Columbia CS Professor: Why LLMs Can’t Discover New Science

0:00:07 Any LLM that was trained on pre-1915 physics would never have come up with a theory of relativity.
0:00:13 Einstein had to sort of reject the Newtonian physics and come up with a space-time continuum.
0:00:15 He completely rewrote the rules.
0:00:20 AGI will be when we are able to create new science, new results, new math.
0:00:25 When an AGI comes up with a theory of relativity, it has to go beyond what it has been trained on
0:00:27 to come up with new paradigms, new science.
0:00:30 That’s my definition of AGI.
0:00:35 Vishal Misra was trying to fix a broken Cricut stats page and accidentally helped spark one
0:00:37 of AI’s biggest breakthroughs.
0:00:43 On this episode of the A16Z podcast, I talk with Vishal and A16Z’s Martin Fasato about how
0:00:48 that moment led to retrieval augmentation generation and how Vishal’s formal models explain what
0:00:50 large language models can and can’t do.
0:00:55 We discuss why LLMs might be hitting their limits, what real reasoning looks like, and what it
0:00:56 would take to go beyond them.
0:00:57 Let’s get into it.
0:01:02 Martin, I knew you wanted to have Vishal on.
0:01:06 What do you find so remarkable about him and his contributions that inspired this?
0:01:08 Vishal actually has very similar backgrounds.
0:01:09 We both come from networking.
0:01:11 He’s a much more accomplished networking guy than I am.
0:01:12 That’s a high bar given to you.
0:01:16 And so we actually view the world in an information-theoretic way.
0:01:18 It is actually part of networking.
0:01:25 And with all this AI stuff, there’s so much work trying to create models that can help us
0:01:27 understand how these LLMs work.
0:01:33 And in my experience over the last three years, the ones that have most impacted my understanding,
0:01:36 and I think have been the most predictive, are the ones that Vishal has come up with.
0:01:40 He did a previous one that we’re going to talk about called Matrix, is it?
0:01:42 Beyond the black box, but yeah.
0:01:43 Beyond the black box.
0:01:48 So actually, we should put this in the notes for this, but the single best talk I’ve ever
0:01:54 seen on trying to understand how LLMs work is one that Vishal did at MIT, which Hari Balakrishnan
0:01:55 pointed me to, and I watched that.
0:02:00 So he did that work, and then he’s doing more recent work that’s actually trying to scope
0:02:05 out not only how LLMs reason, but it has some reflections on humans’ reason, too.
0:02:09 And so I just think he’s doing some of the more profound work in trying to understand and
0:02:12 come up with models, formal models, for how LLMs reason.
0:02:17 On that note, you said his most recent work helped you change how humans think.
0:02:18 Why don’t you flesh that out a little bit?
0:02:19 How did it sort of?
0:02:23 Well, okay, so can I just try to take a rough sketch at it, and then you just tell me how
0:02:23 wrong I am?
0:02:24 Go right ahead.
0:02:30 You’re trying to describe how LLMs work, and one thing that you found is that they reduce
0:02:40 a very, very complex, multidimensional space into basically a geometric manifold that’s
0:02:46 a reduced state space, so it’s reduced degrees of freedom, but you can actually predict where
0:02:49 in the manifold the reasoning can move to, roughly.
0:02:55 So you’ve reduced the dimensionality of the problem to a geometric manifold, and then you
0:03:01 can actually formally specify kind of how far you can reason within that manifold.
0:03:07 And the articulation is that we, or one of the intuitions is that we as humans do the
0:03:12 same thing, is we take this very complex, heavy-tailed, stochastic universe, and we reduce it to
0:03:17 kind of this geometric manifold, and then when we reason, we just move along that manifold.
0:03:20 Yeah, I think you captured it accurately.
0:03:22 That’s kind of the spirit of the work, yeah.
0:03:24 Wait, wait, can I just hear it in your words?
0:03:25 Because I’m a VC, so.
0:03:30 You’re a VC with an H-index of what, 60?
0:03:31 True.
0:03:40 Yeah, so ultimately what all these LLMs are doing, whether the early LLMs or the LLMs that
0:03:46 we have today with all sorts of post-training, RLHF, whatever you do, at the end of the day
0:03:51 what they do is they create a distribution for the next token, right?
0:03:58 So, given a prompt, these LLMs create a distribution for the next token or the next word, and then
0:04:05 they pick something from that distribution using some kind of algorithm to predict the
0:04:07 next token, pick it, and then keep going.
0:04:14 Now, what happens because of the way we train these LLMs, the architecture of the transformers,
0:04:19 and the loss function, the way you put it is, right, it sort of reduces the world into
0:04:21 these Bayesian manifolds.
0:04:30 And as long as the LLM is going in, sort of traversing through these manifolds, it is
0:04:34 confident, and it can produce something which makes sense.
0:04:40 The moment it sort of wears away from the manifold, then it starts hallucinating and starts spotting
0:04:41 nonsense.
0:04:42 Confident nonsense, but nonsense.
0:04:47 So, it creates these manifolds, and the trick is the distribution that is produced.
0:04:51 You can measure the entropy of the distribution.
0:04:53 Right?
0:04:55 But entropy the way Shannon describes it.
0:04:56 Shannon entropy.
0:04:59 Shannon entropy, yeah, not thermodynamic entropy.
0:05:05 So, suppose you have a vocabulary of, let’s say, 50,000 different tokens, and you have a
0:05:09 distribution, next token distribution over these 50,000 tokens.
0:05:12 So, let’s say the cat sat on the, right?
0:05:20 If that is a problem, then the distribution will have a high probability for map, or hat, or table,
0:05:27 and a very low probability of, let’s say, ship, or whale, or something like that, right?
0:05:32 So, because of the way it’s trained, it has these distributions.
0:05:36 Now, the distributions can be low entropy or high entropy.
0:05:42 A high entropy distribution means that there are many different ways that the LLM can go
0:05:46 with a high enough probability for all those paths.
0:05:52 Low entropy means that there are only a small set of choices for the next token.
0:05:58 And the prompts also, you can categorize into two kinds of prompts.
0:06:03 One prompt is, as you can say, high information entropy.
0:06:04 Yeah.
0:06:16 So, the way these manifolds work, the LLMs start paying attention to prompts that have
0:06:21 high information entropy and low prediction entropy.
0:06:23 So, what do I mean by that?
0:06:26 So, when I say, I’m going out for dinner.
0:06:27 Yeah.
0:06:28 Right?
0:06:33 So, when I say, I’m going out for dinner, that phrase, the LLMs have been trained
0:06:39 they’ve seen it a lot, and there are many different directions I can go with it.
0:06:44 I can say, I’m going for dinner tonight, I’m going for dinner to McDonald’s, or I’m going
0:06:45 to dinner, blah, blah, blah.
0:06:45 There are many different.
0:06:46 Yeah.
0:06:53 But, when I say, I’m going to dinner with Martin Cassaro, you know, the LLM, now this is
0:06:54 information rich.
0:06:57 This is sort of a rare phrase.
0:07:02 And now, the sort of realm of possibilities reduces, because Martin is only going to take
0:07:03 me to Michelin star restaurants.
0:07:05 Yep, yep, yep, yep.
0:07:08 I’m not going to go to a McDonald’s.
0:07:09 You get what I’m saying.
0:07:16 The moment you add more context, you make the prompt information rich, the prediction entropy
0:07:16 reduces.
0:07:17 Yep, yep, yep, yep.
0:07:22 And another example that I often cite is…
0:07:25 But just quickly, but what is your takeaway?
0:07:27 What is your implication on that?
0:07:33 Which is, of course, as you’re interested, so, yeah, so you’re, so, sorry, I forgot how you
0:07:40 described it, but so the more precise you are, the more tokens you are, I presume the less
0:07:42 options you have for the next token.
0:07:43 Is that correct or not correct?
0:07:45 Yeah, yeah, essentially.
0:07:52 So, you’re reducing it, you’re reducing it to a very specific state space when it comes
0:07:55 to confidence in an answer.
0:07:57 And this is kind of a manifold that you can go on.
0:08:04 And then, I mean, do you, do you have kind of a conclusion of what that means for systems
0:08:08 or what that means for reasoning, or is it just a nice way to articulate the bounds of
0:08:09 LLMs?
0:08:16 No, there is something, I don’t know, I don’t know if I should say profound, but there is
0:08:21 something about it which tells what these LLMs can or cannot do, right?
0:08:29 So, one of the examples that I often tell is, suppose I ask you, what is 769 times 1025?
0:08:32 You have no idea.
0:08:36 You can have some vague idea, given the two numbers, right?
0:08:43 And so, in your mind, the next token distribution of the answer is going to be diffuse, right?
0:08:45 You don’t know.
0:08:46 You have maybe a vague guess.
0:08:51 If you are mathematically very good, maybe your guess is more precise, but it is still going
0:08:52 to be diffuse.
0:08:53 And it’s not going to be the correct answer.
0:09:00 But, if I, if you say, can I write it down and do it the way we have learned multiplication
0:09:04 tables, now you know exactly what to do next step, right?
0:09:09 You write 769 and then 1025, and then you know exactly.
0:09:15 So, at each stage of that process, your prediction entropy is very low.
0:09:22 You know exactly what to do because you have been taught this algorithm.
0:09:28 And by invoking this algorithm, saying, okay, I’m not going to just guess the answer, but
0:09:30 I’m going to do it step by step.
0:09:33 Then, your prediction and entropy reduces.
0:09:39 And, you can arrive at an answer which you’re confident of and which is correct.
0:09:41 And, the LLMs are pretty much the same way.
0:09:43 That’s why chain of thought works.
0:09:50 What happens with chain of thought is, you ask the LLM to do something chain of thought.
0:09:52 It starts breaking the problem into small steps.
0:09:55 These steps, it has seen in the past.
0:09:56 It has been trained on.
0:10:00 Maybe with some different numbers, but the concept, it has been trained on.
0:10:03 And, once it breaks it down, then it’s confident.
0:10:08 Okay, now I need to do A, B, C, D, and then I arrive at this answer.
0:10:10 Whatever it is.
0:10:12 Let’s zoom back out.
0:10:13 I want to get into LLMs.
0:10:18 But, first, Vishal, maybe you can give more context on your background and how that informs
0:10:20 your work here.
0:10:21 Okay.
0:10:26 So, yeah, as Martin said, my background is very similar to his.
0:10:28 We, you know, we come from doing networking.
0:10:35 So, my PhD thesis, my sort of early work at Columbia has all been in networking.
0:10:42 But there’s another side of me, another hat that I wear, which is both an entrepreneur and
0:10:43 a cricket fan.
0:10:45 I was going to say, don’t you own a cricket team or something?
0:10:51 I’m a minority owner at your, for your local cricket team, the San Francisco Unicorns.
0:10:52 Yeah, that’s right.
0:10:54 I’m very proud to have you.
0:11:04 So, but, so, say, in the 90s, I was one of the people who started this portal called Crickinfo.
0:11:11 And Crickinfo, at one point, it was the most popular website in the world.
0:11:12 It had more hits than Yahoo.
0:11:14 That was before India came on.
0:11:21 And so, you know, we built, cricket is a very start with sport.
0:11:24 You’ll think baseball multiplied by a thousand.
0:11:31 And we had built this free searchable stats database on Crickinfo called Stats Guru.
0:11:35 And this has been available on Crickinfo since 2000.
0:11:43 But because you can search for anything, everything was made available on Stats Guru.
0:11:48 And, you know, you can’t expect people to write SQL queries to query everything.
0:11:50 So how do you, how did we do it?
0:11:56 Well, it was a web form, you know, where you could formulate your query using that form.
0:12:00 And in the back end, that was translated into SQL query, got the results and got it back.
0:12:05 But as a result, that because you could do everything, everything was made available.
0:12:11 The web form had like 25 different checkboxes, 15 text fields, 18 different dropdowns.
0:12:13 The interface was a mess.
0:12:14 It was very daunting.
0:12:23 So, and ESPN acquired Crickinfo in the mid-2006, I think.
0:12:26 But they still kept the same interface.
0:12:29 And that has always sort of nagged me.
0:12:32 And so I still know the people who run ESPN.
0:12:33 Wait, wait, what nagged you?
0:12:35 Is that Crickinfo did not have informal language?
0:12:37 It had a web form for doing queries?
0:12:40 That web form was terrible.
0:12:45 Because of that, only the real nerds use Stats Guru.
0:12:50 Of all the things in the world that bother you, the fact that an old website was a web form.
0:12:54 I appreciate your commitment to aesthetic.
0:13:01 So I’m still friendly with the people who run ESPN Crickinfo.
0:13:05 The editor-in-chief, whenever he comes to New York, you know, we meet up, we go out for a drink.
0:13:08 And so he was here in 2000.
0:13:13 So now the story shifts to how LLMs and me sort of met.
0:13:17 So January 2000, right before the pandemic, he was here.
0:13:21 And I again said, why did you do something about Stats Guru?
0:13:23 And he looks at me and says, why did you do something about Stats Guru?
0:13:30 He was kind of joking, but he thought maybe, you know, I had some ways to fix the interface.
0:13:34 So anyway, then the pandemic hit, the world stopped.
0:13:39 But in July of 2020, the first version of GPT-3 was released.
0:13:51 And I saw someone use GPT-3 to write a SQL query for their own database using natural language.
0:13:56 And I thought, can I use this to fix Stats Guru?
0:14:04 So I got early access to GPT-3, you know, getting access those days was difficult, but somehow I got it.
0:14:08 But soon I realized that, you know, no, I cannot really do it.
0:14:12 Because Stats Guru, the backend databases were so complex.
0:14:17 And if you remember, GPT-3 had only a 2048 token context window.
0:14:23 There was no way in hell I could fit the complexities of that database in that context window.
0:14:29 And GPT-3 also did not do instruction following at that time.
0:14:37 But then in trying to solve this problem, I accidentally invented what’s now called RAG.
0:14:47 Where based on the natural language query, I created a database of natural language queries and structure, the structured queries.
0:14:54 Like I created a DSL, which then translated into a REST call to Stats Guru.
0:15:00 So based on the new query, I would look through my set of natural language queries.
0:15:01 I had about 1,500 examples.
0:15:05 And I would pick the six or seven most relevant ones.
0:15:10 And then that and the structured query, I would send as a prefix and the new query.
0:15:13 And GPT-3 magically completed it.
0:15:15 And the accuracy was very high.
0:15:19 So that had been running in production since September 2021.
0:15:23 You know, about 15 months before, chatGPD came.
0:15:27 And, you know, the whole revolution in some sense started.
0:15:29 And RAG became very popular.
0:15:30 I didn’t call it RAG.
0:15:35 But this is something sort of I accidentally did in trying to solve that problem for ClickInfo.
0:15:42 Now, once I built it, you know, I was thrilled that this worked.
0:15:44 But I had no idea why it worked.
0:15:50 You know, I stared at that transformer architecture diagram.
0:15:51 I read those papers.
0:15:54 But I couldn’t understand how or why it worked.
0:16:00 So then I started in this journey of developing a mathematical model,
0:16:02 trying to understand how it worked.
0:16:09 So that’s been sort of my journey through this world of AI and LLMs
0:16:12 because I was trying to solve this cricket problem.
0:16:13 Yeah, amazing.
0:16:17 And so maybe reflecting back since the release of GPT-3,
0:16:20 what has most surprised you about how LLMs have developed?
0:16:23 So what has most surprised me?
0:16:25 The pace of development.
0:16:30 So GPT-3 was, you know, it was a nice pilot trick.
0:16:33 And you had to jump through hoops to get it to do something useful.
0:16:39 But starting with the, you know, chat GPT was an advance over GPT-3.
0:16:43 And then you had all these things like chain of thought, instruction following.
0:16:46 GPT-4 really made it polished.
0:16:51 And, you know, the pace of development has really surprised me.
0:16:54 Now, you know, when I started working with GPT-3,
0:16:57 I could sort of see what its limitations were,
0:17:00 what I could make it do, what I couldn’t make it do.
0:17:03 But I never thought of it as, you know,
0:17:06 what these LLMs have become for me now
0:17:09 and what have become for millions of people around the world.
0:17:13 We treat these models as our co-workers.
0:17:15 Almost like an intern.
0:17:18 That, you know, you’re constantly chatting with them,
0:17:21 brainstorming, making them do all sorts of work,
0:17:23 which we couldn’t imagine, you know.
0:17:26 Just when ChatGPT was released,
0:17:28 it was nice, it could write poems,
0:17:29 it could write limericks,
0:17:32 it could answer some hallucinated questions.
0:17:36 But the capabilities that have emerged now,
0:17:40 that pace has been very sort of surprising to me.
0:17:42 Do you see progress plateauing?
0:17:45 Or how do you, either now or in the near future,
0:17:46 how do you see it going?
0:17:52 Yes, in some sense, progress is plateauing.
0:17:54 It’s like the iPhone, you know,
0:17:55 when the iPhone came out,
0:17:58 wow, what is this thing?
0:18:01 And the early iterations, you know,
0:18:04 constantly we were amazed by new capabilities.
0:18:07 But the last, you know, seven, eight, nine years,
0:18:11 it’s maybe the camera got a little bit better
0:18:13 or, you know, one thing changed here
0:18:14 or memory is more.
0:18:17 But there has been no fundamental advance
0:18:19 in what it’s capable of.
0:18:24 You can sort of see a similar thing happening with these LLMs.
0:18:29 And this is not true for just one company and one model, right?
0:18:32 You look at what OpenAI is coming up with
0:18:34 or what Anthropik, Google,
0:18:40 or all these open source Chinese model or Mistral.
0:18:44 The capabilities of LLMs has not fundamentally changed.
0:18:46 They’ve become better, right?
0:18:47 They’ve improved.
0:18:53 But they have not crossed into a different realm.
0:18:57 So this is something that I really appreciate about your work.
0:19:01 And so the thing that really struck me is
0:19:03 as soon as these things showed up,
0:19:07 you actually got busy trying to have a formal model
0:19:09 of what they’re capable of,
0:19:12 which was in stark contrast to what everybody else was doing.
0:19:13 Everybody else was like,
0:19:16 AGI, these things are going to, you know,
0:19:18 recursively self-improve.
0:19:20 Like, or they’ll say,
0:19:22 Oh, these are just stochastic parrots,
0:19:23 which doesn’t mean anything.
0:19:25 So everybody had rhetoric.
0:19:27 And sometimes this rhetoric was fanciful.
0:19:30 And sometimes this rhetoric was almost reductionist.
0:19:31 Like, oh, it’s just a database,
0:19:33 which is clearly not true.
0:19:35 And the thing that really struck me about your work
0:19:36 is you’re like,
0:19:38 No, let’s figure out exactly what’s going on.
0:19:39 Let’s come up with a formal model.
0:19:41 And once we have a formal model,
0:19:43 we can reason about what that means.
0:19:47 And then, you know, in my reading of your work,
0:19:48 I kind of break it up in two pieces.
0:19:50 There’s the first one where you basically,
0:19:53 you came up with this, you know, matrix abstraction.
0:19:54 I think it’s worth you talking through.
0:19:57 And then you took in-context learning as an example
0:19:59 and you mapped it to Bayesian reasoning,
0:20:01 which to me was incredibly powerful
0:20:02 because at the time,
0:20:04 nobody knew why in-context learning worked.
0:20:08 So I think it’d be great for you to discuss that
0:20:10 because again, I think it was the first real
0:20:11 kind of formal effect on like,
0:20:13 how are these things working?
0:20:16 And then the more recent work
0:20:17 that you’re working on now
0:20:21 is a kind of more generalized version
0:20:25 of what is the state space
0:20:26 that these models output
0:20:29 when it comes to confidence,
0:20:31 which is the manifold that we’re talking about before.
0:20:33 So I think it’d be great
0:20:36 if you just described your matrix model
0:20:38 and then how you use that
0:20:41 to provide some bounds
0:20:43 what in-context learning is doing,
0:20:44 what’s happening.
0:20:45 Okay.
0:20:49 So, yeah, let’s start with that matrix abstraction.
0:20:51 So the idea behind the matrix is
0:20:53 you have this gigantic matrix
0:20:57 where every row corresponds to a prompt.
0:21:02 And then the number of columns of this matrix
0:21:05 is the vocabulary of the LLM,
0:21:07 the number of tokens it has
0:21:07 that it can emit.
0:21:10 So for every prompt,
0:21:13 this matrix contains
0:21:16 the distribution over this vocabulary.
0:21:19 So when you say the cat sat on the,
0:21:20 you know,
0:21:22 the column that corresponds to mat
0:21:25 will have a high probability.
0:21:26 Most of them will be zero.
0:21:27 But, you know,
0:21:30 reasonable continuations
0:21:32 will have a non-zero probability.
0:21:33 And so you can imagine
0:21:35 that there’s this gigantic matrix.
0:21:38 Now, the size of this matrix is,
0:21:39 you know,
0:21:42 if we just take just the old
0:21:45 first-generation GPT-3 model,
0:21:48 which had a context window of 2,000 tokens
0:21:54 and a vocabulary of 50,000 next tokens
0:21:56 or 50,000 tokens,
0:21:58 then the size of it,
0:22:00 the number of rows in this matrix
0:22:03 is more than the number of atoms
0:22:06 across all galaxies that we know of.
0:22:07 So clearly,
0:22:11 we cannot represent it exactly.
0:22:13 Now, fortunately,
0:22:16 a lot of these rows are,
0:22:19 do not appear in real life, right?
0:22:21 An arbitrary collection of tokens,
0:22:23 you are not going to use that as a prompt.
0:22:25 Similarly,
0:22:28 you saw a lot of these rows are absent
0:22:29 and a lot of the column values
0:22:30 are also zero.
0:22:31 Right?
0:22:33 When you say the cat sat on the,
0:22:35 it’s unlikely to be followed
0:22:37 by the token corresponding
0:22:38 to, let’s say, numbers.
0:22:40 Or, you know,
0:22:41 an arbitrary collection of tokens.
0:22:44 there will be only a very small subset of tokens
0:22:46 that can follow a particular prompt.
0:22:49 So this matrix is very,
0:22:50 very sparse.
0:22:53 But even after that sparsity
0:22:56 and even after removing the sort of gibberish prompts,
0:23:02 the size of this matrix is too much for these models to represent,
0:23:04 even with the trail in parameters.
0:23:05 So what,
0:23:07 in an abstract sense,
0:23:08 what is happening is
0:23:11 the models get trained
0:23:13 on certain,
0:23:14 you know,
0:23:14 data
0:23:16 from the training set
0:23:17 and certain,
0:23:17 some,
0:23:18 a subset,
0:23:20 a small subset of these rows,
0:23:21 you have
0:23:23 reasonable values
0:23:26 for the next token distribution.
0:23:29 whenever you give the prompt
0:23:30 something new,
0:23:31 right,
0:23:33 then it’ll try to interpolate
0:23:34 with what it has learned
0:23:36 and what’s there
0:23:37 in the new prompt
0:23:39 and come up with
0:23:40 a new distribution.
0:23:42 But it’s basically,
0:23:44 so it’s more than a stochastic parrot.
0:23:47 It is sort of based on this,
0:23:50 on this subset of the matrix
0:23:51 that it has been trained on.
0:23:53 So,
0:23:53 so when I say,
0:23:54 you know,
0:23:56 I’m going out for dinner
0:23:57 with Martin
0:23:58 tonight.
0:24:00 Now,
0:24:02 I’m reasonably sure
0:24:04 that it has never encountered
0:24:04 that phrase
0:24:07 in its training data,
0:24:07 right?
0:24:09 But it has encountered
0:24:10 variants of this phrase
0:24:12 and
0:24:14 given that
0:24:15 I’m going out with Martin,
0:24:16 it,
0:24:17 it can produce
0:24:18 a Bayesian posterior.
0:24:20 It uses that evidence
0:24:20 that Martin
0:24:21 is the one
0:24:22 that I’m going for dinner with
0:24:23 and it’ll produce
0:24:24 a next token distribution
0:24:26 that will focus on
0:24:27 the likely places
0:24:28 that we are going.
0:24:30 So,
0:24:31 so this matrix,
0:24:33 because it’s represented
0:24:34 in a compressed way,
0:24:36 yet the models
0:24:37 respond to everything,
0:24:38 every prompt.
0:24:40 How do they do it?
0:24:41 Well,
0:24:41 they,
0:24:42 they go back to what
0:24:43 they’ve been trained on,
0:24:44 interpolate there,
0:24:46 and use the prompt
0:24:46 as sort of
0:24:48 some evidence
0:24:49 to compute
0:24:50 a new distribution.
0:24:53 which,
0:24:53 so right,
0:24:55 so the context
0:24:56 of the prompt
0:24:57 impacts
0:24:59 the posterior distribution.
0:25:00 Exactly,
0:25:00 yeah.
0:25:01 Right.
0:25:02 And you mapped
0:25:03 to Bayesian
0:25:04 learning
0:25:04 where
0:25:06 the,
0:25:07 the context
0:25:08 is
0:25:09 the new
0:25:09 evidence.
0:25:11 New evidence,
0:25:11 exactly.
0:25:12 To learn from.
0:25:12 So,
0:25:13 so I’ll give you,
0:25:13 so,
0:25:14 so for instance,
0:25:15 the cricket example
0:25:16 that I spoke about earlier.
0:25:17 Yeah.
0:25:17 So I,
0:25:19 I created my own DSL.
0:25:20 Yeah.
0:25:21 Which,
0:25:21 you know,
0:25:22 mapped a natural language
0:25:23 query in cricket
0:25:24 to this DSL,
0:25:26 which then I can
0:25:27 translate into a SQL
0:25:28 query or a REST API
0:25:29 or whatever.
0:25:30 But getting the DSL
0:25:31 is important.
0:25:32 Now,
0:25:33 these LLMs
0:25:34 have never seen
0:25:34 that DSL.
0:25:35 I designed it.
0:25:36 Yeah.
0:25:38 But yet,
0:25:39 after showing
0:25:39 a few examples,
0:25:41 it learned it.
0:25:43 How did it learn it?
0:25:43 And this is,
0:25:45 this is in the prompt.
0:25:45 You didn’t,
0:25:46 no training,
0:25:46 no post training.
0:25:46 It’s in the prompt.
0:25:48 100% in the prompt,
0:25:48 right?
0:25:49 So like,
0:25:50 it’s the way to understand it.
0:25:51 Yeah, yeah.
0:25:51 Yeah, yeah.
0:25:52 This is,
0:25:52 this was happening
0:25:53 in October 2020.
0:25:55 Right?
0:25:56 I had no access
0:25:57 to internals
0:25:57 of OpenAI.
0:25:59 I could just,
0:25:59 you know,
0:26:00 access the API.
0:26:01 OpenAI
0:26:02 had no access
0:26:03 to internal
0:26:04 structure of Stats Guru
0:26:06 or the DSL
0:26:07 that I cooked up
0:26:07 in my head.
0:26:08 Yet,
0:26:09 after showing it
0:26:10 only a few examples,
0:26:11 it learned it right away.
0:26:13 So that’s an example
0:26:15 where it has seen
0:26:15 DSLs
0:26:17 or structures
0:26:18 in the past.
0:26:20 And
0:26:21 now using
0:26:22 this evidence
0:26:22 that I show,
0:26:22 okay,
0:26:23 this is what
0:26:24 my DSL looks like.
0:26:25 Now,
0:26:26 a new
0:26:27 natural language query,
0:26:28 it is able to
0:26:30 create the right
0:26:30 posterior distribution
0:26:31 for the tokens
0:26:33 that map
0:26:34 to the example
0:26:35 that I’ve seen.
0:26:35 Now,
0:26:37 the other
0:26:37 beautiful thing
0:26:38 about this is
0:26:40 this is an example
0:26:41 of few-shot learning
0:26:42 or in-context learning.
0:26:43 Right?
0:26:45 But when I
0:26:46 give that prompt
0:26:47 along with
0:26:48 these examples
0:26:49 to this LLM,
0:26:50 I’m not saying
0:26:51 to the LLM,
0:26:51 okay,
0:26:52 this is an example
0:26:52 of few-shot learning,
0:26:53 so learn from
0:26:54 these examples.
0:26:55 Right?
0:26:56 You just pass
0:26:58 this to the LLM
0:26:59 as a prompt
0:27:00 and it
0:27:01 processes it
0:27:02 exactly
0:27:03 the way
0:27:04 it would process
0:27:04 any other
0:27:05 prompt
0:27:05 which is
0:27:05 not an
0:27:06 example
0:27:06 of
0:27:07 in-context
0:27:07 learning.
0:27:09 So,
0:27:09 that really
0:27:10 means that
0:27:10 the underlying
0:27:11 mechanism
0:27:12 is the same.
0:27:15 Whether you
0:27:16 give a set
0:27:17 of examples
0:27:18 and then
0:27:18 ask it to
0:27:19 complete a talk,
0:27:20 a task like
0:27:21 in-context learning
0:27:22 or just
0:27:22 give it
0:27:23 some prompt
0:27:24 for continuation
0:27:24 that I’m
0:27:25 going out
0:27:25 for dinner
0:27:26 with Martin
0:27:26 tonight.
0:27:27 There’s no
0:27:28 in-context learning
0:27:28 there.
0:27:30 But
0:27:32 the process
0:27:33 with which
0:27:34 it’s generating
0:27:34 or doing
0:27:35 this inferencing
0:27:36 is exactly
0:27:36 the same.
0:27:37 And that’s
0:27:38 what I
0:27:38 have been
0:27:39 trying to
0:27:39 model
0:27:40 and come
0:27:40 up with
0:27:41 a formal
0:27:41 model of.
0:27:43 What I’ve
0:27:43 found very
0:27:44 impressive
0:27:46 is you’ve
0:27:46 used this
0:27:48 basic model
0:27:49 to show
0:27:49 a number
0:27:50 of things,
0:27:50 right,
0:27:50 to describe
0:27:51 context learning
0:27:52 and to map
0:27:52 the Bayesian
0:27:52 learning.
0:27:53 But you
0:27:53 did it
0:27:53 for another
0:27:54 one where
0:27:55 you’ve
0:27:55 sketched out
0:27:56 this almost
0:27:57 glib argument
0:27:58 on Twitter,
0:27:58 on X,
0:27:59 where you
0:28:00 made this
0:28:05 rough argument
0:28:06 for why
0:28:06 recursive
0:28:07 self-improvement
0:28:09 can’t happen
0:28:09 without
0:28:10 additional
0:28:11 information.
0:28:11 And so
0:28:13 maybe just
0:28:13 walk through
0:28:14 very quickly
0:28:14 how this
0:28:15 same model
0:28:15 you can
0:28:16 just very
0:28:17 quickly show
0:28:17 that a
0:28:18 model can
0:28:18 never
0:28:19 recursively
0:28:20 self-improve.
0:28:22 So,
0:28:24 you know,
0:28:25 another
0:28:26 phrase that
0:28:29 we’ve been
0:28:29 using
0:28:31 recently is,
0:28:31 you know,
0:28:32 the output
0:28:32 of the
0:28:33 LLM is
0:28:33 the inductive
0:28:34 closure of
0:28:34 what it
0:28:35 has been
0:28:35 trained on.
0:28:36 Yeah.
0:28:37 So when
0:28:38 you say
0:28:38 that it
0:28:39 can
0:28:39 recursively
0:28:40 self-improve,
0:28:44 it could
0:28:45 mean one
0:28:45 of two
0:28:45 things.
0:28:46 So let’s
0:28:46 get back
0:28:47 to the
0:28:47 Well,
0:28:47 actually,
0:28:48 you know
0:28:48 what’s
0:28:48 kind of
0:28:48 interesting
0:28:49 is like
0:28:49 often
0:28:51 most people
0:28:52 agree that
0:28:52 if you have
0:28:53 one LLM
0:28:53 and you
0:28:53 just feed
0:28:54 the output
0:28:54 and the
0:28:54 input,
0:28:55 like it’s
0:28:55 not going
0:28:55 to do
0:28:56 anything.
0:28:57 But then
0:28:58 often people
0:28:58 will say,
0:28:58 well,
0:28:58 what if
0:28:59 you have
0:28:59 two LLMs,
0:28:59 you have
0:29:00 no external
0:29:01 information,
0:29:01 but you have
0:29:01 two LLMs
0:29:02 talking to
0:29:02 each other,
0:29:03 maybe they
0:29:04 can improve
0:29:04 each other
0:29:05 and then you
0:29:05 can have
0:29:05 like,
0:29:06 you know,
0:29:06 a takeoff
0:29:07 scenario.
0:29:07 But again,
0:29:08 you even
0:29:09 address this,
0:29:09 even in the
0:29:10 case of like
0:29:11 N number
0:29:11 of LLMs
0:29:12 using kind
0:29:13 of the
0:29:14 matrix model
0:29:14 to show
0:29:15 that like
0:29:15 you just
0:29:15 aren’t
0:29:16 getting any
0:29:16 information
0:29:18 entropy.
0:29:20 So you
0:29:21 can represent
0:29:21 the sort
0:29:22 of information
0:29:23 contained in
0:29:23 these models.
0:29:25 And let’s
0:29:25 go back to
0:29:26 that matrix
0:29:27 analogy that
0:29:27 I have,
0:29:29 the matrix
0:29:29 abstraction.
0:29:30 So like I
0:29:30 said,
0:29:31 you know,
0:29:33 these models
0:29:35 represent a
0:29:36 subset of
0:29:36 the rows.
0:29:39 So a subset
0:29:39 of the rows
0:29:41 are represented,
0:29:42 but some
0:29:42 of these
0:29:45 rows are
0:29:46 able to
0:29:47 help fill
0:29:48 out
0:29:48 some of
0:29:49 the missing
0:29:49 rows.
0:29:51 For instance,
0:29:51 you know,
0:29:53 if the
0:29:54 model knows
0:29:54 how to do
0:29:55 multiplication
0:29:55 doing the
0:29:56 step-by-step,
0:29:57 then every
0:29:57 row that is
0:29:58 corresponding to
0:29:58 let’s say
0:29:59 769 times
0:30:00 125 or
0:30:00 whatever,
0:30:01 all those
0:30:01 multiplications,
0:30:03 it can fill
0:30:03 out the
0:30:03 answer.
0:30:04 Because it
0:30:05 has those
0:30:06 algorithms
0:30:06 sort of
0:30:07 embedded in
0:30:07 them,
0:30:08 you just
0:30:08 need to
0:30:09 unroll them.
0:30:10 So it
0:30:11 can sort
0:30:11 of self
0:30:12 improve up
0:30:13 to a point.
0:30:14 But beyond
0:30:14 a point,
0:30:16 these models
0:30:17 can only
0:30:19 sort of
0:30:20 generate what
0:30:20 they have
0:30:21 been trained
0:30:21 on.
0:30:21 So let
0:30:21 me give
0:30:22 you,
0:30:24 I’ll give
0:30:24 you three
0:30:25 examples.
0:30:25 Yeah.
0:30:27 So any
0:30:27 model,
0:30:29 any LLM
0:30:29 that was
0:30:31 trained on
0:30:34 pre-1915
0:30:35 physics
0:30:36 would never
0:30:37 have come
0:30:38 up with a
0:30:38 theory of
0:30:38 relativity.
0:30:41 Einstein
0:30:41 had to
0:30:42 sort of
0:30:42 reject the
0:30:43 Newtonian
0:30:44 physics and
0:30:44 come up
0:30:44 with a
0:30:45 space-time
0:30:46 continuum.
0:30:46 He
0:30:47 completely
0:30:47 rewrote the
0:30:47 rules.
0:30:49 So that
0:30:50 is an
0:30:50 example of
0:30:51 AGI,
0:30:54 where you
0:30:55 are generating
0:30:56 or generating
0:30:56 new knowledge.
0:30:57 It’s not
0:30:58 simply
0:30:59 unrolling
0:31:00 what’s already
0:31:01 about the
0:31:01 universe.
0:31:02 It’s actually
0:31:02 discovering
0:31:02 something
0:31:03 fundamental
0:31:03 about the
0:31:04 universe.
0:31:05 And for
0:31:05 that you
0:31:05 have to
0:31:06 go outside
0:31:06 your
0:31:06 training
0:31:06 set.
0:31:07 Similarly,
0:31:09 any LLM
0:31:09 that was
0:31:09 trained on
0:31:10 it would
0:31:10 not have
0:31:11 come up
0:31:11 with quantum
0:31:11 mechanics.
0:31:14 That’s where
0:31:15 particle duality
0:31:15 or this
0:31:16 whole probabilistic
0:31:17 notion or
0:31:18 that energy
0:31:19 is not
0:31:20 continuous but
0:31:20 it is
0:31:20 quantized.
0:31:21 You had
0:31:22 to reject
0:31:22 Newtonian
0:31:23 physics.
0:31:24 Or
0:31:25 Gettel’s
0:31:25 incompleteness
0:31:26 theorem.
0:31:28 He had to
0:31:28 go outside
0:31:29 the axioms
0:31:30 to say that
0:31:30 okay it is
0:31:31 incomplete.
0:31:32 So those
0:31:32 are examples
0:31:33 where you’re
0:31:34 creating
0:31:36 new science
0:31:37 or
0:31:38 fundamentally
0:31:38 new
0:31:38 results.
0:31:40 That kind
0:31:40 of
0:31:41 self-improvement
0:31:42 is not
0:31:43 possible with
0:31:43 these
0:31:44 architectures.
0:31:45 They can
0:31:46 refine these
0:31:46 they can
0:31:47 fill out
0:31:47 these rows
0:31:48 where the
0:31:48 answer
0:31:49 already
0:31:49 exists.
0:31:50 Another
0:31:50 example
0:31:51 which has
0:31:52 received a
0:31:52 lot of
0:31:52 press these
0:31:53 days is
0:31:54 these
0:31:54 IMO
0:31:55 results.
0:31:56 International
0:31:56 math
0:31:56 will
0:31:56 appear.
0:31:58 Whether
0:31:58 it’s a
0:31:59 human
0:31:59 solving it
0:32:01 or the
0:32:02 LLM
0:32:02 solving it
0:32:03 they are
0:32:04 not inventing
0:32:05 new kinds
0:32:05 of math.
0:32:08 they are
0:32:09 able to
0:32:10 connect
0:32:10 known
0:32:11 results
0:32:11 in a
0:32:12 sequence
0:32:12 of
0:32:12 steps
0:32:13 to
0:32:14 come
0:32:14 up
0:32:14 with
0:32:14 the
0:32:14 answer.
0:32:16 So even
0:32:17 the LLMs
0:32:17 what they are
0:32:17 doing is
0:32:18 they are
0:32:18 exploring
0:32:19 all sorts
0:32:19 of
0:32:20 solutions.
0:32:22 In some
0:32:22 of these
0:32:22 solutions
0:32:24 they start
0:32:24 going on
0:32:25 this path
0:32:26 where their
0:32:27 next token
0:32:27 entropy
0:32:28 is low.
0:32:30 So that’s
0:32:31 where I say
0:32:32 they are in
0:32:32 that Bayesian
0:32:33 manifold
0:32:34 where you
0:32:35 have this
0:32:35 entropy
0:32:36 collapse
0:32:37 and by
0:32:37 doing those
0:32:38 steps
0:32:39 you arrive
0:32:40 at the
0:32:40 answer.
0:32:41 But you’re
0:32:42 not inventing
0:32:42 new math.
0:32:43 You’re not
0:32:43 inventing new
0:32:44 axioms or
0:32:45 new branches
0:32:46 of mathematics.
0:32:47 You’re sort
0:32:48 of using
0:32:48 what you’ve
0:32:49 been trained
0:32:50 on to
0:32:50 arrive at
0:32:51 that answer.
0:32:52 So those
0:32:53 things LLMs
0:32:54 can do
0:32:55 you know
0:32:55 they’ll get
0:32:56 better at it
0:32:57 of connecting
0:32:57 the known
0:32:58 dots.
0:33:00 But creating
0:33:01 new dots
0:33:02 I think we
0:33:03 need an
0:33:03 architectural
0:33:05 advance.
0:33:07 So Martin
0:33:07 was talking
0:33:08 earlier about
0:33:08 how the
0:33:09 discourse
0:33:10 was either
0:33:11 stochastic
0:33:12 parrots or
0:33:13 AGI
0:33:13 recursive
0:33:14 solving.
0:33:16 How do you
0:33:16 conceive of
0:33:17 the AGI
0:33:18 discourse or
0:33:20 even the
0:33:20 concept?
0:33:21 What does
0:33:22 it mean to
0:33:22 the extent
0:33:22 that it’s
0:33:23 useful?
0:33:23 How do
0:33:23 you think
0:33:24 about that?
0:33:27 the way
0:33:28 I think
0:33:28 about it
0:33:28 the way
0:33:29 we’ve tried
0:33:29 to formulate
0:33:29 it in our
0:33:30 papers is
0:33:31 it’s beyond
0:33:31 a stochastic
0:33:32 parrot but
0:33:32 it’s not
0:33:32 AGI.
0:33:34 It’s doing
0:33:34 Bayesian
0:33:36 reasoning over
0:33:36 what it has
0:33:37 been trained
0:33:37 on.
0:33:39 It’s a lot
0:33:40 more sophisticated
0:33:40 than just
0:33:41 a stochastic
0:33:42 parrot.
0:33:44 How do you
0:33:44 define AGI?
0:33:46 Okay so
0:33:47 AGI
0:33:49 so how do
0:33:49 I define
0:33:50 AGI?
0:33:52 So the
0:33:52 way I would
0:33:53 say that
0:33:56 LLMs
0:33:56 currently
0:33:58 navigate
0:33:58 through this
0:33:59 known Bayesian
0:34:00 manifold.
0:34:01 AGI will
0:34:02 create new
0:34:03 manifolds.
0:34:05 So right
0:34:05 now these
0:34:06 models navigate
0:34:07 they do
0:34:07 not create.
0:34:09 AGI will
0:34:09 be when
0:34:09 we are able
0:34:10 to create
0:34:12 new science
0:34:13 new results
0:34:14 new math.
0:34:15 When an
0:34:15 AGI comes
0:34:16 up with a
0:34:16 theory of
0:34:16 relativity
0:34:17 I mean it’s
0:34:18 an extremely
0:34:18 high bar
0:34:18 but you get
0:34:19 what I’m
0:34:19 saying.
0:34:20 It has to
0:34:21 go beyond
0:34:21 what it has
0:34:22 been trained
0:34:23 on to
0:34:23 come up
0:34:24 with
0:34:26 new
0:34:26 paradigms
0:34:27 new
0:34:27 science
0:34:28 and
0:34:29 that’s
0:34:30 that’s
0:34:30 my
0:34:30 definition
0:34:31 of AGI.
0:34:32 Vishal can
0:34:32 you do
0:34:33 you think
0:34:33 that based
0:34:33 on the
0:34:34 work you’ve
0:34:34 done can
0:34:35 you bound
0:34:35 the amount
0:34:36 of data
0:34:37 computer
0:34:39 or data
0:34:40 or compute
0:34:40 that would
0:34:41 be needed
0:34:42 in order
0:34:43 for it
0:34:44 to evolve
0:34:44 it.
0:34:45 So what
0:34:45 are the
0:34:46 problems
0:34:48 if you
0:34:48 just take
0:34:49 LLMs as
0:34:49 they exist
0:34:50 is there
0:34:50 was so
0:34:51 much data
0:34:51 used to
0:34:52 create them
0:34:53 to create
0:34:53 a new
0:34:54 manifold
0:34:55 will need
0:34:55 a lot
0:34:56 more data
0:34:56 just because
0:34:57 of the
0:34:57 basic
0:34:58 mechanisms
0:34:58 right
0:34:58 otherwise
0:34:58 it’ll
0:34:59 just
0:34:59 kind of
0:34:59 like
0:35:00 you know
0:35:00 get kind
0:35:01 of consumed
0:35:01 into the
0:35:02 existing
0:35:02 set of
0:35:02 data
0:35:03 like
0:35:04 have you
0:35:04 found
0:35:05 any bounds
0:35:05 of
0:35:06 of
0:35:07 what
0:35:07 would
0:35:07 be
0:35:08 needed
0:35:08 to
0:35:09 actually
0:35:09 evolve
0:35:09 the
0:35:10 manifold
0:35:10 in
0:35:10 a
0:35:10 useful
0:35:11 way
0:35:11 or
0:35:11 do you
0:35:11 think
0:35:11 we
0:35:12 just
0:35:12 need a new
0:35:12 architecture
0:35:14 I personally
0:35:16 think that we
0:35:16 need a new
0:35:17 architecture
0:35:19 the more data
0:35:19 that we have
0:35:20 the more compute
0:35:21 we have
0:35:21 we’ll get maybe
0:35:23 smoother manifolds
0:35:23 so it’s like a map
0:35:25 there’s this view
0:35:26 that people have
0:35:26 they’re like
0:35:27 well
0:35:28 Vishal
0:35:29 this is all
0:35:31 good and well
0:35:32 but you know
0:35:32 I could just
0:35:33 take an LLM
0:35:34 and I can give it
0:35:34 eyes
0:35:35 and I can give it
0:35:35 ears
0:35:36 and I can put it
0:35:36 in the world
0:35:37 and it’ll gain
0:35:38 information
0:35:38 and based on that
0:35:40 it’ll improve
0:35:41 itself
0:35:42 and therefore
0:35:43 it can learn
0:35:44 new things
0:35:44 but the
0:35:45 counterpoint
0:35:46 that I’ve always
0:35:46 just intuitively
0:35:47 thought to that
0:35:48 is
0:35:49 the amount
0:35:49 of data
0:35:50 used to train
0:35:50 these things
0:35:51 is so large
0:35:52 how much
0:35:53 can you actually
0:35:53 evolve that
0:35:54 manifold
0:35:54 given an
0:35:55 incremental
0:35:55 I mean
0:35:56 almost none
0:35:56 at all
0:35:57 right
0:35:57 there has
0:35:58 to be
0:35:59 some other
0:36:00 way to generate
0:36:01 new manifolds
0:36:01 that aren’t
0:36:02 evolving the
0:36:03 existing one
0:36:05 I completely
0:36:06 agree
0:36:06 there has
0:36:07 to be
0:36:07 a new
0:36:08 sort of
0:36:08 architectural
0:36:09 leap
0:36:09 that is
0:36:09 needed
0:36:11 to go
0:36:12 from the
0:36:12 current
0:36:13 you know
0:36:13 just throwing
0:36:14 more data
0:36:14 and more
0:36:15 compute
0:36:16 you know
0:36:16 it’s going
0:36:16 to plateau
0:36:17 it’s you know
0:36:18 the iPhone
0:36:18 15
0:36:18 16
0:36:19 17
0:36:20 and are
0:36:20 there any
0:36:21 research
0:36:21 directions
0:36:22 that are
0:36:22 promising
0:36:23 in your
0:36:23 mind
0:36:23 that
0:36:23 might
0:36:24 help
0:36:24 us
0:36:24 you know
0:36:25 go beyond
0:36:25 LLM
0:36:26 limitations
0:36:28 so I
0:36:29 mean
0:36:31 again
0:36:32 I love
0:36:32 LLMs
0:36:32 they are
0:36:33 fantastic
0:36:33 and they are
0:36:34 going to
0:36:34 increase
0:36:34 productivity
0:36:35 like
0:36:36 nobody’s
0:36:36 business
0:36:37 but I
0:36:37 don’t think
0:36:37 they are
0:36:38 the answer
0:36:39 so
0:36:39 you know
0:36:39 Yard
0:36:41 famously
0:36:41 says
0:36:42 that
0:36:43 LLMs
0:36:43 are a
0:36:44 distraction
0:36:45 on the
0:36:45 road
0:36:45 to AGI
0:36:48 I don’t
0:36:48 think
0:36:49 I’m not
0:36:49 quite
0:36:50 in that
0:36:50 camp
0:36:51 but I
0:36:52 think we
0:36:53 need
0:36:53 a new
0:36:54 new
0:36:55 architecture
0:36:55 to sit
0:36:56 on top
0:36:56 of LLMs
0:36:58 to reach
0:36:58 AGI
0:36:59 you know
0:37:00 a very
0:37:00 basic
0:37:00 thing
0:37:00 you know
0:37:00 what
0:37:01 Martin
0:37:01 just
0:37:01 said
0:37:02 you give
0:37:02 them
0:37:02 eyes
0:37:03 and you
0:37:03 give
0:37:03 them
0:37:03 ears
0:37:03 you make
0:37:04 them
0:37:04 multi-model
0:37:05 of course
0:37:05 they’ll
0:37:05 become more
0:37:06 powerful
0:37:07 but you need
0:37:07 a little bit
0:37:07 more than
0:37:08 that
0:37:09 you know
0:37:09 the way
0:37:10 human brains
0:37:10 learn
0:37:11 with very
0:37:12 few
0:37:12 examples
0:37:13 that’s not
0:37:13 the way
0:37:14 transformers
0:37:14 learn
0:37:16 and
0:37:18 you know
0:37:19 I’m not
0:37:19 saying that
0:37:20 we need
0:37:20 to create
0:37:21 an Einstein
0:37:21 or a Gator
0:37:22 but
0:37:23 there has
0:37:23 to be
0:37:23 an
0:37:24 architectural
0:37:24 leap
0:37:25 that is
0:37:25 able to
0:37:26 create
0:37:26 these
0:37:27 manifolds
0:37:27 and just
0:37:28 throwing
0:37:28 new
0:37:28 data
0:37:29 will not
0:37:29 do it
0:37:29 it’ll
0:37:29 just
0:37:30 smoothen
0:37:30 out
0:37:30 the
0:37:30 already
0:37:31 existing
0:37:31 manifolds
0:37:33 is that
0:37:33 something
0:37:33 so is
0:37:34 your goal
0:37:34 to actually
0:37:35 help
0:37:37 like think
0:37:37 through new
0:37:38 architectures
0:37:39 or are
0:37:39 you
0:37:39 primarily
0:37:40 focused
0:37:40 on
0:37:41 putting
0:37:41 formal
0:37:41 bounds
0:37:41 on
0:37:42 existing
0:37:42 architectures
0:37:45 a bit
0:37:45 of both
0:37:45 I mean
0:37:46 the former
0:37:47 goal is
0:37:47 the more
0:37:47 ambitious
0:37:48 one
0:37:48 that
0:37:49 everybody
0:37:50 is chasing
0:37:51 and yeah
0:37:52 I think
0:37:52 about that
0:37:52 constantly
0:37:54 are there
0:37:55 any new
0:37:55 even
0:37:56 like
0:37:57 sort of
0:37:57 hints
0:37:57 that a new
0:37:58 architect
0:37:58 or like
0:37:58 have we
0:37:59 started to
0:37:59 make any
0:38:00 progress
0:38:01 on new
0:38:02 architectures
0:38:03 or is
0:38:03 it
0:38:10 you
0:38:10 know
0:38:12 Yarn
0:38:12 has been
0:38:13 pushing
0:38:14 at this
0:38:14 JPA architecture
0:38:16 energy based
0:38:17 architectures
0:38:18 they seem
0:38:18 promising
0:38:20 the way
0:38:21 I have
0:38:21 been
0:38:22 sort of
0:38:22 thinking
0:38:23 about it
0:38:23 is
0:38:27 you know
0:38:27 there’s
0:38:28 this
0:38:29 set of
0:38:30 benchmarks
0:38:31 or the
0:38:31 ARC
0:38:32 prize
0:38:32 yeah
0:38:33 right
0:38:33 that
0:38:34 Mike
0:38:35 and
0:38:36 François
0:38:36 Chalet
0:38:36 have
0:38:38 and
0:38:39 if you
0:38:40 understand
0:38:41 why the
0:38:41 LLMs are
0:38:42 failing on
0:38:42 this test
0:38:43 maybe you
0:38:44 can sort of
0:38:44 reverse engineer
0:38:45 a new
0:38:45 architecture
0:38:47 that will
0:38:48 help you
0:38:50 succeed in
0:38:51 that right
0:38:52 and
0:38:53 I agree
0:38:54 with a
0:38:55 lot of
0:38:55 what
0:38:56 several
0:38:56 people say
0:38:57 that you
0:38:57 know
0:38:58 language is
0:38:59 great but
0:38:59 language is
0:39:00 not the
0:39:00 answer
0:39:01 you know
0:39:02 when I’m
0:39:03 looking at
0:39:04 catching a
0:39:04 ball that
0:39:05 is coming
0:39:05 to me
0:39:06 I’m mentally
0:39:07 doing that
0:39:07 simulation in
0:39:08 my head
0:39:09 I’m not
0:39:10 translating it
0:39:11 to language to
0:39:11 figure out where
0:39:12 it will land
0:39:13 I do that
0:39:13 simulation in
0:39:14 my head
0:39:15 so
0:39:16 a way
0:39:17 you know
0:39:17 one of the
0:39:17 new
0:39:18 architectures
0:39:19 architectural
0:39:19 things is
0:39:20 how do we
0:39:21 do
0:39:22 how do we
0:39:22 get these
0:39:23 models to
0:39:23 do approximate
0:39:24 simulations
0:39:26 to test
0:39:26 out that
0:39:27 idea and
0:39:28 whether to
0:39:28 proceed
0:39:30 or not
0:39:31 so
0:39:33 so yeah
0:39:33 we have
0:39:34 you know
0:39:35 another thing
0:39:35 that I’ve
0:39:36 always wondered
0:39:37 about is
0:39:38 did we
0:39:39 develop as
0:39:39 humans
0:39:40 did we
0:39:40 develop
0:39:41 language
0:39:42 because we
0:39:42 were
0:39:42 intelligent
0:39:44 or because
0:39:45 we developed
0:39:45 language
0:39:47 we accelerated
0:39:47 our intelligence
0:39:49 so I
0:39:50 don’t know
0:39:50 which side
0:39:50 of the
0:39:51 camp
0:39:52 you follow
0:39:52 on that
0:39:53 question
0:39:54 what’s
0:39:54 interesting
0:39:55 is
0:39:55 like
0:39:55 you
0:39:55 have
0:39:55 these
0:39:56 anecdotal
0:39:57 examples
0:39:58 of
0:40:00 humans
0:40:00 developing
0:40:01 languages
0:40:01 de novo
0:40:01 that have
0:40:01 been
0:40:02 recorded
0:40:02 right
0:40:02 like
0:40:03 it’s
0:40:03 either
0:40:03 the
0:40:04 Guatemalan
0:40:05 or Nicaraguan
0:40:06 sign language
0:40:06 right
0:40:07 where there
0:40:07 is these
0:40:07 students
0:40:08 that develop
0:40:09 their own
0:40:10 language
0:40:11 without being
0:40:11 taught
0:40:12 and so
0:40:12 that would
0:40:13 suggest
0:40:13 that language
0:40:14 just follows
0:40:14 intelligence
0:40:16 the problem
0:40:16 is they’re
0:40:17 all anecdotal
0:40:17 right
0:40:18 like who
0:40:18 knows
0:40:19 if somebody
0:40:19 didn’t teach
0:40:20 them sign
0:40:20 language
0:40:20 like nobody
0:40:21 really knows
0:40:21 there is
0:40:22 no controls
0:40:23 so this is
0:40:23 all these
0:40:24 observational
0:40:24 studies
0:40:25 and there’s
0:40:26 so few
0:40:26 of them
0:40:27 you have
0:40:28 to wonder
0:40:28 if it’s
0:40:29 just kind
0:40:30 of sloppy
0:40:31 observation
0:40:31 and so I
0:40:31 think that
0:40:32 the question
0:40:32 is still
0:40:33 outstanding
0:40:34 yeah
0:40:35 so
0:40:38 I mean
0:40:38 language
0:40:39 definitely
0:40:40 accelerated
0:40:41 our intelligence
0:40:41 there’s no
0:40:42 question about
0:40:42 that
0:40:43 but which
0:40:43 followed which
0:40:44 we don’t
0:40:44 know
0:40:46 I view it
0:40:47 as a networking
0:40:47 problem
0:40:48 naturally
0:40:48 which is
0:40:48 once you
0:40:49 have languages
0:40:49 you can
0:40:50 communicate
0:40:50 and when
0:40:51 you can
0:40:51 communicate
0:40:51 you can
0:40:52 store
0:40:52 you can
0:40:53 replicate
0:40:53 yeah
0:40:53 yeah
0:40:54 exactly
0:40:55 cool
0:40:57 again this
0:40:57 is kind
0:40:58 of a wonky
0:40:58 question
0:40:59 but
0:41:02 you know
0:41:02 I think
0:41:03 one thing
0:41:03 that you’ve
0:41:03 brought to
0:41:04 the discourse
0:41:04 and for
0:41:04 those that
0:41:04 are listening
0:41:05 to this
0:41:05 I really
0:41:05 think
0:41:05 that you
0:41:05 should
0:41:06 look up
0:41:06 Vishal’s
0:41:06 work
0:41:07 and read
0:41:07 it
0:41:07 I just
0:41:07 think
0:41:08 it’ll
0:41:08 give you
0:41:08 a really
0:41:09 really
0:41:09 especially
0:41:09 if you
0:41:09 have a
0:41:10 systems
0:41:10 background
0:41:10 like a
0:41:11 networking
0:41:11 systems
0:41:11 background
0:41:12 give you
0:41:12 really
0:41:12 really
0:41:12 good
0:41:13 understanding
0:41:13 of
0:41:14 kind
0:41:14 of
0:41:14 the
0:41:14 bounds
0:41:14 on
0:41:15 these
0:41:16 but like
0:41:17 the toolkit
0:41:17 that you
0:41:18 draw from
0:41:19 is like
0:41:20 information
0:41:20 theory
0:41:21 and like
0:41:22 more formal
0:41:24 have you
0:41:25 found that
0:41:25 the AI
0:41:26 community
0:41:27 is receptive
0:41:27 to this
0:41:28 or is it
0:41:28 like
0:41:29 two different
0:41:30 cultures
0:41:30 two different
0:41:31 planets
0:41:31 trying to
0:41:32 communicate
0:41:33 and not a lot
0:41:33 of common
0:41:33 ground
0:41:34 like how
0:41:34 have you
0:41:35 found
0:41:36 like bringing
0:41:36 like the
0:41:36 networking
0:41:37 view of
0:41:37 the world
0:41:38 to the
0:41:38 AI
0:41:39 realm
0:41:40 some
0:41:40 of
0:41:41 them
0:41:41 are
0:41:41 receptive
0:41:42 to it
0:41:42 definitely
0:41:43 but
0:41:47 you know
0:41:49 these
0:41:49 large
0:41:50 conferences
0:41:50 and their
0:41:50 reviewing
0:41:51 processes
0:41:51 it’s so
0:41:52 random
0:41:53 and the
0:41:53 kind of
0:41:54 questions
0:41:54 they ask
0:41:55 you know
0:41:55 I’m a
0:41:56 modeling
0:41:56 person
0:41:57 I like
0:41:58 to bottle
0:41:58 things
0:41:59 and you
0:41:59 know
0:42:00 I submitted
0:42:01 one version
0:42:01 of this
0:42:01 work to
0:42:03 one very
0:42:03 famous
0:42:06 machine learning
0:42:06 or AI
0:42:07 conference
0:42:08 and the
0:42:08 reviewer
0:42:09 said okay
0:42:09 this is
0:42:09 a
0:42:09 model
0:42:09 so
0:42:10 what
0:42:15 so
0:42:16 there
0:42:17 is
0:42:20 that’s
0:42:20 absolutely
0:42:21 remarkable
0:42:21 so like
0:42:22 you’ve
0:42:22 actually
0:42:22 taken a
0:42:23 system
0:42:23 that
0:42:23 nobody
0:42:24 understands
0:42:24 we have
0:42:25 no
0:42:25 models
0:42:25 for
0:42:25 you
0:42:25 actually
0:42:26 provided
0:42:26 some
0:42:26 model
0:42:27 that
0:42:27 we
0:42:27 can
0:42:27 use
0:42:27 to
0:42:27 analyze
0:42:28 it
0:42:28 and
0:42:29 that
0:42:29 alone
0:42:29 wasn’t
0:42:30 sufficient
0:42:30 they’re
0:42:31 asking
0:42:31 so
0:42:31 where
0:42:31 are
0:42:32 the
0:42:32 large
0:42:32 scale
0:42:33 experiments
0:42:33 to
0:42:34 prove
0:42:35 this
0:42:35 I do
0:42:36 listen
0:42:36 I honestly
0:42:37 I mean
0:42:37 I find
0:42:38 there’s
0:42:38 so much
0:42:39 empiricism
0:42:39 in
0:42:40 the
0:42:41 current
0:42:42 AI
0:42:43 community
0:42:43 exactly
0:42:44 because we
0:42:44 don’t
0:42:44 understand
0:42:44 the
0:42:45 systems
0:42:45 you know
0:42:46 it kind
0:42:46 of reminds
0:42:46 me
0:42:47 I feel
0:42:47 like
0:42:49 systems
0:42:49 went the
0:42:50 other way
0:42:50 right
0:42:50 it’s like
0:42:51 we had
0:42:51 all of
0:42:51 these
0:42:51 models
0:42:52 but then
0:42:52 we didn’t
0:42:52 understand
0:42:53 how the
0:42:53 systems
0:42:53 worked
0:42:54 and then
0:42:54 we just
0:42:55 actually did
0:42:55 measurement
0:42:55 it feels
0:42:56 like
0:42:57 ML
0:42:58 or the
0:42:58 AI stuff
0:42:58 is the
0:42:58 opposite
0:42:59 which is
0:42:59 like
0:42:59 we know
0:43:00 we don’t
0:43:00 understand
0:43:00 them
0:43:01 and so
0:43:01 we just
0:43:01 measure
0:43:01 them
0:43:02 but now
0:43:02 we’re
0:43:02 trying
0:43:02 to
0:43:03 come up
0:43:03 with
0:43:03 the
0:43:04 models
0:43:06 yeah
0:43:07 exactly
0:43:07 so
0:43:08 it was
0:43:08 so easy
0:43:09 in some
0:43:09 sense
0:43:10 to build
0:43:10 these
0:43:11 artifacts
0:43:12 and then
0:43:13 just measure
0:43:13 them
0:43:13 that people
0:43:14 have been
0:43:14 going around
0:43:15 trying to
0:43:16 do that
0:43:17 and
0:43:19 one term
0:43:19 I really
0:43:20 dislike
0:43:21 is prompt
0:43:22 engineering
0:43:23 why
0:43:23 you know
0:43:24 engineering
0:43:25 used to
0:43:25 mean
0:43:26 sending
0:43:26 a man
0:43:27 to the
0:43:27 moon
0:43:27 or
0:43:28 providing
0:43:28 five
0:43:28 nines
0:43:29 reliability
0:43:30 prompt
0:43:31 engineering
0:43:31 is prompt
0:43:32 twiddling
0:43:34 you fiddle
0:43:34 with a
0:43:34 prompt
0:43:35 and the
0:43:35 bogus
0:43:36 changes
0:43:37 and the
0:43:37 inference
0:43:38 the output
0:43:39 changes
0:43:39 and you
0:43:40 know
0:43:40 you have
0:43:40 like
0:43:41 hundreds
0:43:41 of
0:43:41 papers
0:43:41 just
0:43:43 just
0:43:44 you know
0:43:44 doing one
0:43:45 experiment
0:43:45 on the
0:43:45 other
0:43:46 changing
0:43:46 a
0:43:46 prompt
0:43:46 this
0:43:46 way
0:43:47 that
0:43:47 way
0:43:47 and
0:43:48 writing
0:43:48 their
0:43:48 observations
0:43:49 and
0:43:50 as a
0:43:50 result
0:43:51 you know
0:43:51 lots of
0:43:52 these
0:43:52 papers
0:43:53 are being
0:43:53 written
0:43:54 are being
0:43:54 submitted
0:43:55 for review
0:43:55 reviewers get
0:43:56 busy
0:43:56 looking at
0:43:57 all this
0:43:57 kind of
0:43:57 empirical
0:43:58 work
0:43:59 and
0:44:00 my
0:44:00 personal
0:44:01 taste
0:44:01 is
0:44:02 to
0:44:02 first
0:44:03 try to
0:44:04 understand
0:44:05 model it
0:44:07 and then
0:44:07 you can
0:44:07 do
0:44:08 the other
0:44:08 thing
0:44:09 so like
0:44:09 a true
0:44:09 theory
0:44:10 guy
0:44:12 I don’t
0:44:12 know about
0:44:12 this bit
0:44:13 twiddling
0:44:15 let me
0:44:16 ask one
0:44:16 more
0:44:16 LLM
0:44:16 question
0:44:17 which is
0:44:18 are there
0:44:18 any
0:44:19 benchmarks
0:44:19 or real
0:44:20 world
0:44:20 tasks
0:44:21 that if
0:44:21 they
0:44:22 occurred
0:44:22 you’d
0:44:23 sort of
0:44:23 reevaluate
0:44:23 and say
0:44:24 hey
0:44:25 maybe
0:44:25 LLMs
0:44:25 are
0:44:26 closer to
0:44:26 the path
0:44:27 to AGI
0:44:27 than I
0:44:27 thought
0:44:31 if
0:44:32 the
0:44:33 real
0:44:33 world
0:44:33 tasks
0:44:41 good
0:44:41 question
0:44:48 you know
0:44:48 which
0:44:50 for
0:44:52 LLMs
0:44:54 or these
0:44:54 models
0:44:55 the
0:44:56 one
0:44:58 domain
0:44:58 where you
0:44:58 have
0:44:59 the
0:44:59 most
0:44:59 training
0:45:00 data
0:45:00 is
0:45:01 probably
0:45:01 coding
0:45:04 and
0:45:04 coding
0:45:04 is
0:45:05 where
0:45:08 you
0:45:08 can
0:45:08 also
0:45:09 have
0:45:09 the
0:45:09 most
0:45:10 structure
0:45:11 and
0:45:12 yet
0:45:13 anyone
0:45:13 who
0:45:14 has
0:45:14 used
0:45:15 these
0:45:16 tools
0:45:16 whether
0:45:16 it’s
0:45:16 cursor
0:45:17 or
0:45:17 whatever
0:45:18 or
0:45:18 cloud
0:45:18 code
0:45:21 LLMs
0:45:21 continue
0:45:21 to
0:45:22 hallucinate
0:45:22 continue
0:45:23 to
0:45:23 generate
0:45:24 unreasonable
0:45:24 code
0:45:25 you
0:45:25 know
0:45:25 you
0:45:25 have
0:45:26 to
0:45:28 you
0:45:28 have
0:45:28 to
0:45:28 constantly
0:45:31 babysit
0:45:31 these
0:45:32 models
0:45:33 so
0:45:34 the
0:45:34 day
0:45:35 an
0:45:35 LLM
0:45:35 can
0:45:36 create
0:45:36 a
0:45:37 large
0:45:37 software
0:45:37 project
0:45:38 without
0:45:38 any
0:45:42 babysitting
0:45:43 is the
0:45:43 day I’ll
0:45:44 be a little
0:45:44 bit convinced
0:45:45 that it’s
0:45:46 too much
0:45:46 easier
0:45:46 but
0:45:47 again
0:45:49 I don’t
0:45:49 think
0:45:50 it’ll be
0:45:51 able to
0:45:51 create
0:45:51 new
0:45:52 science
0:45:52 if it
0:45:52 does
0:45:54 that’s
0:45:54 when I’ll
0:45:54 be
0:45:54 convinced
0:45:54 I
0:45:55 think
0:45:56 you can
0:45:56 almost
0:45:56 take
0:45:56 a
0:45:57 definitional
0:45:57 approach
0:45:57 to answer
0:45:58 this
0:45:58 question
0:45:58 Vishal
0:45:59 like
0:45:59 the
0:46:00 problem
0:46:00 with
0:46:00 these
0:46:00 types
0:46:00 of
0:46:01 questions
0:46:01 is
0:46:02 if you
0:46:02 have
0:46:02 billions
0:46:02 of
0:46:03 dollars
0:46:03 and you
0:46:03 can
0:46:04 collect
0:46:04 whatever
0:46:04 data
0:46:04 you
0:46:05 want
0:46:05 you
0:46:05 can
0:46:05 make
0:46:05 a
0:46:05 model
0:46:05 do
0:46:06 anything
0:46:06 you
0:46:06 want
0:46:07 right
0:46:07 and so
0:46:08 so
0:46:08 you
0:46:09 know
0:46:09 at
0:46:10 some
0:46:10 level
0:46:11 you
0:46:11 have
0:46:11 got
0:46:11 this
0:46:12 entire
0:46:13 capital
0:46:15 structure
0:46:15 machinery
0:46:16 behind
0:46:16 these
0:46:16 models
0:46:16 so
0:46:17 you’re
0:46:17 like
0:46:17 oh
0:46:17 it
0:46:17 can
0:46:18 be
0:46:18 good
0:46:18 at
0:46:18 science
0:46:18 well
0:46:18 sure
0:46:19 you put
0:46:19 a billion
0:46:19 dollars
0:46:19 in solving
0:46:20 materials
0:46:21 science
0:46:21 and collect
0:46:21 all this
0:46:21 data
0:46:22 you’ll be
0:46:22 good
0:46:22 at
0:46:22 material
0:46:22 science
0:46:23 or
0:46:23 whatever
0:46:24 it
0:46:24 is
0:46:24 but
0:46:25 there
0:46:25 is
0:46:25 a
0:46:26 definitional
0:46:26 answer
0:46:26 which
0:46:27 is
0:46:28 and
0:46:29 I’m
0:46:29 going
0:46:29 to
0:46:29 draw
0:46:30 from
0:46:30 your
0:46:30 work
0:46:30 which
0:46:31 is
0:46:31 there
0:46:31 is
0:46:31 a
0:46:31 manifold
0:46:32 that’s
0:46:32 in
0:46:32 there
0:46:33 based
0:46:33 on
0:46:33 the
0:46:33 data
0:46:34 it’s
0:46:34 been
0:46:34 trading
0:46:34 on
0:46:35 and
0:46:35 then
0:46:35 the
0:46:36 question
0:46:36 is
0:46:36 if
0:46:36 it
0:46:37 ever
0:46:37 produces
0:46:37 something
0:46:38 that’s
0:46:38 off
0:46:38 like
0:46:38 a
0:46:38 new
0:46:39 manifold
0:46:40 so
0:46:41 considering
0:46:41 the
0:46:41 existing
0:46:42 traded
0:46:42 data
0:46:42 if
0:46:42 it
0:46:43 ever
0:46:43 does
0:46:43 that
0:46:43 if
0:46:43 it
0:46:43 does
0:46:44 something
0:46:44 that’s
0:46:44 outside
0:46:44 of
0:46:44 that
0:46:45 distribution
0:46:46 then
0:46:47 clearly
0:46:47 we’re
0:46:47 on
0:46:47 a
0:46:47 path
0:46:48 to
0:46:48 learning
0:46:49 new
0:46:49 things
0:46:49 and
0:46:49 if
0:46:50 not
0:46:50 then
0:46:50 everything
0:46:50 is
0:46:50 just
0:46:50 a
0:46:51 computational
0:46:51 step
0:46:51 from
0:46:52 what’s
0:46:52 already
0:46:52 known
0:46:55 I guess
0:46:57 the counter
0:46:57 to that
0:46:58 would be
0:46:58 maybe
0:46:58 all
0:46:58 humans
0:46:59 do
0:46:59 is
0:47:00 work
0:47:00 on
0:47:00 their
0:47:00 own
0:47:00 manifold
0:47:01 and
0:47:02 Einstein
0:47:04 was
0:47:05 lucky
0:47:05 or
0:47:05 something
0:47:06 I guess
0:47:06 would be
0:47:06 the
0:47:06 counter
0:47:06 to
0:47:07 that
0:47:09 there’s
0:47:09 several
0:47:10 many
0:47:11 Einstein
0:47:11 examples
0:47:11 and
0:47:12 it’s
0:47:13 creating
0:47:13 this
0:47:13 new
0:47:13 manifold
0:47:13 I
0:47:14 didn’t
0:47:14 want
0:47:14 to
0:47:14 use
0:47:14 that
0:47:15 definitional
0:47:15 answer
0:47:15 I
0:47:15 thought
0:47:16 it
0:47:16 might
0:47:16 sound
0:47:17 too
0:47:19 wonky
0:47:19 too
0:47:20 mathematical
0:47:21 but
0:47:21 essentially
0:47:23 if
0:47:24 LLMs
0:47:24 really
0:47:25 created
0:47:25 this
0:47:25 new
0:47:25 manifold
0:47:27 then
0:47:27 I
0:47:27 would
0:47:27 be
0:47:28 convinced
0:47:29 but
0:47:30 so far
0:47:30 they have
0:47:30 just
0:47:31 gotten
0:47:31 better
0:47:31 at
0:47:32 navigating
0:47:32 the
0:47:32 existing
0:47:33 manifold
0:47:33 the
0:47:33 existing
0:47:34 training
0:47:34 set
0:47:34 which
0:47:34 is
0:47:35 hugely
0:47:35 powerful
0:47:35 and
0:47:35 is
0:47:35 going
0:47:35 to
0:47:36 change
0:47:36 the
0:47:36 world
0:47:37 I’m
0:47:37 not
0:47:38 denying
0:47:38 that
0:47:38 I
0:47:38 think
0:47:38 they
0:47:38 are
0:47:39 extremely
0:47:40 good
0:47:41 at
0:47:41 what
0:47:41 they
0:47:41 can
0:47:41 do
0:47:42 but
0:47:42 there’s
0:47:42 a
0:47:42 limit
0:47:42 to
0:47:43 what
0:47:43 they
0:47:43 can
0:47:43 do
0:47:44 so
0:47:44 I
0:47:44 have
0:47:44 one
0:47:44 quick
0:47:44 question
0:47:45 what’s
0:47:45 next
0:47:45 for
0:47:45 you
0:47:46 you’ve
0:47:47 tackled
0:47:47 in
0:47:48 context
0:47:48 learning
0:47:48 you’ve
0:47:48 got
0:47:48 a
0:47:49 model
0:47:49 for
0:47:49 LLMs
0:47:49 and
0:47:49 I’ve
0:47:50 got
0:47:50 a
0:47:50 generalized
0:47:51 model
0:47:51 for
0:47:52 their
0:47:53 solution
0:47:53 space
0:47:54 what
0:47:54 what
0:47:54 what are
0:47:54 you
0:47:54 thinking
0:47:55 about
0:47:55 tackling
0:47:56 next
0:47:58 in
0:47:58 terms
0:47:58 of
0:47:59 modeling
0:48:00 or
0:48:02 academically
0:48:02 an
0:48:03 LLM
0:48:03 academically
0:48:06 I’m
0:48:09 you know
0:48:09 I’m
0:48:09 thinking
0:48:10 of
0:48:10 this
0:48:10 what
0:48:11 is
0:48:11 the
0:48:11 architectural
0:48:12 leap
0:48:12 that is
0:48:13 needed
0:48:15 to create
0:48:15 this
0:48:16 new
0:48:16 manifold
0:48:17 and
0:48:17 how
0:48:17 do
0:48:18 we
0:48:18 use
0:48:19 multimodal
0:48:20 data
0:48:21 awesome
0:48:22 to
0:48:23 expand
0:48:23 I’ll
0:48:24 come back
0:48:24 and talk
0:48:25 to us
0:48:25 that’s
0:48:26 right
0:48:27 so I
0:48:28 mean
0:48:29 you know
0:48:29 even
0:48:29 with
0:48:31 LLMs
0:48:31 you know
0:48:31 in the
0:48:32 paper
0:48:32 we say
0:48:32 that
0:48:33 you can
0:48:34 improve
0:48:35 the
0:48:36 inference
0:48:37 by
0:48:37 following
0:48:38 this
0:48:39 low
0:48:39 or
0:48:39 minimum
0:48:40 entropy
0:48:40 path
0:48:42 so that’s
0:48:42 a very
0:48:43 sort of
0:48:43 small
0:48:43 step
0:48:44 that we
0:48:44 are
0:48:44 taking
0:48:44 you know
0:48:45 we are
0:48:46 building
0:48:46 and
0:48:46 training
0:48:46 models
0:48:47 that will
0:48:47 do
0:48:47 inference
0:48:48 based on
0:48:50 the
0:48:50 entropy
0:48:50 path
0:48:51 yeah
0:48:52 by the way
0:48:52 is
0:48:53 model probe
0:48:54 still up
0:48:56 token probe
0:48:56 yeah yeah
0:48:56 token probe
0:48:57 is still up
0:48:58 and you
0:48:58 can see
0:48:59 actually
0:48:59 the
0:49:00 you know
0:49:01 token probe
0:49:02 is a
0:49:02 software
0:49:03 that we
0:49:03 built
0:49:03 and thanks
0:49:04 to
0:49:04 Martin
0:49:04 and
0:49:05 A16Z’s
0:49:06 generosity
0:49:07 it’s running
0:49:07 on your
0:49:08 servers
0:49:09 and anyone
0:49:09 can go
0:49:10 and test
0:49:11 and what
0:49:11 we have
0:49:11 done there
0:49:12 is we
0:49:12 actually
0:49:13 show
0:49:13 the
0:49:14 entropy
0:49:15 yeah
0:49:16 it is so
0:49:16 enlightening
0:49:17 I recommend
0:49:17 anybody
0:49:17 listening to
0:49:18 this
0:49:18 who’s
0:49:18 interested
0:49:18 actually
0:49:19 check out
0:49:19 token probe
0:49:20 it only
0:49:21 shows you
0:49:21 the confidence
0:49:22 yeah
0:49:23 as you go
0:49:23 along
0:49:24 it’s remarkable
0:49:25 so in context
0:49:25 learning
0:49:26 you know
0:49:26 you create
0:49:27 your new
0:49:27 DSL
0:49:28 and you
0:49:29 give it
0:49:29 to the
0:49:29 prompt
0:49:30 and you
0:49:30 can see
0:49:31 the confidence
0:49:32 rising
0:49:33 with each
0:49:33 new example
0:49:34 the entropy
0:49:35 reducing
0:49:36 and that
0:49:36 sort of
0:49:37 is a validation
0:49:38 of the
0:49:38 model
0:49:39 you can
0:49:39 see it
0:49:40 sort of
0:49:41 unfurling
0:49:41 and right
0:49:42 in front
0:49:42 of your
0:49:42 eyes
0:49:43 so token
0:49:43 probe
0:49:43 is
0:49:44 thanks
0:49:45 thanks again
0:49:46 Michelle
0:49:46 thanks so
0:49:47 much for
0:49:47 coming on
0:49:47 the podcast
0:49:48 it was a great
0:49:48 conversation
0:49:49 it was great
0:49:50 fun
0:49:50 thank you
0:49:51 thank you so
0:49:51 much again
0:49:54 thanks for
0:49:55 listening to
0:49:55 this episode
0:49:56 of the
0:49:56 A16Z
0:49:57 podcast
0:49:58 if you
0:49:58 liked this
0:49:58 episode
0:49:59 be sure
0:49:59 to like
0:50:00 comment
0:50:00 subscribe
0:50:01 leave us
0:50:01 a rating
0:50:02 or review
0:50:03 and share
0:50:03 it with
0:50:03 your friends
0:50:04 and family
0:50:05 for more
0:50:05 episodes
0:50:06 go to
0:50:06 YouTube
0:50:07 Apple
0:50:07 Podcast
0:50:08 and Spotify
0:50:09 follow us
0:50:10 on X
0:50:10 at A16Z
0:50:11 and subscribe
0:50:12 to our
0:50:12 Substack
0:50:12 at
0:50:14 a16z.substack.com
0:50:15 thanks again
0:50:16 for listening
0:50:16 and I’ll see
0:50:17 you in the
0:50:17 next episode
0:50:19 as a reminder
0:50:20 the content
0:50:20 here is for
0:50:21 informational
0:50:22 purposes only
0:50:22 should not
0:50:23 be taken
0:50:23 as legal
0:50:23 business
0:50:24 tax
0:50:24 or investment
0:50:25 advice
0:50:26 or be used
0:50:26 to evaluate
0:50:27 any investment
0:50:27 or security
0:50:28 and is not
0:50:29 directed at
0:50:29 any investors
0:50:30 or potential
0:50:30 investors
0:50:31 in any
0:50:32 A16Z
0:50:32 fund
0:50:33 please note
0:50:34 that A16Z
0:50:34 and its
0:50:34 affiliates
0:50:35 may also
0:50:35 maintain
0:50:35 investments
0:50:36 in the
0:50:36 companies
0:50:36 discussed
0:50:37 in this
0:50:37 podcast
0:50:38 for more
0:50:38 details
0:50:39 including
0:50:39 a link
0:50:40 to our
0:50:40 investments
0:50:41 please
0:50:41 see
0:50:42 a16z.com
0:50:44 forward slash
0:50:44 disclosures
0:51:21 Thank you.