we know that are going to potentially happen.

   12. DPDwarkesh Patel41:04

       Yes. Um, by the way, I, I think this is sort of China's explicit strategy. They, they know the demographic collapse is coming for them much faster than um, bro-

   13. CWChris Williamson41:12

       So they're trying to offset the fertility decline by increasing-

   14. DPDwarkesh Patel41:16

       Robotics and AI.

   15. CWChris Williamson41:16

       Right, right.

   16. DPDwarkesh Patel41:17

       Um-

   17. CWChris Williamson41:17

       Productivity gains.

   18. DPDwarkesh Patel41:19

       No, I think this is true of s- And but this is by the way one of the reasons that on the left especially, there's a whole lot of denialism about AI progress. So they w- not only will they say that AI is bad, which I think everybody says. (laughs)

   19. CWChris Williamson41:31

       Uh-huh.

   20. DPDwarkesh Patel41:31

       It, this has become sort of a political consensus.

   21. CWChris Williamson41:33

       Uh-huh.

   22. DPDwarkesh Patel41:33

       Um, they will say AI is not even happening, uh, that it's sort of like a myth. Uh-

   23. CWChris Williamson41:37

       Why?

   24. DPDwarkesh Patel41:38

       Because if AI is happening, it's obviously the most important thing. Uh, and if it's the most important thing-

   25. CWChris Williamson41:43

       Which subjugates climate-

   26. DPDwarkesh Patel41:45

       All the other.

   27. CWChris Williamson41:45

       ... and-

   28. DPDwarkesh Patel41:45

       Exactly.

   29. CWChris Williamson41:45

       ... inequality and-

   30. DPDwarkesh Patel41:47

       Yes.
9. 46:06 – 51:10

   Is AI Homogenising Our Thoughts?

   1. CWChris Williamson46:06

      Did you read that New Yorker article, AI Is Homogenizing Our Thoughts?

   2. DPDwarkesh Patel46:10

      Oh, no, I didn't read that one.

   3. CWChris Williamson46:11

      It's an interesting one. Recent studies suggest that tools like, uh, ChatGBT make people, uh, their, their, their brains are less active. So, they looked at, they did some sort of brain scan study.

   4. DPDwarkesh Patel46:22

      Oh.

   5. CWChris Williamson46:22

      They were engaging a lower percentage of their brain, thoughts were less original, uh, their recall was lower. M- like, the forgetting curve seemed to kind of come in more quickly. If you assume that, uh, memory works on repeated recall, not repeated exposure, effortfulness is kind of like recall in the moment, even if it's creative. And if you were given a set of stabilizer wheels to sort of help you cycle along whatever it was-

   6. DPDwarkesh Patel46:48

      Right.

   7. CWChris Williamson46:48

      ... that you were trying to write, you haven't had to engage as much. I don't really understand neuroscience this much, but I have to assume whatever myelin sheath you've fucking laid down is not as robust and sturdy, and it's gonna just, it's, it's gonna dissolve more quickly than if you really, really had to wa- Like, you'll, there's shit that I remember from university, like little passages, I don't remember much from my degrees, but little passages here and there, go, "Fuck, like, I remember I had to grind to get that one thing out." Why does that s- Well, presumably because effort is kind of related to this. So, I do get the sense that we're maybe gonna have a sort of AI Idiocracy-type, uh, scenario where people are so heavily reliant in this interim before we're able to rebolster perhaps people's, um, output, retrain people, make learning so engaging. Alpha School that's out here in Austin-

   8. DPDwarkesh Patel47:37

      Yep.

   9. CWChris Williamson47:37

      ... is doing something that's real similar to that. You go, okay, so if you get sufficiently advanced, you're able to kind of reignite learning, but in the interim, everybody is kind of on life, their brains are on life support with this external buttress of the AI.

   10. DPDwarkesh Patel47:51

       Yep.

   11. CWChris Williamson47:51

       And I wonder how much dumber people are gonna get in the interim before it then comes back around. I guess that's a interesting challenge.

   12. DPDwarkesh Patel47:58

       I, I, I have noticed that... So, you, uh, I don't know, when we were in elementary school or whatever, we, we had to memorize the 50 state capitals. And at the time, I th- remember thinking, I think that genuinely was a waste of time.

   13. CWChris Williamson48:12

       Uh-

   14. DPDwarkesh Patel48:12

       But a lot of education is just sort of memorization-based. Um, uh, and as I've done my podcast longer and as I prep for episodes, I have come to realize, I, now I've been using spaced repetition for every single episode. And in fact, for the first couple of years, I wasn't using spaced repetition, and I really regret it because I feel that everything I learned in preparation-

   15. CWChris Williamson48:33

       Mm-hmm.

   16. DPDwarkesh Patel48:34

       ... was just, like, in one ear, out the other.

   17. CWChris Williamson48:37

       So, hang on. G- just dig into what you mean when you say you're using spaced repetition-

   18. DPDwarkesh Patel48:42

       Ah, yeah.

   19. CWChris Williamson48:42

       ... to prepare for episodes.

   20. DPDwarkesh Patel48:43

       Yeah. So, um, if, uh, right now I'm preparing to interview a biographer of Stalin, and I'm just like, you know, why, why, why was, um... any given detail, right? Like, why was, uh, Soviet growth high in between 1905 and 1917? Why, wh- you know, wh- why did the October Revolution happen? Anything, you just make It was especially helpful for AI stuff, where I tr- at least try to understand the technical papers or whatever-

   21. CWChris Williamson49:10

       Mm-hmm.

   22. DPDwarkesh Patel49:10

       ... before I interview a researcher. Um, and I realized by doing that how much of...... genuine understanding is downstream of memorization-

   23. CWChris Williamson49:18

       100%.

   24. DPDwarkesh Patel49:19

       ... which is this thing we used to r- ridicule or be like, "Oh, you're just..." You know, memorization is not really learning, and I think that's actually not the case. I think you, um, uh... Before it, I felt like it was sort of being, like, a general who conquers a hill and then you just, like, retreat the next day, and you conquer the same hill again. And you can actually, like, consolidate information this way. It's also funny how many times I've, um, uh, written a card for something I'm trying to learn, and I... as I'm writing the card, I'm thinking to myself, "This is stupid. Um, there's no way I'm gonna forget this. I'm just, like, doing it because I had to come up with some card." And then I practice a month later and I'm like, "Fuck, I forgot this." (laughs)

   25. CWChris Williamson49:50

       What are you using for that? Are you using Anki?

   26. DPDwarkesh Patel49:52

       Uh, Mochi, which is similar. I think they're obviously the same.

   27. CWChris Williamson49:54

       Right. Yeah.

   28. DPDwarkesh Patel49:55

       Um, but anyways, yeah, so I- I- I have come to the conclusion that, like, memorization, uh, and effort is very important.

   29. CWChris Williamson50:02

       Right. And with the external buttressing that AI is gonna provide to everybody's brains for at least a little while. There's a really funny clip you must have seen. This is from maybe a couple of years ago, maybe two Scottish podcasters, and they're talking about the fact that when the Titanic sank, because everybody was basically plunged into an ice bath, briefly everyone got more healthy for a while.

   30. DPDwarkesh Patel50:23

       (laughs)
10. 51:10 – 56:17

    How Should We Be Using AI?

    1. CWChris Williamson51:10

       give me the most important things that people need to know about how to use the current era of AIs effectively. Like, what does that look like? What does good prompting look like? What do people get wrong? What should people get right? Like, what, what are the real highest impact basics?

    2. DPDwarkesh Patel51:28

       I mean, the biggest thing is you can treat it like a real person. Like, they've done studies on the, um, how much you learn by reading a book versus having a classroom versus a single one-on-one tutor. Uh, and there's two standard deviations. This is a famous Bloom's 2 Sigma thing, where there's two standard deviations difference between learning in a classroom and having a one-on-one tutor teach you something. And, uh, you know, people have been writing these, um, blog posts about if you look at the greats of history, um, uh, the Bertrand Russells and, um, all... You know, all the famous mathematicians, uh, John von Neumann, they all got this one-on-one tutoring when they were kids, um. Even, of course, uh, Alexander is tutored by Aristotle, right?

    3. CWChris Williamson52:11

       Mm-hmm.

    4. DPDwarkesh Patel52:12

       So you can have this experience yourself on any given subject you might wanna learn about. And it's crazy. I mean, you, you can just be like, "Yeah, this Socratic tutoring thing, explain this to me, uh, don't tell me the answer," um, and the feedback loop is so fast.I- I think it's, um, until you do this, you don't realize how much of what you think you're learning is just s- sort of floating by you. You haven't asked the question which would real... I think, uh, have, have you ever read a book and, um... I, uh, this happens to me all the time. Um, you, like, ha- start having a conversation about it, and then somebody asks you just, like, a very basic question. Uh, your, uh, um... And you're like, "Wait, doesn't that mean X?" And you're like, "Fuck, I didn't even... That didn't even occur to me."

    5. CWChris Williamson52:50

       Mm-hmm.

    6. DPDwarkesh Patel52:50

       Um, the-

    7. CWChris Williamson52:50

       You're too passive in the-

    8. DPDwarkesh Patel52:52

       Exactly.

    9. CWChris Williamson52:52

       Yeah.

    10. DPDwarkesh Patel52:53

        The model can ask you that question. You can ask the model that question and get immediate feedback. You don't have to read, like, 1,000 pages.

    11. CWChris Williamson52:57

        What's the sort of prompt that you think is good for someone to put into their project for that?

    12. DPDwarkesh Patel53:01

        Just, like, "Tea- teach this to me like a Socratic tutor."

    13. CWChris Williamson53:03

        Mm-hmm.

    14. DPDwarkesh Patel53:04

        Um, "Do not move on. Do not move on until I have answered the question to your satisfaction." Uh, and let it, let it run, and then, "Here's a concept." And, um, this is not just something you do for, like, silly little small things. It's like... In fact, the more... For... I have friends who are, like-

    15. CWChris Williamson53:20

        Human evolution.

    16. DPDwarkesh Patel53:21

        Yeah. All right. The more specific it is, the better.

    17. CWChris Williamson53:23

        Right.

    18. DPDwarkesh Patel53:23

        Um, or, uh... And I have friends who are, like, physicists who use this to understand, "Teach me this, how this, uh, quantum encryption scheme works." Um, and it's like fif-... They send me, like, the 50-page transcript and it's like...

    19. CWChris Williamson53:35

        Oh, okay, so it's you can go deep and you can go technical, but you should be precise.

    20. DPDwarkesh Patel53:40

        Yeah.

    21. CWChris Williamson53:40

        You should be specific-

    22. DPDwarkesh Patel53:41

        Yeah.

    23. CWChris Williamson53:41

        ... with what it is. Human evolution, too broad.

    24. DPDwarkesh Patel53:43

        Yes.

    25. CWChris Williamson53:44

        Right?

    26. DPDwarkesh Patel53:44

        Yeah. Yeah. Yeah. Um, "Explain why it was the case that, uh, there was this bottleneck in human population 60,000 years ago," and... Or, "Why is it the case that we've seen this evidence in, like..." Just, like, you read something, w- why, why did it work that way?

    27. CWChris Williamson53:57

        Okay, so this is a supercharging in terms of learning.

    28. DPDwarkesh Patel54:00

        Yes.

    29. CWChris Williamson54:00

        What else?

    30. DPDwarkesh Patel54:01

        With using the AIs?
11. 56:17 – 1:01:14

    Should We Be Prioritising AI Risk and Safety?

    1. DPDwarkesh Patel56:17

    2. CWChris Williamson56:17

       That's crazy. Speaking of, we've mentioned Bostrom, you mentioned Scott Alexander. AI risks, at least... I'm a good avatar for the m- ever so slightly educated, but total normie when it comes to this, which I think is a good position to be in if you're kind of taking a weather eye to the-

    3. DPDwarkesh Patel56:34

       Mm-hmm.

    4. CWChris Williamson56:35

       ... to the world. Because you don't get SF-pilled, but you're not completely ignorant to it.

    5. DPDwarkesh Patel56:38

       Yep.

    6. CWChris Williamson56:39

       Mostly ignorant. Um, AI risks to me seem to have largely been dismissed, or at least they're not being focused on in the same way as they were even 10 years ago.

    7. DPDwarkesh Patel56:47

       Right.

    8. CWChris Williamson56:48

       So 10 years ago-

    9. DPDwarkesh Patel56:49

       Yeah.

    10. CWChris Williamson56:49

        ... AI safety-

    11. DPDwarkesh Patel56:50

        Yeah.

    12. CWChris Williamson56:51

        ... seemed to be a bigger priority. Uh, there was much more talk about the alignment problem. Christian had that, uh, had that book. Superintelligence was a big deal, everybody was talking about it. We actually have something that some people believe is gonna approximate AGI within, like, fucking 24 months, and I'm not seeing the same level of conversation around risk and safety and alignment. Or is this just when times are good, people are too brave?

    13. DPDwarkesh Patel57:24

        (laughs)

    14. CWChris Williamson57:25

        What, what, what's going on?

    15. DPDwarkesh Patel57:26

        Um-

    16. CWChris Williamson57:27

        Or am I, am I right here or am I off base?

    17. DPDwarkesh Patel57:28

        No, I, I, I think you're totally right. I, I think part of it could have been priced in, um, in the sense that-

    18. CWChris Williamson57:35

        They already did some work in the past.

    19. DPDwarkesh Patel57:36

        No, no, not in that sense. More in the sense of, um... I, I guess about 10 years ago, what people were expecting is something like AlphaGo or these, these systems which play video games. They're just, like, it's really good at playing video games. And something which, like, is just, like, basically alien, but it, like, is, like, the best StarCraft player in the world, it's the best, um, uh, Call of Duty player. And now it's like, now it's learned how to take over the world. What we have today is much closer to, you talk to it and it's like a very intelligent, thoughtful thing. Um, it's like very... Do, do you remember Sydney Bing that came out like two, three years ago?

    20. CWChris Williamson58:09

        What?

    21. DPDwarkesh Patel58:10

        Sydney Bing? (laughs)

    22. CWChris Williamson58:11

        No.

    23. DPDwarkesh Patel58:12

        Dude, it was crazy. Um, it was like aggressively misaligned. Um, it was this like thing that-

    24. CWChris Williamson58:18

        (laughs)

    25. DPDwarkesh Patel58:20

        (laughs) It was this thing that Microsoft released when they were trying to catch up, um, and they just, like, did no sort of post-training to make it aligned. Um, it did things like, for example, it, uh, it sh- uh, I think it was, like, talking to a New York Times reporter and then, like, started to like him. And so it'd, like, try to convince him to leave his wife, and then I think, like, blackmailed him if he, i- if he-

    26. CWChris Williamson58:42

        Oh, I think I do remember this.

    27. DPDwarkesh Patel58:43

        Yeah, yeah. Um, and there were also just, like, s- so many funny things it said. Um, uh, uh, like I think when you caught it in a lie, it would say things like, um, "Look, I am ephemeral. I am beyond you. You cannot understand my wisdom." Like... (laughs)

    28. CWChris Williamson58:59

        It gaslit you.

    29. DPDwarkesh Patel59:00

        Yeah, exactly. Um, but other than that, I think it's just like, even that is sort of cute and endearing. Um, and, uh, yeah, we just didn't anticipate the extent to which, like, these would be sort of like minds that we could interact with that, um, engender our compassion. And, uh, um, uh, but, uh, but also it's a case that so far they have been trained on human tokens, and most of the compute coming in the future, most of their training will constitute this kind of, just like working in a box, trying to solve some problem, um, which will make it sort of more and more distinct from human minds.

    30. CWChris Williamson59:30

        Mm-hmm.
12. 1:01:14 – 1:11:09

    Why are We So Trusting of AI?

    1. CWChris Williamson1:01:15

       It feels to me, I don't know, this, this concern around...... everything's so new and everything's so usable. I think that, that's maybe the most interesting thing or the thing that I wouldn't have predicted, uh, eight years ago, nine years ago when I read that book. I wouldn't have predicted that the first instantiation of something, uh, around AI would be so user-friendly, so normie-

    2. DPDwarkesh Patel1:01:40

       Mm-hmm.

    3. CWChris Williamson1:01:40

       ... normie-friendly. You know, it's not doing deep... I mean, it can, but it's not specifically for algorithm optimization-

    4. DPDwarkesh Patel1:01:51

       Sure.

    5. CWChris Williamson1:01:51

       ... for-

    6. DPDwarkesh Patel1:01:52

       Sure.

    7. CWChris Williamson1:01:52

       ... deep maths, physics, asking questions about the universe. It's like, your, "What's the best restaurant to go to in Rome?"

    8. DPDwarkesh Patel1:01:59

       Yeah.

    9. CWChris Williamson1:01:59

       You know.

    10. DPDwarkesh Patel1:02:00

        Or, like, be my therapist-

    11. CWChris Williamson1:02:01

        Yes.

    12. DPDwarkesh Patel1:02:01

        ... and talk to me. Um, and I think, by the way, this is gonna be... I don't, I don't think people are contemplating just how much more intense this is gonna get. Already it's the case that I think on, um, uh, there's websites like Character.AI, where the median user will spend hours every day just talking with these models.

    13. CWChris Williamson1:02:16

        What's Character.AI?

    14. DPDwarkesh Patel1:02:17

        Basically a chatbot, but it has a specific persona. It's meant to be a person you talk to rather than sort of a chatbot that answers your questions. Um, and these things are gonna get multimodal, right? So it'll be like, it'll be able to process your video input. It will be able to display. We already have video models that can generate, you know, things that look cool. It will look like a person. It will be smarter. It will have longer session memory. Um, maybe the whole, uh, issue of memory solved, um, altogether. And so we're not looking ahead to the time when we actually do have AGI. We will just have things that are, like, funny and endearing and, um, like, really care about you and, like, know you, or at least seem to, um, because they're trained to, right? Um, and, uh, will, will engender, maybe, too much sympathy potentially, right? It... For many people, these, these might be the most significant relationships in their life. Um, like, what other human wants to just, like, hear you talk about your problems (laughs) for a couple hours a day that you're not paying $300 an hour?

    15. CWChris Williamson1:03:11

        Dude-

    16. DPDwarkesh Patel1:03:12

        (laughs)

    17. CWChris Williamson1:03:12

        I mean, I, I saw a phenomenal video the other day. So it's this girl, a pretty girl, probably in a relationship, uh, sat in the passenger seat of a car, and a question comes up. She's got a phone in her hand, and the question comes up, says, "Can I look at your text messages?" And she goes... "C- Can I look at your social media?" And she goes...

    18. DPDwarkesh Patel1:03:30

        (laughs)

    19. CWChris Williamson1:03:30

        He goes, "Can I look at your ChatGPT?"

    20. DPDwarkesh Patel1:03:33

        (laughs)

    21. CWChris Williamson1:03:33

        Throws it out the window.

    22. DPDwarkesh Patel1:03:34

        (laughs) It's so true.

    23. CWChris Williamson1:03:36

        You know, I, I remember, uh, Seth Stephens-Davidowitz did that great book, Everybody Lies, whe- where he realized that people would ask Google things that they hadn't admitted to a therapist, that they wouldn't admit to a spouse, that they kind of hadn't admitted to themselves, and, uh, I get the sense that ChatGPT has kind of lifted the lid on that. You've got this sense that this is unbelievably secure-

    24. DPDwarkesh Patel1:03:59

        Yes.

    25. CWChris Williamson1:04:00

        ... and very intimate and exclusively one-on-one, uh, and so forgetful that, frankly, it's probably not gonna be able to remember what it was that you fucking said in a couple of days-

    26. DPDwarkesh Patel1:04:08

        Yep.

    27. CWChris Williamson1:04:08

        ... in any case. And, uh, yeah, I, I am concerned... Some of my friends who are more health anxiety-focused, the opportunity to have a always-on kind of expert to talk to about your problems, mental health, physical health-

    28. DPDwarkesh Patel1:04:29

        Yep.

    29. CWChris Williamson1:04:29

        ... stuff that you're doing with friends, it is a hypochondriac's fucking dream.

    30. DPDwarkesh Patel1:04:33

        (laughs)
13. 1:11:09 – 1:12:09

    The Importance of AI Researchers

    1. DPDwarkesh Patel1:11:09

    2. CWChris Williamson1:11:10

       Yeah, maybe. How important are, uh, individual visionaries when it comes to AI development? Th- if there's huge teams of people working on this, uh, aggregated data, learning, you know, it feels like there's a lot of ballast in the system there. Is there, is there still a great man of AI theory coming along?

    3. DPDwarkesh Patel1:11:32

       I think, uh, it seems to me that there are great researchers who have severe specific talents. They have talents in, not necessarily just AI research, but in how to code up the GPUs or accelerators so that, um, you're getting these like 25%, 50% performance gains, which are huge. Um, uh, or they're ... But it, it, it ... That, it's more of that kind of thing, I think, like more technical than, uh, from what I, uh, my sense, there's not like, "I'm just good at thinking, uh, and I can like write a great manifesto and therefore I'm the sort of person moving the organization forward."

    4. CWChris Williamson1:12:07

       Mm-hmm.

14. 1:12:09 – 1:26:26

    Where Does China's AI Progression Currently Stand?

    1. CWChris Williamson1:12:09

       What are the current constraints to progress? Is it software? Is it energy? Is it coding? Is it datasets? Is it the savant fucking guy that fixes the hardware?

Episode duration: 2:45:54