Latest / Razib Khan's Unsupervised Learning / Nikolai Yakovenko: the stillborn promise of the LLM age
Transcript
- 0:00SPEAKER_02: This podcast is brought to you by the Albany Public Library Main Branch and the
- 0:04SPEAKER_02: generosity of listeners like you.
- 0:09SPEAKER_02: What is a podcast?
- 0:16SPEAKER_02: God, Daddy, these people talk as much as you do.
- 0:26SPEAKER_02: Razeeb Khan's Unsupervised Learning
- 0:29SPEAKER_00: Even if you and your partner are healthy,
- 0:30SPEAKER_00: there's still a chance your child can develop a serious genetic disease.
- 0:33SPEAKER_00: This is because every embryo has new changes not present in either parent.
- 0:37SPEAKER_00: Most of the time, these are benign, but sometimes they can be catastrophic.
- 0:40SPEAKER_00: Orkin's whole genome embryo reports directly screen the embryo and analyze these de
- 0:45SPEAKER_00: novo genetic mutations.
- 0:46SPEAKER_00: Discuss embryo screening and IVF with a genetics expert.
- 0:50SPEAKER_00: Hey, everybody.
- 0:50SPEAKER_00: This is Razeeb with the Unsupervised Learning Podcast.
- 0:54SPEAKER_00: And I am here now with a returning guest.
- 0:56SPEAKER_00: I think – my friend Nikolai Yakovenko here.
- 0:59SPEAKER_00: I think Nikolai might be tied with Samuel Buria as the most common, the most frequent –
- 1:07SPEAKER_00: guest on unsupervised learning.
- 1:10SPEAKER_00: And the reason that he is such a frequent guest here is I like to talk about
- 1:16SPEAKER_00: technology and artificial intelligence with him.
- 1:19SPEAKER_00: Obviously, if you read what I write, I am not a specialist or a
- 1:26SPEAKER_00: I don't feel like I have specialized knowledge,
- 1:28SPEAKER_00: so that's why I would bring Nikolai and people like Nick Casimides on the podcast.
- 1:33SPEAKER_00: And I mentioned Nikolai and both Nick Casimides because I advise both their companies.
- 1:38SPEAKER_00: I want to just get that out of the way.
- 1:40SPEAKER_00: You see it if you follow me on X. You see it if you go to resume.com.
- 1:44SPEAKER_00: And so there's that connection there to make it explicit.
- 1:48SPEAKER_00: So Nikolai right now is going to be –
- 1:51SPEAKER_00: He'll be talking about artificial intelligence, and we'll get to all those great details.
- 1:55SPEAKER_00: But he also has a company that he's leaning pretty hard into right now,
- 1:59SPEAKER_00: Deep News,
- 2:00SPEAKER_00: which is tied to artificial intelligence.
- 2:03SPEAKER_00: Makes sense because he has been in the field for a while now.
- 2:06SPEAKER_00: And, yeah, Nikolai, could you –
- 2:09SPEAKER_00: Talk about deep news, why you're excited.
- 2:13SPEAKER_00: Obviously, I am excited.
- 2:14SPEAKER_00: I am an advisor.
- 2:16SPEAKER_00: I just told the listeners that.
- 2:18SPEAKER_00: Tell people why you're excited,
- 2:19SPEAKER_00: and then let's also talk about all the excitement that's been going on in AI and
- 2:25SPEAKER_00: the various big players that we have gotten to know over the last year and a half.
- 2:31SPEAKER_01: There is.
- 2:32SPEAKER_01: Yeah.
- 2:32SPEAKER_01: So as a lot of you sort of know,
- 2:35SPEAKER_01: these these LLM models,
- 2:37SPEAKER_01: large language models are pretty amazing and can do a lot of stuff.
- 2:41SPEAKER_01: But,
- 2:41SPEAKER_01: you know,
- 2:42SPEAKER_01: one thing that you notice is that they explicitly have a knowledge cut off and they
- 2:46SPEAKER_01: don't really talk about recent things.
- 2:47SPEAKER_01: Like if you ask chat GBT, what's what happened today?
- 2:51SPEAKER_01: It doesn't give you very good answers.
- 2:52SPEAKER_01: I mean, it can query Bing and whatever.
- 2:55SPEAKER_01: So I thought it would be kind of interesting starting a year ago to go the exact
- 3:00SPEAKER_01: opposite and say,
- 3:01SPEAKER_01: hey,
- 3:01SPEAKER_01: can we use AI to actually curate exactly what's going on?
- 3:05SPEAKER_01: So that's kind of where Deep News was born from.
- 3:07SPEAKER_01: So we look at a large number of sources.
- 3:10SPEAKER_01: We write stories and maintain it.
- 3:12SPEAKER_01: And I think what's really cool is you can use AI to really cover all kinds of topics,
- 3:17SPEAKER_01: including pretty niche things.
- 3:19SPEAKER_01: So we cover...
- 3:21SPEAKER_01: like hundreds of topics that we label.
- 3:24SPEAKER_01: And one thing that people really like is you can get your specific things.
- 3:28SPEAKER_01: You can say, hey, what's going on in sports or golf or San Francisco or AI or genomics.
- 3:35SPEAKER_01: We will have that coverage.
- 3:36SPEAKER_01: Now,
- 3:36SPEAKER_01: obviously you can look at a lot of specialized sources that do that as well,
- 3:40SPEAKER_01: but I think it's pretty cool to have that succinct,
- 3:43SPEAKER_01: done by AI,
- 3:44SPEAKER_01: you know,
- 3:45SPEAKER_01: more or less in real time,
- 3:47SPEAKER_01: you know,
- 3:47SPEAKER_01: super,
- 3:48SPEAKER_01: like I said,
- 3:48SPEAKER_01: succinct,
- 3:49SPEAKER_01: these articles are short,
- 3:50SPEAKER_01: you know,
- 3:50SPEAKER_01: they do link to the longer sources when that's appropriate.
- 3:53SPEAKER_01: And also,
- 3:54SPEAKER_01: I think that there is like an element of sort of kind of an open source neutrality
- 3:58SPEAKER_01: to it as well as we really try to be to cover sort of everything you may want to know.
- 4:03SPEAKER_01: But people do seem to be,
- 4:04SPEAKER_01: you know,
- 4:05SPEAKER_01: very,
- 4:05SPEAKER_01: very excited about the ability to get their specific niches as well as the big things.
- 4:10SPEAKER_00: Yeah, I think that's key.
- 4:12SPEAKER_00: One of the things with AI is obviously it's not perfect,
- 4:16SPEAKER_00: but there are a lot of these LLMs,
- 4:20SPEAKER_00: there are a lot of tasks where good enough is sufficient.
- 4:24SPEAKER_00: So if I am developing engineering for a nuclear submarine, I do want perfect.
- 4:30SPEAKER_00: But if I want the latest stories on HDNA or something,
- 4:35SPEAKER_00: I think good enough would actually be sufficient if there's a few random things
- 4:38SPEAKER_00: Oddballs in there, it doesn't really matter.
- 4:40SPEAKER_00: I really want 90-95%.
- 4:42SPEAKER_00: And so I think there are certain issues that we're seeing.
- 4:45SPEAKER_00: AI and LLMs are already starting to be useful.
- 4:48SPEAKER_00: I have a search on Feedly for certain topics, mostly science-related.
- 4:54SPEAKER_00: So there are already...
- 4:57SPEAKER_00: attempts to do this sort of thing with more primitive recommendations, top-down queries and whatnot.
- 5:04SPEAKER_00: I think in the future, though, you know what this is?
- 5:07SPEAKER_00: Do you remember push technology?
- 5:11SPEAKER_01: Is that a specific name?
- 5:13SPEAKER_00: Well,
- 5:14SPEAKER_00: this was like a specific phase on the internet,
- 5:16SPEAKER_00: probably 1996,
- 5:17SPEAKER_00: a little before maybe you were paying a lot of attention.
- 5:20SPEAKER_00: And there was this thing where you would install that,
- 5:23SPEAKER_00: well,
- 5:24SPEAKER_00: it's an app,
- 5:24SPEAKER_00: but we would call it an app then.
- 5:26SPEAKER_00: And it would like bring all this news and email and everything onto your computer.
- 5:31SPEAKER_00: The problem with push technology is it just wasn't very specific.
- 5:36SPEAKER_00: And also like the bandwidth was like out of control.
- 5:40SPEAKER_00: So eventually it just faded away.
- 5:42SPEAKER_00: And people went back to the web portal idea.
- 5:44SPEAKER_00: But in some ways, the need for both push and pull are always there.
- 5:53SPEAKER_00: Younger people, for example, they do not like to proactively subscribe to things.
- 5:58SPEAKER_00: They do love the YouTube recommendation algorithm, right?
- 6:02SPEAKER_01: Absolutely.
- 6:02SPEAKER_01: That's right.
- 6:03SPEAKER_01: Yeah.
- 6:03SPEAKER_01: And you're right.
- 6:04SPEAKER_01: The news is something where it's definitely a combination of push and pull.
- 6:08SPEAKER_01: You want to search things,
- 6:09SPEAKER_01: you want to find things,
- 6:10SPEAKER_01: but you also just want to check,
- 6:11SPEAKER_01: hey,
- 6:11SPEAKER_01: what's going on?
- 6:13SPEAKER_01: And I think to your point about being good enough,
- 6:15SPEAKER_01: I mean,
- 6:16SPEAKER_01: that sounds like a little bit of a faint praise,
- 6:18SPEAKER_01: but it's actually the right idea.
- 6:21SPEAKER_01: A few years ago, this was impossible.
- 6:23SPEAKER_01: Now it's relatively,
- 6:25SPEAKER_01: I shouldn't say easy,
- 6:26SPEAKER_01: of course,
- 6:26SPEAKER_01: I shouldn't undersell what we're doing,
- 6:28SPEAKER_01: but with a relatively small team and a
- 6:30SPEAKER_01: fairly reasonable budget,
- 6:32SPEAKER_01: you can sort of cover everything that's going on,
- 6:34SPEAKER_01: not just at a surface level,
- 6:36SPEAKER_01: but relatively deeply and in a way where you can't really miss anything.
- 6:40SPEAKER_01: So one example I use personally is something like law news.
- 6:44SPEAKER_01: I've always been a bit of a law nerd.
- 6:46SPEAKER_01: I'm very interested in what's going on with the SCOTUS, interesting decisions, things like that.
- 6:50SPEAKER_01: That's Supreme Court for...
- 6:52SPEAKER_01: uh, for, for non-Americans.
- 6:53SPEAKER_01: And I'm like,
- 6:55SPEAKER_01: you could always,
- 6:55SPEAKER_01: you know,
- 6:55SPEAKER_01: read law blogs,
- 6:56SPEAKER_01: subscribe to things,
- 6:58SPEAKER_01: listen to podcasts,
- 6:59SPEAKER_01: but it's a little bit too much into the weeds.
- 7:00SPEAKER_01: Like I've tried to do it.
- 7:01SPEAKER_01: I've never gotten into it.
- 7:02SPEAKER_01: And now with our coverage,
- 7:04SPEAKER_01: I just know that if I check that tab,
- 7:05SPEAKER_01: I'm going to know what are the decisions that matter,
- 7:07SPEAKER_01: where the pending decisions that matter.
- 7:09SPEAKER_01: I can always read more.
- 7:11SPEAKER_01: It brings in the commentary from the law blogs where appropriate.
- 7:13SPEAKER_01: Obviously, sometimes things cross over for various reasons.
- 7:17SPEAKER_01: You may have a Supreme Court case that involves a celebrity.
- 7:20SPEAKER_01: For example,
- 7:21SPEAKER_01: there was a music IP case a couple weeks ago that was kind of interesting where the
- 7:24SPEAKER_01: plaintiff was Flo Rida,
- 7:26SPEAKER_01: basically.
- 7:26SPEAKER_01: So that obviously got more coverage.
- 7:28SPEAKER_01: And you're going to get all of that stuff in a succinct, super readable format, which I think...
- 7:34SPEAKER_01: A human could do, but it would be a lot of work.
- 7:37SPEAKER_01: And I think the AI is quite good at it.
- 7:38SPEAKER_01: It's consistent.
- 7:40SPEAKER_01: I mean, I started my career at Google Search 20 years ago.
- 7:43SPEAKER_01: And that's something that Amit Singhal would always say,
- 7:45SPEAKER_01: who was the head of Google Search,
- 7:47SPEAKER_01: was like,
- 7:47SPEAKER_01: hey,
- 7:49SPEAKER_01: guys,
- 7:49SPEAKER_01: a human could do this better,
- 7:52SPEAKER_01: but it doesn't really make sense.
- 7:54SPEAKER_01: So our machine learning, it was kind of like a humbling thing.
- 7:56SPEAKER_01: We're trying to do a decent estimation of what an expert would do, but do it very fast.
- 8:02SPEAKER_01: very consistently.
- 8:03SPEAKER_01: And the consistency I think is key.
- 8:05SPEAKER_01: When you're reading a law blog, you are sort of swayed by that person's opinion.
- 8:09SPEAKER_01: But if you're like,
- 8:10SPEAKER_01: hey,
- 8:10SPEAKER_01: you've got seven to 10 law professors talking about it and you can see them all in
- 8:15SPEAKER_01: one place and have that be synthesized,
- 8:17SPEAKER_01: I think it's pretty cool.
- 8:18SPEAKER_00: Yeah, can you – I want to update.
- 8:22SPEAKER_00: Let's do a status check on the big AI companies right now,
- 8:27SPEAKER_00: like OpenAI,
- 8:27SPEAKER_00: et cetera,
- 8:28SPEAKER_00: because Claude,
- 8:29SPEAKER_00: as we're recording,
- 8:30SPEAKER_00: Claude just had an update.
- 8:33SPEAKER_00: Can you do a quick refresher for my listener?
- 8:37SPEAKER_00: Because this is the first time you've been on here.
- 8:39SPEAKER_00: I do have more questions.
- 8:41SPEAKER_00: you know, people, more subscribers, et cetera, et cetera.
- 8:45SPEAKER_00: What is an LLM?
- 8:46SPEAKER_00: And talk about really quickly how it's,
- 8:49SPEAKER_00: because artificial intelligence is a big,
- 8:50SPEAKER_00: broad category of a bunch of different things,
- 8:53SPEAKER_00: a bundle of different methods.
- 8:55SPEAKER_00: Talk about how LLMs, like their trajectory of like when they got big and what they do in general.
- 9:00SPEAKER_01: That's right.
- 9:00SPEAKER_01: Yeah.
- 9:02SPEAKER_01: So obviously, like deep learning AI are things that people are pretty aware of.
- 9:05SPEAKER_01: We're really focusing pretty much on LLM, which stands for large language models.
- 9:09SPEAKER_01: So at a high level, this technology hasn't really changed in the last couple of years.
- 9:12SPEAKER_01: The idea is you train this very large model on, let's just say, the whole Internet, more or less.
- 9:19SPEAKER_01: And it's able to do a few things.
- 9:23SPEAKER_01: Most specifically is that it can answer questions and follow directions,
- 9:27SPEAKER_01: but it also has a tremendous amount of knowledge.
- 9:29SPEAKER_01: So just as a stylized fact,
- 9:32SPEAKER_01: any of these top LLMs,
- 9:33SPEAKER_01: and I'll get into what those are in a second,
- 9:35SPEAKER_01: if you ask it like a basic question that can be answered from a popular Wikipedia page,
- 9:40SPEAKER_01: Like it can basically recite all that information verbatim.
- 9:43SPEAKER_01: Like it has that much knowledge.
- 9:45SPEAKER_01: So like think of maybe not the whole internet, but a lot of the information is in there.
- 9:50SPEAKER_01: You know,
- 9:51SPEAKER_01: every like lyrics to a famous song,
- 9:52SPEAKER_01: even though that's an interesting case because of these restrictions,
- 9:55SPEAKER_01: this hobbling stuff,
- 9:56SPEAKER_01: which we'll also get into.
- 9:57SPEAKER_01: Sometimes it'll be like, hey, I know the song, but I won't tell you.
- 10:01SPEAKER_01: So you can sort of get into why that happens.
- 10:04SPEAKER_01: But at a high level,
- 10:05SPEAKER_01: the LLM technology was just this kind of amazing intuition that you could train a model.
- 10:10SPEAKER_01: I mean, these models are quite large.
- 10:12SPEAKER_01: They take many millions of dollars in just electricity to train on clusters that
- 10:17SPEAKER_01: cost at this point into the billions of dollars.
- 10:21SPEAKER_01: But the point is, you take the thing, you run it through the internet.
- 10:27SPEAKER_01: And it basically like some,
- 10:29SPEAKER_01: you know,
- 10:29SPEAKER_01: zips all of that information into something that can be accessed basically in real time.
- 10:34SPEAKER_00: Yeah.
- 10:34SPEAKER_00: Um, so I think, uh,
- 10:37SPEAKER_00: I think the first podcast probably we did was either at the end of 2022 or very early in 2023.
- 10:45SPEAKER_00: The LLMs had kind of – well, it was ChatGPT.
- 10:51SPEAKER_00: ChatGPT 3 slash 3.5.
- 10:54SPEAKER_00: And so the genius of OpenAI and ChatGPT is they realized, oh, we can get the public involved.
- 11:01SPEAKER_00: And, you know, public doesn't care about an API.
- 11:04SPEAKER_00: Public doesn't know what API stands for.
- 11:06SPEAKER_00: But public knows how to interface in a chat window.
- 11:10SPEAKER_00: That exploded really quickly, and the other companies kind of raced behind it.
- 11:16SPEAKER_00: Obviously, Google with Gemini tried to do that.
- 11:19SPEAKER_00: And now Claude out of Anthropic has come up, which is Claude is the AI from Anthropic.
- 11:25SPEAKER_00: And so there's all of these companies that are releasing and then matching each other.
- 11:31SPEAKER_00: And there's also a whole proliferation of companies like MidJourney, which is like Dolly.
- 11:37SPEAKER_00: It does the visualization stuff.
- 11:40SPEAKER_00: Those aren't LLMs, though, right?
- 11:41SPEAKER_00: Those are different.
- 11:42SPEAKER_00: The stable diffusion.
- 11:44SPEAKER_00: What does the visual ones do?
- 11:46SPEAKER_00: Like what are they using?
- 11:48SPEAKER_01: Right.
- 11:49SPEAKER_01: So there's just sort of this parallel path of basically image understanding and image generation.
- 11:56SPEAKER_01: Like you need to have understanding to do generation,
- 11:58SPEAKER_01: but fundamentally the product there is you type something in and it draws a picture
- 12:02SPEAKER_01: that matches your description.
- 12:05SPEAKER_01: It's a slightly different path.
- 12:06SPEAKER_01: Now,
- 12:06SPEAKER_01: these things do get merged because the standard now for GPT-4.0 and I think O
- 12:13SPEAKER_01: somehow stands for multimodal.
- 12:14SPEAKER_01: I'm not sure exactly how that translates into O,
- 12:18SPEAKER_01: but it also does have this nice feature where you can understand an image and
- 12:21SPEAKER_01: communicate in one model with images and text.
- 12:26SPEAKER_01: But I think, not to say that's not important, I think that's...
- 12:30SPEAKER_01: Not that important, honestly.
- 12:32SPEAKER_01: I think just the text form itself is quite interesting, quite powerful.
- 12:37SPEAKER_01: I can sort of explain why I think that's the most important.
- 12:40SPEAKER_01: But words are sort of how we think and how we record information.
- 12:44SPEAKER_01: The vast majority of human information on the web is text.
- 12:48SPEAKER_01: So,
- 12:50SPEAKER_01: you know,
- 12:50SPEAKER_01: we can get into the weeds on the image stuff and some of the multi-language stuff,
- 12:54SPEAKER_01: but even if you restrict yourself just to language and just to English,
- 12:58SPEAKER_01: still quite a bit to do.
- 12:59SPEAKER_01: So you said it correctly,
- 13:03SPEAKER_01: like GPT-3 was really sort of the breakthrough,
- 13:06SPEAKER_01: but at first it just had all this knowledge,
- 13:08SPEAKER_01: but it wasn't a product.
- 13:10SPEAKER_01: So chat GPT with GPT 3.5,
- 13:13SPEAKER_01: what they did is there's a human alignment where they said,
- 13:15SPEAKER_01: hey,
- 13:15SPEAKER_01: take this model and tune it to get good at question-answer pairs,
- 13:22SPEAKER_01: which basically means they gave it some examples they cooked up,
- 13:24SPEAKER_01: and then they sort of did this like...
- 13:26SPEAKER_01: self-repeating loop where once it became decent at generating answers and they
- 13:30SPEAKER_01: would sort of have a generate different answers and have humans rate,
- 13:33SPEAKER_01: which is the best one.
- 13:34SPEAKER_01: And you do a few cycles of that.
- 13:35SPEAKER_01: Um, and that's how we get these, these modern, you know, LLMs that we're used to like, um,
- 13:41SPEAKER_01: like ChatGPT, Gemini, and Unplugged, like you said.
- 13:45SPEAKER_01: So fundamentally, yes, there's two components.
- 13:47SPEAKER_01: You have this pre-training where you train it on the whole web where it gains the knowledge,
- 13:50SPEAKER_01: you know,
- 13:51SPEAKER_01: the knowledge,
- 13:51SPEAKER_01: right?
- 13:52SPEAKER_01: And then you do this thing where you align it to human responses,
- 13:56SPEAKER_01: everything from just teaching basic Q&A to also embedding all kinds of,
- 14:01SPEAKER_01: well,
- 14:02SPEAKER_01: let's just say value judgments.
- 14:04SPEAKER_01: It's both inevitable, but I would argue that some of these companies may be...
- 14:09SPEAKER_01: you know focus on that in in a way that you know isn't always great yeah well uh i
- 14:15SPEAKER_00: think okay the most extreme case obviously was gemini's uh uh anti-racist uh
- 14:22SPEAKER_00: historical personages and so uh do you want to talk about the black vikings and the
- 14:28SPEAKER_00: black nazis
- 14:32SPEAKER_01: Yeah, that was – honestly, well, you can explain that if you want.
- 14:37SPEAKER_01: But I think that was – it was great because it was so extreme.
- 14:40SPEAKER_01: Like it's sort of – what is it?
- 14:44SPEAKER_01: What is the London quote, right, that heightened the differences?
- 14:46SPEAKER_01: So to the point where it actually triggered a reorg at Google where it was such an
- 14:51SPEAKER_01: embarrassment for Google that they actually kind of –
- 14:55SPEAKER_01: had to reorganize their organization and actually change the leadership of the
- 15:00SPEAKER_01: alignment team and bring it more into alignment with the goals of the company
- 15:04SPEAKER_01: because it was just too ridiculous.
- 15:07SPEAKER_00: Yeah,
- 15:08SPEAKER_00: basically it was – you would put in a query like show me some Waffen-SS and Google
- 15:14SPEAKER_00: would be – Google would say make them black.
- 15:22SPEAKER_01: That's right.
- 15:23SPEAKER_01: It would rewrite it, right?
- 15:24SPEAKER_01: So like Ben Thompson had a great analysis of this.
- 15:26SPEAKER_01: He even had a screenshot because there was a bug where it actually showed the
- 15:29SPEAKER_01: rewriting that was supposed to be hidden,
- 15:30SPEAKER_01: but briefly it showed it.
- 15:31SPEAKER_01: So you would type in, hey, my favorite was, you know...
- 15:36SPEAKER_01: this is a Canadian GPT, right?
- 15:38SPEAKER_01: Like style.
- 15:38SPEAKER_01: You said, Hey, like make us a picture of a, um, of a bunch of people playing hockey.
- 15:44SPEAKER_01: And it would rewrite your, it would take your query.
- 15:46SPEAKER_01: It could generate the answer perfectly reasonably well,
- 15:48SPEAKER_01: but it would rewrite it as,
- 15:50SPEAKER_01: you know,
- 15:50SPEAKER_01: show us a hockey scene,
- 15:52SPEAKER_01: make sure that you include like diversity of gender and,
- 15:55SPEAKER_01: um,
- 15:56SPEAKER_01: ethnic identity.
- 15:57SPEAKER_01: So you, so you literally ended up with a, with an Indian lady in a COVID mask playing hockey.
- 16:04SPEAKER_01: Which makes no sense, but it's kind of hilarious.
- 16:06SPEAKER_00: I mean, yeah, that's the issue.
- 16:07SPEAKER_00: It's inadvertently hilarious.
- 16:10SPEAKER_00: Inadvertently hilarious.
- 16:13SPEAKER_00: Yeah, so, you know, we'll go back to the technical aspects.
- 16:18SPEAKER_00: I've been talking to you about this.
- 16:20SPEAKER_00: Well,
- 16:20SPEAKER_00: I mean,
- 16:21SPEAKER_00: I think we've been talking in a way,
- 16:22SPEAKER_00: like,
- 16:22SPEAKER_00: on and off for two years,
- 16:23SPEAKER_00: but,
- 16:23SPEAKER_00: you know,
- 16:23SPEAKER_00: it definitely broke open with OpenAI and,
- 16:28SPEAKER_00: you know,
- 16:29SPEAKER_00: 3.3.5,
- 16:30SPEAKER_00: ChatGPT.
- 16:34SPEAKER_00: Do you think we are where we were supposed to be in terms of world transformative
- 16:44SPEAKER_00: impact with artificial intelligence in summer of 2024,
- 16:47SPEAKER_00: a year and a half out?
- 16:47SPEAKER_00: Where are we?
- 16:52SPEAKER_01: Yeah, so the short answer is no, absolutely not.
- 16:56SPEAKER_01: I think that a lot of interesting things have happened.
- 16:58SPEAKER_01: And I think it's actually interesting to go through what has happened in the past 18 months.
- 17:03SPEAKER_01: But the trajectory has been absolutely not what was predicted, right?
- 17:08SPEAKER_01: The prediction was that these models are going to get like crazy better every three
- 17:11SPEAKER_01: months or every six months.
- 17:13SPEAKER_01: The previous version blows the future, the past one away.
- 17:16SPEAKER_01: You have this like runaway winner with open AI, etc.
- 17:20SPEAKER_01: And that's not what's happened at all.
- 17:23SPEAKER_01: What has happened is that the models have gotten better,
- 17:27SPEAKER_01: but the pace of improvement at these core tasks has almost certainly slowed down.
- 17:32SPEAKER_01: But as a result of that, you have two interesting effects.
- 17:37SPEAKER_01: One is you have a lot more interesting technology developed sort of around
- 17:40SPEAKER_01: applications and also measurement.
- 17:43SPEAKER_01: For example, LMSIS has been created, which is kind of gold standard now.
- 17:46SPEAKER_01: I'll get into why that's interesting in a second.
- 17:48SPEAKER_01: But I think in a very good way,
- 17:52SPEAKER_01: it's sort of debatable whether open AI sort of like let their lead slip or whether
- 17:56SPEAKER_01: it was just inevitable,
- 17:57SPEAKER_01: but now you have a situation where it's more of a,
- 18:00SPEAKER_01: you know,
- 18:00SPEAKER_01: of an oligarchy where you have,
- 18:02SPEAKER_01: you know,
- 18:03SPEAKER_01: three to six to maybe if you're really,
- 18:05SPEAKER_01: really aggressive,
- 18:05SPEAKER_01: maybe in as many as 12 or 13 organizations that have the possibility of training a
- 18:10SPEAKER_01: world-class AI model that's basically on par with GPT-4.
- 18:15SPEAKER_01: Now, some are a little bit better than others.
- 18:18SPEAKER_01: Some have advantages.
- 18:19SPEAKER_01: And then you can get into really specific things like having it be small enough to be,
- 18:22SPEAKER_01: to run locally on your,
- 18:24SPEAKER_01: You know, on your MacBook or next year, even your iPhone.
- 18:27SPEAKER_01: But even for the high end models,
- 18:29SPEAKER_01: like the truth is they haven't really gotten that much bigger and they haven't
- 18:33SPEAKER_01: gotten that much better.
- 18:35SPEAKER_01: So as a result, it's a little bit more like cars, right?
- 18:37SPEAKER_01: Like the analogy would be instead of like Ford running away with a Ford model,
- 18:42SPEAKER_01: you know,
- 18:43SPEAKER_01: with a consumer car market the way they did with the Model T.
- 18:46SPEAKER_01: at the beginning i'm not even saying this is true but sort of people think this is
- 18:49SPEAKER_01: what happened right yeah then you have sort of you have a bunch of entries you had
- 18:53SPEAKER_01: gm enter and you end up in a situation where you have you know um at first you know
- 18:59SPEAKER_01: at peak probably probably before the consolidation how many car companies that we
- 19:03SPEAKER_01: have in america that were legit probably like you know between six to twelve right
- 19:07SPEAKER_01: yeah yeah it got consolidated to three to four so
- 19:11SPEAKER_01: And each model has a little bit of a difference.
- 19:14SPEAKER_01: They're all kind of more or less the same.
- 19:16SPEAKER_01: They do the same thing.
- 19:17SPEAKER_01: I like this car analogy because LLMs are a bit more like cars than like airplanes.
- 19:22SPEAKER_01: Airplanes are very specialized, right?
- 19:24SPEAKER_01: Cars kind of all do the same thing, but they have small differences in features.
- 19:27SPEAKER_01: Like your F-150 versus your smart car are kind of different, but they're kind of the same.
- 19:32SPEAKER_01: Like they follow the same rules, they drive in the same rows, they do the same stuff.
- 19:37SPEAKER_01: So that's kind of what we've seen.
- 19:39SPEAKER_01: So for example,
- 19:39SPEAKER_01: if you look at the LMS leaderboards,
- 19:41SPEAKER_01: which we should explain what that is,
- 19:42SPEAKER_01: I think it's actually very interesting.
- 19:44SPEAKER_01: Yeah, I mean, the lead is OpenAI, followed by Google, followed by OpenAI, followed by Anthropic.
- 19:50SPEAKER_01: And then you have an entry from like 01AI, which is a company out of China.
- 19:58SPEAKER_01: You have a model from Facebook,
- 20:01SPEAKER_01: you have a model from Nvidia,
- 20:02SPEAKER_01: and then it sort of tails off a little bit.
- 20:05SPEAKER_01: And these models are all, according to, again, LMSIS, which is blind user ratings, about the same.
- 20:11SPEAKER_01: There's differences.
- 20:12SPEAKER_01: I'm not saying one isn't better than the other one.
- 20:14SPEAKER_01: If you set up a chess tournament and you had them play 20 games,
- 20:19SPEAKER_01: the best model would win, but it wouldn't win all 20 games, if that makes sense, right?
- 20:23SPEAKER_01: You would think it's, again, it's like a chess tournament.
- 20:24SPEAKER_01: You'd have a lot of draws and like, you know, you'd win.
- 20:27SPEAKER_01: Yeah, there's parity.
- 20:28SPEAKER_01: There's parity.
- 20:29SPEAKER_01: Tremendous amount of parity.
- 20:30SPEAKER_01: And I think that's kind of interesting from the product perspective,
- 20:33SPEAKER_01: because that's when you get into differentiation.
- 20:35SPEAKER_01: And I think already,
- 20:36SPEAKER_01: like OpenAI has,
- 20:37SPEAKER_01: I think,
- 20:37SPEAKER_01: tacitly admitted this,
- 20:38SPEAKER_01: or at least part of their organization,
- 20:40SPEAKER_01: Um, they've become more of a consumer company.
- 20:43SPEAKER_01: Like their advantage at this point is already branding.
- 20:45SPEAKER_01: You know,
- 20:46SPEAKER_01: the stylized fact is I think if you go and you talk to random people in America who
- 20:49SPEAKER_01: are outside of sort of the Silicon Valley bubble,
- 20:51SPEAKER_01: I'd say a high percentage of them,
- 20:53SPEAKER_01: I don't know if it's 20% or 60%,
- 20:56SPEAKER_01: but a lot of them would have heard of Chad GPT.
- 20:58SPEAKER_01: Um,
- 20:59SPEAKER_01: I'd be willing to bet that none of them have heard of Anthropic, like less than 1%.
- 21:04SPEAKER_01: But these are competitive models.
- 21:07SPEAKER_01: So I think OpenAI is already taking advantage of that.
- 21:09SPEAKER_01: They have to deal with Apple for distribution.
- 21:12SPEAKER_01: So it's already becoming this sort of consumer tech thing where you differentiate
- 21:15SPEAKER_01: on speed,
- 21:16SPEAKER_01: on cost,
- 21:16SPEAKER_01: on convenience,
- 21:17SPEAKER_01: on branding,
- 21:18SPEAKER_01: as opposed to what the story was supposed to be,
- 21:20SPEAKER_01: which is like...
- 21:21SPEAKER_01: You know, every new model from OpenAI is going to be revolutionary.
- 21:24SPEAKER_01: And that has just not been the case.
- 21:27SPEAKER_01: Now,
- 21:27SPEAKER_01: I think some people are hoping or thinking or expecting that there's a chance that
- 21:31SPEAKER_01: GPT-5 will blow people out of the water.
- 21:33SPEAKER_01: But the longer that doesn't happen, the less likely that becomes.
- 21:36SPEAKER_01: And to be honest, already GPT-3 to GPT-4, while a big improvement,
- 21:41SPEAKER_01: was a lot less than zero to GPT-3.5.
- 21:45SPEAKER_01: And in fact,
- 21:47SPEAKER_01: if you actually break it down by tasks,
- 21:49SPEAKER_01: there's many useful tasks that became possible with LLMs that were impossible before,
- 21:54SPEAKER_01: where GPT-3.5 does just fine.
- 21:57SPEAKER_01: So for example,
- 21:57SPEAKER_01: if you ask for sentiment,
- 21:58SPEAKER_01: if you ask for some sort of summarization,
- 22:01SPEAKER_01: like a lot of people still use GPT-3.5 until OpenAI phases that out in API setting.
- 22:06SPEAKER_01: Because if you want to set up some sort of just LLM flow and say your app,
- 22:11SPEAKER_01: the user is not really going to know the difference in many cases.
- 22:14SPEAKER_01: Like, even GPT-4, while bigger and better and stuff, it wasn't a revolution, really.
- 22:18SPEAKER_00: Yeah, let me just reiterate something.
- 22:22SPEAKER_00: So the secret sauce of these companies is their models, right?
- 22:27SPEAKER_00: But really, the key with these big companies is the training.
- 22:33SPEAKER_00: The training is expensive.
- 22:35SPEAKER_00: Very expensive.
- 22:37SPEAKER_00: That's their moat.
- 22:38SPEAKER_00: That's really their moat, right?
- 22:39SPEAKER_00: It's the training.
- 22:41SPEAKER_01: So I think I think the data is a really interesting case.
- 22:45SPEAKER_01: I don't really want to get into the weeds and the law stuff because I frankly, I'm not an expert on it.
- 22:49SPEAKER_01: But, you know, in a stylized way, basically think of it like pre 2021 data.
- 22:56SPEAKER_01: You know,
- 22:56SPEAKER_01: you have common crawl,
- 22:57SPEAKER_01: you had all this stuff,
- 22:58SPEAKER_01: basically anything that Google could index,
- 23:00SPEAKER_01: you could realistically train an LLM on.
- 23:04SPEAKER_01: you had even giant data sets of things like Reddit, right?
- 23:07SPEAKER_01: Like you could download all of Reddit and train models on it.
- 23:10SPEAKER_01: We were doing this 10 years ago, we were doing this seven years ago.
- 23:13SPEAKER_01: This data set was kind of undervalued,
- 23:15SPEAKER_01: underutilized because before LLMs,
- 23:17SPEAKER_01: it really wasn't like that valuable.
- 23:19SPEAKER_01: You know, it was cool,
- 23:21SPEAKER_01: But now all these things are shut off.
- 23:24SPEAKER_01: So you've probably heard about the deals that OpenAI has made with the publishers,
- 23:28SPEAKER_01: you know,
- 23:28SPEAKER_01: with Reddit,
- 23:29SPEAKER_01: things like that.
- 23:29SPEAKER_01: So like going forward,
- 23:31SPEAKER_01: if you want models that sort of have knowledge after a certain sort of like,
- 23:35SPEAKER_01: you know,
- 23:35SPEAKER_01: think of it as like a BCC kind of thing,
- 23:38SPEAKER_01: right?
- 23:38SPEAKER_01: There's like a hard date.
- 23:39SPEAKER_01: And after that, data access becomes a lot more difficult.
- 23:41SPEAKER_01: So I think the remote...
- 23:43SPEAKER_01: is probably going to be if you want your model to, you know, it's a data, right?
- 23:48SPEAKER_01: Same for coding.
- 23:49SPEAKER_01: Sorry,
- 23:49SPEAKER_01: I'm jumping around a little bit,
- 23:50SPEAKER_01: but like probably the biggest,
- 23:52SPEAKER_01: so where I will sort of go say there was kind of a real breakthrough was on the
- 23:57SPEAKER_01: coding side.
- 23:58SPEAKER_01: GPT three and a half was not very good at writing code.
- 24:01SPEAKER_01: Like the fact that it could do it at all was okay.
- 24:03SPEAKER_01: But since then it's gotten like way better.
- 24:05SPEAKER_01: And part of it is access to data, right?
- 24:08SPEAKER_01: So GitHub Copilot,
- 24:10SPEAKER_01: Okay, I don't know what the thing is, but presumably it trained on GitHub, right?
- 24:14SPEAKER_01: Like actually getting the data, training on the good code, getting access to that.
- 24:18SPEAKER_01: So I think going forward,
- 24:21SPEAKER_01: yeah,
- 24:21SPEAKER_01: like there's probably unlikely to be a giant breakthrough in modeling,
- 24:24SPEAKER_01: although you never know.
- 24:25SPEAKER_01: So the real boundary is, yes, the cost of training and the access to great data.
- 24:32SPEAKER_00: So I do have to say, in terms of coding, okay, this is...
- 24:39SPEAKER_00: Maybe you can offer some insight.
- 24:41SPEAKER_00: Maybe some listeners can.
- 24:43SPEAKER_00: You know, I have always been not the greatest at regular expressions, you know.
- 24:49SPEAKER_00: And so – but I have to say these – I mostly use ChatGPT 4.0 now.
- 24:55SPEAKER_00: I was using Cloud for a while.
- 24:56SPEAKER_00: I go back and forth.
- 24:57SPEAKER_00: I have –
- 24:58SPEAKER_00: Pro accounts to all of these.
- 25:00SPEAKER_00: But I don't – they do regular expressions really well.
- 25:02SPEAKER_00: I have to say it's made a big difference in some of the work that I do – well,
- 25:09SPEAKER_00: candidly,
- 25:09SPEAKER_00: some of the consulting projects and genomics that I do where there's some pattern
- 25:13SPEAKER_00: matching and some data format shifts and stuff like that.
- 25:16SPEAKER_00: And I always had –
- 25:18SPEAKER_00: I could get it done.
- 25:19SPEAKER_00: I could write a script that would do the conversion, but now I have to say it's much faster.
- 25:23SPEAKER_00: So there are all these little things that are happening on the margin.
- 25:25SPEAKER_00: I don't want to undersell,
- 25:29SPEAKER_00: I think,
- 25:30SPEAKER_00: where we are and all the little things that are happening in the aggregate due to
- 25:35SPEAKER_00: the accessibility of this – accessibility of the power of this AI technology.
- 25:40SPEAKER_00: I will say –
- 25:41SPEAKER_00: I said this on social media,
- 25:42SPEAKER_00: I said this on X,
- 25:43SPEAKER_00: there were people at the end of 2022 in group chats,
- 25:47SPEAKER_00: not in public that I saw,
- 25:49SPEAKER_00: although some people are crazy in public were saying that,
- 25:51SPEAKER_00: but there were people at the end of 2022 who thought that AGI would be here by the summer,
- 25:58SPEAKER_00: that AGI would arrive at the summer of 2023 at the latest,
- 26:04SPEAKER_00: and that our species would be over.
- 26:07SPEAKER_00: You know,
- 26:08SPEAKER_00: this is from people in the Bay Area who are in particular circles at parties where
- 26:15SPEAKER_00: Altman would be,
- 26:16SPEAKER_00: where people like Nat Friedman would be.
- 26:18SPEAKER_00: I'm not going to name the person, but I'm not making this up as a generic person.
- 26:21SPEAKER_00: This is like a real person.
- 26:23SPEAKER_00: They're not stupid.
- 26:24SPEAKER_00: They're smart.
- 26:25SPEAKER_00: But there was some crazy thinking that got unleashed by ChatGPT.
- 26:31SPEAKER_01: Oh, totally.
- 26:32SPEAKER_01: And I mean, and you still sort of see remnants of that.
- 26:35SPEAKER_01: And I'm not saying that idea is dead or will ever die.
- 26:38SPEAKER_01: And I do think we should sort of take it seriously.
- 26:42SPEAKER_01: Even if you don't believe in AGI at all, I think at least like the metagame of it is interesting.
- 26:46SPEAKER_01: If enough people do, then it's a big deal, sort of like even in a cynical perspective.
- 26:50SPEAKER_01: But I mean, even now, I mean, I don't live in the Bay Area, but when I visit...
- 26:54SPEAKER_01: I think it's almost inevitable to go to an event and a party and you'll have somebody,
- 26:57SPEAKER_01: almost always a very smart person who's quite young,
- 27:00SPEAKER_01: hence idealistic,
- 27:02SPEAKER_01: who will give a version of either A,
- 27:05SPEAKER_01: AGI is here and we just didn't notice and it just sort of got lost.
- 27:09SPEAKER_01: Basically, it's like if we had the right incantation, we would have it already.
- 27:12SPEAKER_01: Or,
- 27:14SPEAKER_01: or they will basically give a version of either like,
- 27:16SPEAKER_01: with one small improvement,
- 27:17SPEAKER_01: it can do these amazing things,
- 27:19SPEAKER_01: or what people will often say to me,
- 27:20SPEAKER_01: actually,
- 27:20SPEAKER_01: specifically,
- 27:22SPEAKER_01: almost always happens at one of these events is they will claim with a straight face,
- 27:25SPEAKER_01: and they clearly believe it,
- 27:26SPEAKER_01: that like AI can write,
- 27:28SPEAKER_01: you know,
- 27:28SPEAKER_01: better music than professional musicians or something along those lines.
- 27:32SPEAKER_01: which is clearly not... Like pseudo.
- 27:37SPEAKER_01: Yes, exactly.
- 27:38SPEAKER_01: And that's always their take.
- 27:38SPEAKER_01: And they're like, have you heard of pseudo?
- 27:40SPEAKER_01: I'm like, yeah, I've heard of pseudo.
- 27:41SPEAKER_01: I've looked at it.
- 27:43SPEAKER_01: But it's just a weird thing where if you poll the audience among informed people,
- 27:49SPEAKER_01: Very few people would actually agree with that statement,
- 27:51SPEAKER_01: but there always are a few people who,
- 27:53SPEAKER_01: like,
- 27:53SPEAKER_01: extremely believe it.
- 27:54SPEAKER_01: And I think,
- 27:55SPEAKER_01: you know,
- 27:55SPEAKER_01: not to play psychologist,
- 27:56SPEAKER_01: but I think the reason that they do is that,
- 28:00SPEAKER_01: you know,
- 28:00SPEAKER_01: it's consistent with their other beliefs,
- 28:02SPEAKER_01: right?
- 28:03SPEAKER_01: Like, I think that, I mean, like...
- 28:06SPEAKER_01: Because it's hard, right?
- 28:07SPEAKER_01: I mean, what's going on now is hard to comprehend.
- 28:09SPEAKER_01: On the one hand,
- 28:10SPEAKER_01: the models are getting better and they're clearly doing certain jobs,
- 28:12SPEAKER_01: like some software engineering jobs that clearly can do.
- 28:16SPEAKER_01: Another thing is you see these tests, right?
- 28:18SPEAKER_01: I think OpenAI has been talking a lot about this recently.
- 28:20SPEAKER_01: If you've seen the, you know, these interview clips with like Amira Maradi.
- 28:24SPEAKER_01: But they're like,
- 28:25SPEAKER_01: well,
- 28:25SPEAKER_01: you know,
- 28:25SPEAKER_01: you have the path and you're basically like measuring the intelligence level,
- 28:28SPEAKER_01: giving an IQ test,
- 28:29SPEAKER_01: like good ones,
- 28:30SPEAKER_01: like proper IQ tests,
- 28:31SPEAKER_01: giving it test exams.
- 28:32SPEAKER_01: And you're like, well, you know, before it was the intelligence of a middle schooler.
- 28:37SPEAKER_01: Now it's the intelligence of a high schooler.
- 28:38SPEAKER_01: Now it's a college.
- 28:40SPEAKER_01: Next, it'll get a Ph.D.
- 28:42SPEAKER_01: And I think the problem is, you know, like it's it's you.
- 28:47SPEAKER_01: I think people who are more from a prediction market or finance background or just
- 28:53SPEAKER_01: kind of like,
- 28:54SPEAKER_01: I would just say practical people don't really have a problem synthesizing this
- 28:57SPEAKER_01: information that that is clearly true.
- 29:00SPEAKER_01: And yet other things are not true.
- 29:01SPEAKER_01: But I think people who want to have a very simple,
- 29:03SPEAKER_01: very cohesive world vision,
- 29:04SPEAKER_01: you sort of have to say that AGI is going to happen because you just draw the line
- 29:07SPEAKER_01: and you extrapolate.
- 29:09SPEAKER_01: But the truth is, A, you can't extrapolate that far.
- 29:12SPEAKER_01: not everything behaves like Moore's law, you know?
- 29:15SPEAKER_01: Um,
- 29:15SPEAKER_01: and two,
- 29:16SPEAKER_01: I just think that these are like really useful heuristics,
- 29:19SPEAKER_01: but computers just are intelligent differently.
- 29:22SPEAKER_01: For example,
- 29:22SPEAKER_01: like as a chess player,
- 29:25SPEAKER_01: pretty much my whole life,
- 29:26SPEAKER_01: computers have been better at chess than like most humans.
- 29:30SPEAKER_01: And for a lot of it better than all humans, you know?
- 29:32SPEAKER_01: Um, and so I, has that really, has that, you know, has have computer transformed society?
- 29:37SPEAKER_01: Absolutely.
- 29:38SPEAKER_01: By, uh,
- 29:39SPEAKER_01: We've seen nothing remotely resembling AGI from that,
- 29:42SPEAKER_01: like crossover between those tasks has not happened.
- 29:45SPEAKER_01: So,
- 29:46SPEAKER_01: you know,
- 29:47SPEAKER_01: another thing I'll say and I'll stop rambling is like for a human,
- 29:50SPEAKER_01: I mean,
- 29:50SPEAKER_01: you know,
- 29:51SPEAKER_01: this a lot of things are super correlated.
- 29:53SPEAKER_01: For example,
- 29:53SPEAKER_01: you can give somebody a memory test and it's very highly correlated with like a
- 29:56SPEAKER_01: proper IQ test.
- 29:57SPEAKER_01: There's noise, but the correlation is very, very high.
- 30:00SPEAKER_01: I know they gave me these tests when the brain injury stuff, right?
- 30:04SPEAKER_01: Like, but for computers, it's not correlated at all.
- 30:08SPEAKER_01: Like you have models of perfect recall 10,
- 30:09SPEAKER_01: 20 years ago that can do any reasoning at all,
- 30:14SPEAKER_01: but there is no human alive that can remember all kinds of things that has no
- 30:18SPEAKER_01: reasoning ability.
- 30:19SPEAKER_01: So the computer is just sort of like think differently.
- 30:23SPEAKER_01: And I, having said that, I do think the general, the, the generalization of the LLM is different.
- 30:30SPEAKER_01: I do think these models are quite different than the specialty chess playing models.
- 30:34SPEAKER_01: But again,
- 30:36SPEAKER_01: I just personally,
- 30:37SPEAKER_01: I guess,
- 30:37SPEAKER_01: don't really have a problem with this conflicting information that they're clearly
- 30:41SPEAKER_01: getting smarter and doing more tasks.
- 30:43SPEAKER_01: And yet it doesn't necessarily mean that they'll take over the world.
- 30:47SPEAKER_01: But I think people who just haven't been in the space as long,
- 30:50SPEAKER_01: it's easier to have an aligned vision,
- 30:52SPEAKER_01: right?
- 30:53SPEAKER_00: So I think one thing that I have started to think,
- 30:56SPEAKER_00: it's kind of what you said,
- 30:57SPEAKER_00: you know,
- 30:58SPEAKER_00: we use cognitive science uses computational analogy or metaphor for the brain,
- 31:05SPEAKER_00: human brain.
- 31:06SPEAKER_00: And obviously with artificial intelligence, we think of it as like our brain.
- 31:12SPEAKER_00: Yeah, we can't help it.
- 31:15SPEAKER_00: Yeah.
- 31:16SPEAKER_00: But the issue is like perhaps we're creating a – we're not creating –
- 31:23SPEAKER_00: Okay, what the hell is AGI?
- 31:25SPEAKER_00: Whatever.
- 31:26SPEAKER_00: We kind of know what we're talking about,
- 31:27SPEAKER_00: but it seems like we are creating a new form of operation,
- 31:34SPEAKER_00: of cognitive operation of the universe that's just qualitatively different from ours.
- 31:39SPEAKER_00: And so these attempts to, one, benchmark against humans, two,
- 31:46SPEAKER_00: cause problems because they're just qualitatively different now you know it's let
- 31:52SPEAKER_00: me use a metaphor or analogy it's kind of like saying a bird versus a flying fish
- 32:00SPEAKER_00: you know they can both like stay aloft
- 32:04SPEAKER_00: But biomechanically, there's all sorts of differences.
- 32:08SPEAKER_00: Genealogically, they're different.
- 32:10SPEAKER_00: But yeah, they can both go maybe like 40 feet.
- 32:14SPEAKER_00: The bird can go further, obviously.
- 32:16SPEAKER_00: But they can both do this thing.
- 32:18SPEAKER_00: And so...
- 32:20SPEAKER_00: Like I asked – actually,
- 32:22SPEAKER_00: while you were talking about the Gemini and all that and OpenAI,
- 32:26SPEAKER_00: I asked them about the breeder's equation,
- 32:27SPEAKER_00: which is something in genetics.
- 32:28SPEAKER_00: I just like to test them.
- 32:30SPEAKER_00: I'm surprised.
- 32:30SPEAKER_00: Gemini Pro actually did better than Anthropic, did better than Claude or Sunday.
- 32:36SPEAKER_00: I don't know.
- 32:373.5.
- 32:37SPEAKER_00: OpenAI did the best, 4.0.
- 32:41SPEAKER_00: Did the best.
- 32:42SPEAKER_00: But this is something that I understand.
- 32:45SPEAKER_00: And so that's why I asked you the question about something I understand.
- 32:47SPEAKER_00: And it gave an output that was pretty reasonable.
- 32:49SPEAKER_00: As you said, it can read Wikipedia, right?
- 32:52SPEAKER_00: But that's the output.
- 32:54SPEAKER_00: Like what the LLM and how it operates is like nothing like my brain.
- 33:00SPEAKER_00: Nothing.
- 33:01SPEAKER_00: And so I think that's the fundamental –
- 33:04SPEAKER_00: issue that as you move away from – as you move from an AI scientist like you are
- 33:09SPEAKER_00: who understands the guts a bit on some – I'm not going to say surface level because
- 33:16SPEAKER_00: I just want to be careful with what I say because when ChatGPT came out in the fall
- 33:22SPEAKER_00: of 2022,
- 33:23SPEAKER_00: some people in your field who are accomplished scientists
- 33:26SPEAKER_00: were very clear that they had not predicted this,
- 33:28SPEAKER_00: which means that on some level,
- 33:31SPEAKER_00: you guys did not have a full understanding,
- 33:33SPEAKER_00: whatever that means,
- 33:35SPEAKER_00: of the underlying architecture there,
- 33:38SPEAKER_00: right?
- 33:38SPEAKER_00: That was developing organically and then made a breakthrough, quote, quote, a breakthrough, right?
- 33:44SPEAKER_00: And so there's still...
- 33:46SPEAKER_00: We still have to have some uncertainty in our predictions here,
- 33:48SPEAKER_00: but it's summer of 2024.
- 33:50SPEAKER_00: I feel...
- 33:53SPEAKER_00: I feel like things you said,
- 33:55SPEAKER_00: you – I've had several people on,
- 33:58SPEAKER_00: like Nick Cassimides,
- 33:59SPEAKER_00: David Mackay,
- 34:00SPEAKER_00: from a computer science background that have gotten into LLMs or were already
- 34:04SPEAKER_00: well-versed in them.
- 34:05SPEAKER_00: And the technologists tend to be far less into the we are creating a god thing.
- 34:15SPEAKER_00: that will take over and do all these things and more like we are creating a very
- 34:19SPEAKER_00: powerful tool that will be revolutionary.
- 34:23SPEAKER_00: And so I feel it's only a year and a half or whatever.
- 34:27SPEAKER_00: It's not quite two years yet for me to pay attention to this LLM world.
- 34:32SPEAKER_00: But I have to say that I think,
- 34:34SPEAKER_00: yeah,
- 34:35SPEAKER_00: we need to – a little less Yudkowsky and more other things I guess is where I'm at,
- 34:42SPEAKER_00: at least for the next decade.
- 34:43SPEAKER_01: Yeah, well, I think you said it well.
- 34:45SPEAKER_01: I mean, we are creating something new and it is distributed.
- 34:50SPEAKER_01: It's not like a single person created or can stop it.
- 34:53SPEAKER_01: So I think it's actually very reasonable to take some of the cautionary stuff seriously,
- 34:58SPEAKER_01: because at least it's interesting,
- 34:59SPEAKER_01: you know,
- 35:00SPEAKER_01: where it is new.
- 35:01SPEAKER_01: It is kind of a runaway thing.
- 35:04SPEAKER_01: It will transform society eventually.
- 35:06SPEAKER_01: I think all of these things are actually true.
- 35:08SPEAKER_01: Um, and it is a new form of intelligence and it could be very interesting.
- 35:12SPEAKER_01: It'll obviously be very good for some people, very bad for others.
- 35:15SPEAKER_01: Um, but more importantly, it's, it's somewhat inevitable.
- 35:18SPEAKER_01: Now,
- 35:19SPEAKER_01: Eliezer would say it's not,
- 35:21SPEAKER_01: he would say that,
- 35:21SPEAKER_01: you know,
- 35:22SPEAKER_01: if you,
- 35:22SPEAKER_01: uh,
- 35:23SPEAKER_01: if you,
- 35:23SPEAKER_01: what is it,
- 35:24SPEAKER_01: the latest thing that he is saying,
- 35:25SPEAKER_01: which,
- 35:25SPEAKER_01: which I think is quite interesting is,
- 35:27SPEAKER_01: you know,
- 35:28SPEAKER_01: what he wants is he wants governor,
- 35:29SPEAKER_01: uh,
- 35:30SPEAKER_01: governments to put all the GPUs in like one,
- 35:34SPEAKER_01: in one place,
- 35:34SPEAKER_01: uh,
- 35:35SPEAKER_01: So then you can press a single button and destroy them all.
- 35:38SPEAKER_01: He's saying you don't have to destroy them, but you have to have the option of destroying it.
- 35:42SPEAKER_01: I guess he's really against the decentralization of it,
- 35:46SPEAKER_01: like the anti-crypto,
- 35:48SPEAKER_01: essentially,
- 35:50SPEAKER_01: which is an interesting point of view.
- 35:52SPEAKER_01: And I do think that it is a new thing and new things always make differences and new things are scary.
- 35:59SPEAKER_01: And I don't think that's like a wrong opinion necessarily, but it doesn't.
- 36:04SPEAKER_01: Yeah.
- 36:04SPEAKER_01: The point of view that is most,
- 36:06SPEAKER_01: I guess,
- 36:06SPEAKER_01: annoying to me,
- 36:07SPEAKER_01: for lack of a better word,
- 36:09SPEAKER_01: is when people are like,
- 36:10SPEAKER_01: yeah,
- 36:10SPEAKER_01: it's a new thing,
- 36:11SPEAKER_01: but we know what it is.
- 36:12SPEAKER_01: We obviously don't.
- 36:13SPEAKER_01: And the lazy analogy, it's like the human brain and it's just not.
- 36:17SPEAKER_01: But it could still be very powerful, and it could do some stuff.
- 36:21SPEAKER_01: Who knows?
- 36:22SPEAKER_00: I mean, you think it will be powerful.
- 36:27SPEAKER_00: But whatever.
- 36:27SPEAKER_00: I mean, a jet plane is powerful.
- 36:29SPEAKER_00: I mean, there's a lot of technologies out there that are powerful.
- 36:32SPEAKER_00: I think the key is we thought that it was going to create an artificial –
- 36:36SPEAKER_00: analog of us like our brain like our you know i'll give you kowski i give i give
- 36:42SPEAKER_01: him but i i let's give a little credence to his argument um because i do think that
- 36:47SPEAKER_01: there is a fundamental difference between a more powerful car and a more powerful
- 36:51SPEAKER_01: ai model right like one is trying explicitly to be intelligent to take actions on
- 36:56SPEAKER_01: behalf of humans and the other one is not
- 36:58SPEAKER_01: You know,
- 37:00SPEAKER_01: like I think that I'm open to the idea that these are just,
- 37:03SPEAKER_01: you know,
- 37:04SPEAKER_01: it's just input output,
- 37:05SPEAKER_01: right?
- 37:05SPEAKER_01: It's just answering questions.
- 37:07SPEAKER_01: But I don't think that's a convincing argument for the skeptics, right?
- 37:13SPEAKER_01: Because like the AI could,
- 37:14SPEAKER_01: if it was a little bit better,
- 37:15SPEAKER_01: it could go and it can go on Slack in theory,
- 37:18SPEAKER_01: right?
- 37:19SPEAKER_01: Or kind of almost in practice and like take meetings for you.
- 37:22SPEAKER_01: And they could take over my job and tell my team what to do.
- 37:25SPEAKER_01: And you can say, yes, the humans don't have to listen.
- 37:28SPEAKER_01: But you and I know they will.
- 37:29SPEAKER_01: I mean, if we had an AI tell us good ideas, we would probably pay attention to it.
- 37:35SPEAKER_01: It would be like, hey, Razeeb, you should write an essay about this.
- 37:38SPEAKER_01: Actually, I wrote the essay for you.
- 37:40SPEAKER_01: Is this good?
- 37:41SPEAKER_00: All right, all right.
- 37:43SPEAKER_00: Listen to a little bit of this brilliant AI musical.
- 37:46SPEAKER_02: In a light so bright and meaning
- 38:06SPEAKER_02: Coding data,
- 38:08SPEAKER_02: never sleeping Machine lines and circuits creeping Future world in his hands
- 38:13SPEAKER_02: keeping Nikolai's dream,
- 38:15SPEAKER_02: never weeping Programs rise,
- 38:18SPEAKER_02: algorithms fly Nikolai,
- 38:21SPEAKER_02: reaching for the sky He writes the future,
- 38:24SPEAKER_02: don't ask why Nikolai's dream will never die
- 38:37SPEAKER_02: Innovation always flowing through the wires.
- 38:40SPEAKER_02: Knowledge growing.
- 38:42SPEAKER_02: Nikolai keeps on going.
- 38:44SPEAKER_02: In his life, the light is glowing.
- 38:52SPEAKER_02: Endless code he keeps on writing,
- 38:54SPEAKER_02: AI's path he's always guiding,
- 38:57SPEAKER_02: in his eyes a fire igniting,
- 38:59SPEAKER_02: future tech he's always fighting.
- 39:03SPEAKER_02: Programs rise,
- 39:04SPEAKER_02: algorithms fly,
- 39:05SPEAKER_02: Nikolai,
- 39:06SPEAKER_02: reaching for the sky,
- 39:08SPEAKER_02: he writes the future,
- 39:09SPEAKER_02: don't ask why,
- 39:10SPEAKER_02: Nikolai's dream will never die.
- 39:33SPEAKER_02: Yeah, that's seriously not bad.
- 39:35SPEAKER_02: So, I mean, this is a...
- 40:01SPEAKER_00: Yeah, we can stop now.
- 40:02SPEAKER_00: But anyway, some of you have heard – actually, not some of you.
- 40:08SPEAKER_00: You guys have heard the song.
- 40:11SPEAKER_00: So this is the kind of thing – I just gave it a – I gave it –
- 40:16SPEAKER_00: you know, uh, I gave a very simple prop about Nikolai Yakovenko.
- 40:20SPEAKER_00: Uh,
- 40:21SPEAKER_00: uh,
- 40:22SPEAKER_00: and it's created this,
- 40:24SPEAKER_00: uh,
- 40:24SPEAKER_00: you know,
- 40:25SPEAKER_00: AI scientist,
- 40:25SPEAKER_00: Nikolai Yakovenko,
- 40:26SPEAKER_00: I decided to like create the song,
- 40:28SPEAKER_00: the song.
- 40:30SPEAKER_00: It's not bad.
- 40:31SPEAKER_00: Uh, and it, I kind of find it amusing.
- 40:35SPEAKER_00: A lot of the pseudo songs are actually quite amusing and funny.
- 40:38SPEAKER_00: Uh, but, uh,
- 40:41SPEAKER_00: It's very,
- 40:42SPEAKER_00: very hard in my experience,
- 40:44SPEAKER_00: because I do have pro-sudo pro,
- 40:46SPEAKER_00: and I have written a bunch of songs.
- 40:47SPEAKER_00: Or not written, but I've generated songs about my friends just for fun.
- 40:51SPEAKER_00: It is hard to use pseudo to let it just kind of run wild and create something emotionally powerful.
- 41:01SPEAKER_00: It's like a song that pushes you.
- 41:04SPEAKER_01: It's completely vacuous, right?
- 41:06SPEAKER_01: Yeah.
- 41:07SPEAKER_01: But here's the thing.
- 41:08SPEAKER_01: So to give it credit, and I think a lot of the...
- 41:11SPEAKER_01: the researchers that I know they're,
- 41:12SPEAKER_01: you know,
- 41:13SPEAKER_01: like the kind of stuff that you see them get into be like,
- 41:15SPEAKER_01: Hey,
- 41:15SPEAKER_01: yes,
- 41:16SPEAKER_01: but you can sort of give it more iteration,
- 41:18SPEAKER_01: right?
- 41:18SPEAKER_01: Instead of writing just one song,
- 41:20SPEAKER_01: you know,
- 41:20SPEAKER_01: it could,
- 41:21SPEAKER_01: it could write a hundred,
- 41:22SPEAKER_01: it could do its own evaluation,
- 41:23SPEAKER_01: or maybe even you vote like,
- 41:24SPEAKER_01: Ooh,
- 41:25SPEAKER_01: this,
- 41:25SPEAKER_01: this part here is actually good.
- 41:27SPEAKER_01: So like iterate on that, like workshop it.
- 41:29SPEAKER_01: Right.
- 41:30SPEAKER_01: And actually that that's been kind of a big, a big thing in the past year.
- 41:34SPEAKER_01: That's probably most of the excitement and the investments that you see at the
- 41:38SPEAKER_01: you know, at sort of the early stage level, it's things like that.
- 41:41SPEAKER_01: It's like here,
- 41:42SPEAKER_01: basically put more effort,
- 41:43SPEAKER_01: you know,
- 41:44SPEAKER_01: don't just output it,
- 41:45SPEAKER_01: but like iterate on it,
- 41:46SPEAKER_01: spend it,
- 41:47SPEAKER_01: spend an hour or a day on it.
- 41:49SPEAKER_01: This idea of agents where you actually have different AIs, like critique each other.
- 41:53SPEAKER_01: One writes a song and the other one's like, yeah, but the song sucks.
- 41:55SPEAKER_01: And here's why.
- 41:56SPEAKER_01: So try it again.
- 41:57SPEAKER_01: I mean, it's, it's this kind of like moon challenge an idea, but it does work.
- 42:01SPEAKER_01: So, so actually like you can't, you know, you can't squeeze blood out of the stone as they say, but yeah,
- 42:08SPEAKER_01: You can make things, let's just say, 10 times better, 100 times better, whatever, through iteration.
- 42:14SPEAKER_01: And the best example is coding, like you said.
- 42:17SPEAKER_01: You can actually test if the code worked.
- 42:19SPEAKER_01: You can write it, clean it up, et cetera.
- 42:21SPEAKER_01: I mean, I don't personally think that...
- 42:25SPEAKER_01: You know,
- 42:25SPEAKER_01: you can I mean,
- 42:26SPEAKER_01: here,
- 42:26SPEAKER_01: OK,
- 42:27SPEAKER_01: all that otherwise wise here,
- 42:28SPEAKER_01: you take a student of a certain ability or intelligence level.
- 42:31SPEAKER_01: You can coach them up, but only to a certain point.
- 42:33SPEAKER_01: Right.
- 42:34SPEAKER_01: Like any good teacher, any good coach can make an athlete or a student or anyone better.
- 42:39SPEAKER_01: But like within reason.
- 42:42SPEAKER_01: So I think those differences are starting to be squeezed out.
- 42:46SPEAKER_01: But I don't think that they're transformative.
- 42:49SPEAKER_01: But I do think it would be a lot better.
- 42:50SPEAKER_01: Like, to be fair, this is the first cut.
- 42:53SPEAKER_01: And actually,
- 42:55SPEAKER_01: the place where this really comes from is actually going back to the games research,
- 42:58SPEAKER_01: where,
- 42:59SPEAKER_01: you know,
- 42:59SPEAKER_01: if you take the best models for,
- 43:01SPEAKER_01: you know,
- 43:02SPEAKER_01: backgammon or chess or Go or any of that,
- 43:05SPEAKER_01: like,
- 43:05SPEAKER_01: they can't just give you the best move.
- 43:07SPEAKER_01: And that's a pretty good model.
- 43:09SPEAKER_01: What they actually do is they do a lot of local search and simulation.
- 43:13SPEAKER_01: And that's what makes them superhuman, right?
- 43:16SPEAKER_01: Is actually rolling it out, trying, editing, things like that.
- 43:20SPEAKER_01: So there is a big push.
- 43:24SPEAKER_01: Some of you have heard about the Q star concept.
- 43:26SPEAKER_01: That's basically the idea there is, hey,
- 43:29SPEAKER_01: yeah these chatbots are okay but if you actually care about something give it a
- 43:33SPEAKER_01: task and let it sort of you know go to town and actually self-improvement
- 43:37SPEAKER_00: essentially yeah uh one thing i'll say is also uh with sudo uh which is uh you guys
- 43:43SPEAKER_00: can just check it out google sudo ai but basically you know you give it a simple
- 43:47SPEAKER_00: prompt and it creates some genre of music and you know they're not bad it's just
- 43:52SPEAKER_00: they're not great usually but
- 43:54SPEAKER_00: You can put in your custom lyrics.
- 43:55SPEAKER_00: I'm not a great lyricist, but you know who's a better lyricist than me?
- 43:59SPEAKER_00: Claude or Chachi PT.
- 44:00SPEAKER_00: Actually, Claude was the better lyricist than Chachi PT.
- 44:03SPEAKER_00: And so what I do is I actually give a more complicated prompt to Claude,
- 44:08SPEAKER_00: tell Claude to produce lyrics,
- 44:10SPEAKER_00: like produce verses.
- 44:11SPEAKER_00: Claude actually produces good verses compared to Suno, in my opinion.
- 44:14SPEAKER_00: And then I put those into Suno, and then Suno sings it with the instrumentals according to genre.
- 44:20SPEAKER_00: So bringing it together...
- 44:22SPEAKER_00: You can imagine kind of like some sort of module where it's like Claude does the lyrics,
- 44:29SPEAKER_00: Sudo does the vocals,
- 44:31SPEAKER_00: et cetera,
- 44:31SPEAKER_00: et cetera.
- 44:32SPEAKER_00: It will get better.
- 44:33SPEAKER_00: And I think there are parts of artistic creativity where –
- 44:38SPEAKER_00: Yeah, we'll see.
- 44:38SPEAKER_00: I mean,
- 44:39SPEAKER_00: look,
- 44:39SPEAKER_00: a lot of I think a lot of the mass produced lowest common denominator pop music,
- 44:43SPEAKER_00: too,
- 44:44SPEAKER_00: is just frankly not that great.
- 44:46SPEAKER_00: Like, we don't remember the song, you know, the song of the summer from like, you know, 2018.
- 44:52SPEAKER_00: What was it?
- 44:52SPEAKER_00: You know, I mean,
- 44:54SPEAKER_00: I'm not saying that the music was better in the past,
- 44:56SPEAKER_00: but there's something – I think that there is something about quantification,
- 45:01SPEAKER_00: mass production,
- 45:02SPEAKER_00: the economies of scale that have kind of made music explore a narrower parameter
- 45:08SPEAKER_00: space of bubblegum pop geared towards,
- 45:11SPEAKER_00: you know,
- 45:12SPEAKER_00: let's say like mid-teens.
- 45:13SPEAKER_00: I don't know.
- 45:14SPEAKER_00: And I think that –
- 45:16SPEAKER_00: A lot of things changed when the sound scan,
- 45:20SPEAKER_00: you could see music producers and music executives could see exactly what's sold.
- 45:25SPEAKER_00: And in some ways, that was really great because it was giving people what they wanted.
- 45:28SPEAKER_00: But sometimes people want things like Marvel movies indefinitely to the point where
- 45:35SPEAKER_00: then they're just like,
- 45:35SPEAKER_00: wait,
- 45:35SPEAKER_00: this is crap.
- 45:37SPEAKER_00: But it's because you kept narrow focusing on one little part of the creative space.
- 45:42SPEAKER_00: And so I do wonder about that.
- 45:44SPEAKER_00: I think humans, we're good at being unpredictable.
- 45:47SPEAKER_00: We'll see how good AI is in the future.
- 45:50SPEAKER_01: Yeah, I think all of that is true.
- 45:54SPEAKER_01: At the same time,
- 45:55SPEAKER_01: I think these models,
- 45:56SPEAKER_01: even with all of these sort of hacks and improvements and agent models and,
- 46:00SPEAKER_01: ah,
- 46:01SPEAKER_01: I can plug an API into this and to that,
- 46:02SPEAKER_01: you know,
- 46:03SPEAKER_01: I love the idea,
- 46:04SPEAKER_01: but if you really zoom out,
- 46:05SPEAKER_01: it sounds like somebody looking for,
- 46:07SPEAKER_01: you know,
- 46:08SPEAKER_01: a cheap code or,
- 46:08SPEAKER_01: you know,
- 46:09SPEAKER_01: kind of like it's,
- 46:10SPEAKER_01: you know,
- 46:10SPEAKER_01: Alibaba and the 40 Thieves,
- 46:12SPEAKER_01: you know,
- 46:12SPEAKER_01: you just got to say the magic word,
- 46:13SPEAKER_01: the magic incantation,
- 46:14SPEAKER_01: and it gets better,
- 46:15SPEAKER_01: right?
- 46:15SPEAKER_01: Like, the...
- 46:17SPEAKER_01: So those improvements are good,
- 46:20SPEAKER_01: and maybe we get lost in the weeds,
- 46:22SPEAKER_01: but the reality is these models still have,
- 46:23SPEAKER_01: like,
- 46:24SPEAKER_01: not just a vacuousness,
- 46:25SPEAKER_01: but a little bit of a mode collapse,
- 46:27SPEAKER_01: which is basically saying that they kind of revert to the same thing.
- 46:30SPEAKER_01: So one thing you'll definitely notice with...
- 46:34SPEAKER_01: with the song we played from Suno,
- 46:35SPEAKER_01: but really all of the earlier LLM stuff,
- 46:38SPEAKER_01: is if you give it a lot of guidance or if you give it specific lyrics,
- 46:41SPEAKER_01: it'll be pretty good.
- 46:42SPEAKER_01: But if you just say,
- 46:43SPEAKER_01: hey,
- 46:43SPEAKER_01: write a song about something,
- 46:45SPEAKER_01: it all kind of sounds really,
- 46:46SPEAKER_01: really,
- 46:47SPEAKER_01: really similar.
- 46:47SPEAKER_01: It just goes to your point, the same words, the same rhymes.
- 46:52SPEAKER_01: It has this averaging effect,
- 46:55SPEAKER_01: which is kind of inherent in how the models are sort of trained,
- 46:59SPEAKER_01: where it goes to what is sort of like lowest common denominator,
- 47:03SPEAKER_01: which I think,
- 47:04SPEAKER_01: is not really what people like the most.
- 47:09SPEAKER_01: You know, I think people like that spark of originality and consistency.
- 47:13SPEAKER_01: And I think for all of our attempts with math,
- 47:15SPEAKER_01: and I'm not saying it won't get better,
- 47:17SPEAKER_01: but it's,
- 47:19SPEAKER_01: it's been hard to do that.
- 47:20SPEAKER_01: You know,
- 47:20SPEAKER_01: like it's actually a lot,
- 47:22SPEAKER_01: it's actually very difficult to get the LMS to say something interesting.
- 47:25SPEAKER_01: Like you can get it to say nonsense or you can make it sound like something else, but
- 47:31SPEAKER_01: But,
- 47:32SPEAKER_01: you know,
- 47:32SPEAKER_01: one things that the models I think are shockingly bad at is if you just say,
- 47:37SPEAKER_01: hey,
- 47:38SPEAKER_01: like join this conversation and just like make like snazzy responses and just say
- 47:41SPEAKER_01: something witty in the context of a real human conversation.
- 47:46SPEAKER_01: or dating app or anything just like that, it's incredibly bad at it.
- 47:51SPEAKER_01: And as much as you try, it's just flat and corny.
- 47:56SPEAKER_01: And I'm not good at that either as a human, but we all know people who are amazing at that.
- 48:00SPEAKER_01: We can do that every day.
- 48:04SPEAKER_01: So why is that?
- 48:05SPEAKER_00: I don't know.
- 48:07SPEAKER_00: So stepping back into a broad macroeconomic scale, so NVIDIA is a massive company.
- 48:14SPEAKER_00: A company you worked at now.
- 48:16SPEAKER_00: NVIDIA was not a household name company-wise a couple of years ago.
- 48:21SPEAKER_00: Today it is kind of, especially if you read the Wall Street Journal business pages, stuff like that.
- 48:26SPEAKER_00: I think it is now the most valuable company.
- 48:28SPEAKER_00: And I texted a friend of mine and I told her, I was like –
- 48:34SPEAKER_00: I think we're at the peak of this AI hype cycle.
- 48:38SPEAKER_00: AI is here to stay.
- 48:39SPEAKER_00: My friends are AI scientists.
- 48:42SPEAKER_00: A lot of AI scientists that I know that work in the corporate world, they are – it's good times.
- 48:48SPEAKER_00: Good times are rolling.
- 48:48SPEAKER_00: They just want to know how long the good times will roll.
- 48:51SPEAKER_00: I think AI is going to be here to stay, obviously, just like the internet was here to stay.
- 48:54SPEAKER_00: But there is a lot of money sloshing around.
- 48:58SPEAKER_00: I was –
- 49:00SPEAKER_00: Talking to a friend of mine who's a scientist,
- 49:02SPEAKER_00: and one of his crappier postdocs somehow got a job at Meta.
- 49:08SPEAKER_00: And this guy is wanting to do a startup now, and he was asking his old postdoc how to recruit.
- 49:13SPEAKER_00: And the postdoc was laughing at him because he said, look, I'm making $600,000 a year.
- 49:19SPEAKER_00: Unless you can pay people that much, you're not going to get anyone good.
- 49:22SPEAKER_00: And this guy was kind of shocked because this was not a great – this was not one of his –
- 49:27SPEAKER_00: Great postdocs.
- 49:28SPEAKER_00: So what is going on?
- 49:31SPEAKER_00: How are you feeling?
- 49:32SPEAKER_00: You have your own company.
- 49:34SPEAKER_00: It's kind of a separate thing.
- 49:35SPEAKER_00: In a way,
- 49:37SPEAKER_00: the hype cycle crashing in some ways can be good for smaller companies because then
- 49:41SPEAKER_00: all of a sudden you compete for talent.
- 49:43SPEAKER_00: The oxygen is not being swallowed up by these big companies.
- 49:45SPEAKER_00: But talk about that a little bit.
- 49:47SPEAKER_00: How do you feel it?
- 49:48SPEAKER_01: Yeah, and I think some people that become kind of a popular, a more popular meme to talk about.
- 49:55SPEAKER_01: My friend Clem,
- 49:55SPEAKER_01: who's a founder of Hugging Face,
- 49:57SPEAKER_01: has been sort of big on this since the beginning of the year.
- 50:01SPEAKER_01: You are seeing, frankly, way too many AI companies that got started.
- 50:05SPEAKER_01: For a while, it was really easy to raise at least the first round.
- 50:09SPEAKER_01: And now I think a lot of them are, you know, the space is crowded.
- 50:13SPEAKER_01: They don't necessarily have anything original to offer.
- 50:16SPEAKER_01: No companies are really making profits.
- 50:18SPEAKER_01: I mean, some are generating revenue.
- 50:20SPEAKER_01: Some are generating a tremendous amount of revenue.
- 50:22SPEAKER_01: I mean, OpenAI went from zero to 3.5 billion.
- 50:26SPEAKER_01: AIR, I think, was the latest thing, the latest report.
- 50:30SPEAKER_01: But even Anthropic,
- 50:31SPEAKER_01: which,
- 50:31SPEAKER_01: again,
- 50:32SPEAKER_01: most people outside of Silicon Valley have never even heard of,
- 50:34SPEAKER_01: is doing close to a billion.
- 50:35SPEAKER_01: AIR, Cohere is doing hundreds of millions.
- 50:39SPEAKER_01: So these companies are growing.
- 50:41SPEAKER_01: But yeah, they're competing for talent.
- 50:43SPEAKER_01: They're overall losing money, as is always natural in the early cycle.
- 50:47SPEAKER_01: And the salaries, to your point, are very high.
- 50:49SPEAKER_01: I mean,
- 50:49SPEAKER_01: I'm obviously not going to name names,
- 50:51SPEAKER_01: but I know very junior people at OpenAI whose total pay package is in the low seven
- 50:57SPEAKER_01: figures now.
- 50:58SPEAKER_01: And even they would admit that that makes no sense.
- 51:02SPEAKER_01: It's just kind of absurd.
- 51:03SPEAKER_01: So I think even if things keep growing, I mean, I think there will be some sort of crash.
- 51:09SPEAKER_01: I mean, there has to be just inevitable.
- 51:11SPEAKER_01: It's always the case.
- 51:13SPEAKER_01: And,
- 51:14SPEAKER_01: you know,
- 51:14SPEAKER_01: a lot of companies are just going to find that they can't,
- 51:16SPEAKER_01: to your point,
- 51:17SPEAKER_01: operate or stay in business.
- 51:19SPEAKER_01: There'll be a lot of consolidation.
- 51:21SPEAKER_01: Salaries at some point will crash, especially when people realize that, you know.
- 51:25SPEAKER_01: Like, the reason they're paying high salaries is because of the rush, right?
- 51:29SPEAKER_01: You're not paying high salaries because people are differentiated.
- 51:32SPEAKER_01: It's because you need something now.
- 51:33SPEAKER_01: So, for example, if you are, you know, I mean, these guys are my friends, you know, just disclosure.
- 51:41SPEAKER_01: But, you know, you look at XAI, right?
- 51:43SPEAKER_01: The Grok model associated, but a separate company from X, i.e.
- 51:49SPEAKER_01: Twitter.
- 51:50SPEAKER_01: You know, they raised $6 billion.
- 51:52SPEAKER_01: You have Elon publicly saying they're going to have a model better than ChatGPT at some point soon.
- 51:58SPEAKER_01: So there's a rush, right?
- 52:00SPEAKER_01: You need to make this happen today.
- 52:01SPEAKER_01: So the reason to spend the money, to hire the people.
- 52:04SPEAKER_01: Also,
- 52:04SPEAKER_01: if you're spending tens of millions,
- 52:07SPEAKER_01: hundreds of millions of dollars on modeling,
- 52:10SPEAKER_01: it does kind of make sense to not underpay your staff.
- 52:13SPEAKER_01: So I think you've got a lot of positive feedback cycles,
- 52:16SPEAKER_01: which are,
- 52:16SPEAKER_01: of course,
- 52:17SPEAKER_01: great for Silicon Valley and for tech and for good engineers.
- 52:20SPEAKER_01: I mean...
- 52:22SPEAKER_01: A lot of these people are not even AI people at all.
- 52:25SPEAKER_01: They're just good systems engineers, which is kind of what you need a lot to build these complex systems.
- 52:30SPEAKER_01: I mean, you mentioned NVIDIA before.
- 52:32SPEAKER_01: I mean,
- 52:33SPEAKER_01: my team and really me and my intern were the first people to train LLMs,
- 52:38SPEAKER_01: uh,
- 52:38SPEAKER_01: deep learning as the,
- 52:39SPEAKER_01: the transformer models at Nvidia,
- 52:42SPEAKER_01: you know,
- 52:42SPEAKER_01: and now that team is,
- 52:43SPEAKER_01: you know,
- 52:44SPEAKER_01: a hundred people,
- 52:45SPEAKER_01: right.
- 52:46SPEAKER_01: I read their technical report from earlier this week.
- 52:48SPEAKER_01: It was pretty awesome.
- 52:49SPEAKER_01: Basically like the first actual open source version of chat GPT explaining all the
- 52:53SPEAKER_01: tricks and things like that.
- 52:54SPEAKER_01: I mean, there's a hundred people on the project, give or take, you know, um,
- 52:59SPEAKER_01: So these things have grown and it makes sense, right?
- 53:02SPEAKER_01: But yeah, I think it's interesting to look at what these AI LLM companies are really doing.
- 53:09SPEAKER_01: I would say there's... You can sort of put them into really just a few categories.
- 53:14SPEAKER_01: So...
- 53:16SPEAKER_01: I'll say the first one,
- 53:17SPEAKER_01: the so-called GPT wrappers that get a very hard time,
- 53:20SPEAKER_01: especially from smart,
- 53:22SPEAKER_01: high IQ scientist people.
- 53:24SPEAKER_01: I think those companies actually make a tremendous amount of sense because GPT,
- 53:29SPEAKER_01: to your point,
- 53:30SPEAKER_01: people don't know how to use an API.
- 53:33SPEAKER_01: So imagine taking the GPT API and actually turning it into like a useful mini product,
- 53:38SPEAKER_01: you know,
- 53:39SPEAKER_01: like receive GPT or something like that.
- 53:42SPEAKER_01: I think that's cool.
- 53:43SPEAKER_01: People like that.
- 53:44SPEAKER_00: There is a receive GPT.
- 53:46SPEAKER_01: Exactly.
- 53:46SPEAKER_01: Exactly.
- 53:47SPEAKER_01: So, you know, you do need that connective layer to, I think, take these things and actually make them,
- 53:53SPEAKER_01: you know, turn them into like a tool, especially, you know, it's a UX thing fundamentally.
- 53:57SPEAKER_01: Right.
- 53:58SPEAKER_01: Um,
- 53:59SPEAKER_01: so I think those are,
- 54:00SPEAKER_01: you know,
- 54:01SPEAKER_01: those are,
- 54:02SPEAKER_01: none of them are going to be huge companies,
- 54:03SPEAKER_01: but that's sort of a,
- 54:03SPEAKER_01: you know,
- 54:04SPEAKER_01: kind of a,
- 54:04SPEAKER_01: just a,
- 54:05SPEAKER_01: um,
- 54:06SPEAKER_01: a good cottage industry,
- 54:06SPEAKER_01: right?
- 54:07SPEAKER_01: You have the,
- 54:08SPEAKER_01: you had some companies for a while that were doing pretty good business,
- 54:11SPEAKER_01: fine tuning these models.
- 54:13SPEAKER_01: So basically taking an existing large,
- 54:15SPEAKER_01: large model,
- 54:16SPEAKER_01: um,
- 54:17SPEAKER_01: usually from stability or mixed trial,
- 54:19SPEAKER_01: because those are open source,
- 54:20SPEAKER_01: um,
- 54:20SPEAKER_01: but occasionally even some of the open AI type models and basically actually
- 54:24SPEAKER_01: editing the model a little bit.
- 54:27SPEAKER_01: I would say those companies,
- 54:28SPEAKER_01: if not dead,
- 54:29SPEAKER_01: are all dying right now because it turns out the latest models are just like better optimized.
- 54:34SPEAKER_01: So there's less meat on the bone.
- 54:36SPEAKER_01: Um, those companies are not, no new ones are being funded and the current ones are probably in trouble.
- 54:41SPEAKER_01: Um,
- 54:42SPEAKER_01: you do,
- 54:43SPEAKER_01: you do have the companies building original new models,
- 54:47SPEAKER_01: but like I said,
- 54:47SPEAKER_01: I mean,
- 54:47SPEAKER_01: it really feels like right now there's probably three to six to maybe 12
- 54:52SPEAKER_01: organizations max that actually like can bear the,
- 54:55SPEAKER_01: can bear the expense of training large models from scratch,
- 54:58SPEAKER_01: which means,
- 54:59SPEAKER_01: you know,
- 54:59SPEAKER_01: getting the data,
- 55:00SPEAKER_01: piping it in.
- 55:02SPEAKER_01: You know,
- 55:02SPEAKER_01: either having your own cloud or spending,
- 55:04SPEAKER_01: you know,
- 55:05SPEAKER_01: basically spending hundreds of millions of dollars at the very least to rent enough
- 55:10SPEAKER_01: capacity to train it.
- 55:11SPEAKER_01: You know, that's why you see companies like Oracle and CoreWeave doing really well out of this.
- 55:16SPEAKER_01: They're like, hey, we bought a bunch of GPUs that we don't use.
- 55:20SPEAKER_01: And, you know, they're kind of printing money and renting them to people who need them.
- 55:25SPEAKER_01: You know,
- 55:26SPEAKER_01: even Microsoft has been buying and renting spare capacity from CoreWeave and Oracle,
- 55:31SPEAKER_01: which is kind of amazing.
- 55:34SPEAKER_01: By the way,
- 55:34SPEAKER_01: that's why,
- 55:35SPEAKER_01: I mean,
- 55:35SPEAKER_01: obviously,
- 55:36SPEAKER_01: you know,
- 55:37SPEAKER_01: I'm an NVIDIA stockholder,
- 55:38SPEAKER_01: but it's not entirely obvious to me that it's peaked yet because,
- 55:42SPEAKER_01: yeah.
- 55:43SPEAKER_01: it's at a ridiculous multiple by you literally have people making billions of
- 55:47SPEAKER_01: dollars of,
- 55:49SPEAKER_01: of,
- 55:49SPEAKER_01: of renting their spare capacity on G on Nvidia GPUs.
- 55:52SPEAKER_01: Like the market is, is, is, is that sort of supply constrained?
- 55:57SPEAKER_01: Yeah.
- 55:59SPEAKER_01: But, but yeah, but, but, but, but I think that's it.
- 56:01SPEAKER_01: Like, I think you, you, I think that the, you know, the,
- 56:06SPEAKER_01: I think ultimately the companies that are going to thrive and survive is a very
- 56:11SPEAKER_01: small number of companies that can actually train and differentiate their model
- 56:15SPEAKER_01: where it actually makes sense.
- 56:16SPEAKER_01: For most of them, it doesn't.
- 56:19SPEAKER_01: And then everything else is going to be applications,
- 56:20SPEAKER_01: either large applications or very small niche ones like the received GPT kind of stuff.
- 56:28SPEAKER_01: So I don't know.
- 56:29SPEAKER_01: I mean, look, again, it's weird to me, again, not not living in Silicon Valley anymore.
- 56:33SPEAKER_01: I mean, you go there and you just go to like a random fills and a random part of town.
- 56:37SPEAKER_01: And it feels like half of the half of the coffee shop is doing something with, you know, with AI models.
- 56:43SPEAKER_01: Right.
- 56:43SPEAKER_01: I mean, I don't know if that feels the same way to you.
- 56:45SPEAKER_01: Probably.
- 56:46SPEAKER_01: Right.
- 56:47SPEAKER_00: Yeah.
- 56:47SPEAKER_00: Yeah.
- 56:48SPEAKER_00: Yeah.
- 56:48SPEAKER_00: No, it was.
- 56:49SPEAKER_00: I was at SF actually recently, actually several times.
- 56:53SPEAKER_00: and it does definitely like, uh, feel that, feel that way.
- 56:57SPEAKER_00: And, uh, you know, it's, it's exciting.
- 57:00SPEAKER_00: It's exciting.
- 57:00SPEAKER_00: Although, you know, people are always, uh, worried about, worried about that.
- 57:04SPEAKER_00: So, you know, there is the, Ooh, AGI will destroy civilization.
- 57:08SPEAKER_00: I think like that, my P doom is dropping.
- 57:11SPEAKER_00: That's what I'm, that's what I'm saying.
- 57:12SPEAKER_00: It's not, it wasn't like, that's really super high, but I got a low P doom.
- 57:15SPEAKER_00: Uh, but you know, the economic, the macroeconomic stuff is, uh, that's also relevant.
- 57:20SPEAKER_00: And, uh, you know,
- 57:22SPEAKER_00: I think people are a little worried about the hangover after.
- 57:26SPEAKER_00: Now,
- 57:27SPEAKER_00: I guess one thing I want to talk to you about before we pivot back into speaking of
- 57:31SPEAKER_00: applications and the deep news,
- 57:32SPEAKER_00: what you're doing there,
- 57:34SPEAKER_00: is these are really big companies.
- 57:40SPEAKER_00: What about just nationalization?
- 57:42SPEAKER_00: I mean, what about China just being like, okay, you know what?
- 57:46SPEAKER_00: We're just going to have a national championship, and we're basically going to buy stock of this company.
- 57:50SPEAKER_00: Yeah.
- 57:51SPEAKER_00: It's going to be like the government of China's AI.
- 57:54SPEAKER_00: What about governments basically getting into the competition?
- 57:59SPEAKER_01: Yeah, no, that's a great point.
- 58:01SPEAKER_01: And I mean, to a large extent, that is happening in China.
- 58:07SPEAKER_01: And that is sort of a little bit how they do business, right?
- 58:10SPEAKER_01: So for people who've read like the chip war book,
- 58:13SPEAKER_01: which I think is really the best thing you can read about the superconducting industry,
- 58:17SPEAKER_01: and especially the later parts,
- 58:18SPEAKER_01: how it was handled differently in different countries,
- 58:20SPEAKER_01: you know,
- 58:20SPEAKER_01: in Japan and Taiwan and China.
- 58:24SPEAKER_01: I mean, you're seeing this now, like you're so...
- 58:28SPEAKER_01: After this,
- 58:29SPEAKER_01: we should probably get into LMS and why it's so interesting,
- 58:31SPEAKER_01: because I think it's fascinating,
- 58:33SPEAKER_01: actually.
- 58:33SPEAKER_01: One of the cool things that come out of it.
- 58:34SPEAKER_01: But,
- 58:34SPEAKER_01: I mean,
- 58:35SPEAKER_01: you do see,
- 58:36SPEAKER_01: like I said,
- 58:36SPEAKER_01: in the top 10,
- 58:38SPEAKER_01: you know,
- 58:38SPEAKER_01: of the best,
- 58:39SPEAKER_01: you know,
- 58:40SPEAKER_01: according to,
- 58:41SPEAKER_01: you know,
- 58:42SPEAKER_01: many,
- 58:42SPEAKER_01: many,
- 58:42SPEAKER_01: many,
- 58:42SPEAKER_01: many people voting on this stuff in a blind test.
- 58:46SPEAKER_01: Yeah,
- 58:46SPEAKER_01: I mean,
- 58:46SPEAKER_01: you do see a Chinese company,
- 58:48SPEAKER_01: the O1 AI,
- 58:49SPEAKER_01: which was founded by Kai-Fu Lee,
- 58:52SPEAKER_01: who,
- 58:52SPEAKER_01: you know,
- 58:52SPEAKER_01: old school tech people will remember from Microsoft and Google.
- 58:56SPEAKER_01: But they have a model that in English is incredibly competitive.
- 59:00SPEAKER_01: And that is very much a... I mean, it's a company, but it is a Chinese national champion.
- 59:07SPEAKER_01: And it is very good.
- 59:08SPEAKER_01: Interestingly enough, for reasons that I don't quite understand, I mean, I can sort of guess, but...
- 59:13SPEAKER_01: Weirdly, Japan has like completely missed the LLM wave for some reason.
- 59:20SPEAKER_01: But China has been incredibly involved and their models are very good.
- 59:23SPEAKER_01: I mean,
- 59:23SPEAKER_01: even before,
- 59:25SPEAKER_01: I mean,
- 59:25SPEAKER_01: I remember using some of the open source Chinese models,
- 59:28SPEAKER_01: especially in between when OpenAI had this funny period where they had some things
- 59:33SPEAKER_01: that are really out of date.
- 59:34SPEAKER_01: And I mean, their models are really not bad at all.
- 59:37SPEAKER_01: Like they're very comparable to be honest.
- 59:39SPEAKER_01: So I think that, you know, the Chinese government is definitely a strong player there.
- 59:44SPEAKER_01: They can absolutely build a giant cluster and take all the data.
- 59:48SPEAKER_01: I mean,
- 59:48SPEAKER_01: you can even argue that they could do things like practical things,
- 59:52SPEAKER_01: like,
- 59:52SPEAKER_01: for example,
- 59:53SPEAKER_01: you know,
- 59:53SPEAKER_01: take things from the Internet Archive and not worry too much about whether the New
- 59:56SPEAKER_01: York Times is going to sue them because maybe they don't care.
- 59:58SPEAKER_01: You know,
- 1:00:01SPEAKER_01: certainly an argument to be made that that that they're not just copying,
- 1:00:04SPEAKER_01: that they're actually doing things well.
- 1:00:06SPEAKER_01: So it's it's interesting.
- 1:00:08SPEAKER_00: So you wanted to talk more about limbsis?
- 1:00:11SPEAKER_01: Yeah, so I think...
- 1:00:13SPEAKER_01: Yeah.
- 1:00:14SPEAKER_01: While the big things that people sort of promised and hoped or feared mostly
- 1:00:18SPEAKER_01: haven't really happened,
- 1:00:19SPEAKER_01: I think around the edges,
- 1:00:20SPEAKER_01: we built kind of an interesting ecosystem.
- 1:00:21SPEAKER_01: And one of the coolest things to emerge is people always want to compare models, right?
- 1:00:27SPEAKER_01: Which is model A better than model B?
- 1:00:29SPEAKER_01: And we have all these tests and people releasing data sets on very hard questions and things like that.
- 1:00:34SPEAKER_01: But very much the gold standard has become LMSIS, which is
- 1:00:38SPEAKER_01: You have tens of thousands of volunteers that go and fundamentally just look side by side.
- 1:00:43SPEAKER_01: They type in anything they want.
- 1:00:45SPEAKER_01: These questions get cleaned up and classified later.
- 1:00:48SPEAKER_01: They're very restrictive.
- 1:00:50SPEAKER_01: If you go there and you ask anything that asks about the demographics,
- 1:00:53SPEAKER_01: they will not answer your question,
- 1:00:54SPEAKER_01: things like that.
- 1:00:55SPEAKER_01: But fundamentally, it's just an API, and it sends it to two models, and you vote which one is better.
- 1:01:01SPEAKER_01: And that has become the gold standard.
- 1:01:03SPEAKER_01: You know,
- 1:01:03SPEAKER_01: it uses ELO ratings,
- 1:01:04SPEAKER_01: which chess nerds were like,
- 1:01:06SPEAKER_01: you know,
- 1:01:06SPEAKER_01: to rank these models separately in coding and things like that,
- 1:01:10SPEAKER_01: but also overall.
- 1:01:11SPEAKER_01: And I think that's pretty interesting that that has emerged as like the gold
- 1:01:17SPEAKER_01: standard for comparing models.
- 1:01:18SPEAKER_01: I think that's pretty cool.
- 1:01:20SPEAKER_00: Yeah.
- 1:01:22SPEAKER_00: I mean, yeah.
- 1:01:24SPEAKER_00: So you said like,
- 1:01:25SPEAKER_00: but like,
- 1:01:25SPEAKER_00: it's like basically the top six or so,
- 1:01:29SPEAKER_00: like how far down do we go before we just don't pay attention?
- 1:01:34SPEAKER_00: Like, how's Meta AI, how's Meta's, their Lama, how's Lama doing?
- 1:01:39SPEAKER_01: Yeah, yeah.
- 1:01:40SPEAKER_01: So, it's a good question.
- 1:01:42SPEAKER_01: So, if you look at the ELO ratings, yeah.
- 1:01:44SPEAKER_01: So, number one is GPT-4.0.
- 1:01:47SPEAKER_01: Then there's three Gemini models that are basically all the same,
- 1:01:50SPEAKER_01: according to Elemsys,
- 1:01:51SPEAKER_01: that are tied for sort of second...
- 1:01:55SPEAKER_01: through fourth um and yeah if you if you look all the way down to yeah so llama uh
- 1:02:02SPEAKER_01: llama 3 70 billion which is an excellent model by the way um it's a 1207 so it's 80
- 1:02:08SPEAKER_01: elo points lower so that's like a substantial difference that you would notice but
- 1:02:13SPEAKER_01: it's it's it's in a big tie for 11th place along with bard the sort of the older
- 1:02:17SPEAKER_01: gemini model and also the nvidia open source model um
- 1:02:23SPEAKER_01: Claw 3 Sonnet, which is sort of the worst model of Claw 3.
- 1:02:27SPEAKER_01: So it's still, I mean, the way I would describe, it's like even less than 100 point ELO difference.
- 1:02:33SPEAKER_01: The way I would describe it is that on a bunch of tasks that have sort of moved
- 1:02:38SPEAKER_01: into the relatively easy mode is going to do exactly great.
- 1:02:43SPEAKER_01: So for example,
- 1:02:43SPEAKER_01: all of these models,
- 1:02:45SPEAKER_01: if you ask it a question like,
- 1:02:46SPEAKER_01: here's a news headline,
- 1:02:47SPEAKER_01: is this positive or negative?
- 1:02:49SPEAKER_01: Is this bullish or bearish for a stock?
- 1:02:50SPEAKER_01: Something like that.
- 1:02:53SPEAKER_01: not not only is something like gpt3 good at it you're actually not going to notice
- 1:02:58SPEAKER_01: a statistical difference between that and gpt4 so a lot of the stuff we actually
- 1:03:03SPEAKER_01: use these models for in practice these you know those type models are perfectly
- 1:03:08SPEAKER_01: fine like you're not even going to notice the difference um you know but obviously
- 1:03:12SPEAKER_01: like the more difficult tasks you are going to notice a difference um but you're
- 1:03:17SPEAKER_01: not going to tell 100 of the time if that makes sense
- 1:03:21SPEAKER_00: Well, I mean, so let's talk about Deep News a little bit.
- 1:03:24SPEAKER_00: So, you know, you've been working on this.
- 1:03:27SPEAKER_00: How many years have you been working on Deep News?
- 1:03:29SPEAKER_00: A year or more than a year?
- 1:03:30SPEAKER_01: I'd say a little bit less.
- 1:03:32SPEAKER_01: I mean,
- 1:03:33SPEAKER_01: I,
- 1:03:34SPEAKER_01: you know,
- 1:03:35SPEAKER_01: with our company,
- 1:03:36SPEAKER_01: you know,
- 1:03:36SPEAKER_01: we were doing some other stuff and kind of like knew we had the team to do this.
- 1:03:39SPEAKER_01: Like, I really, really wanted to do this a year ago.
- 1:03:42SPEAKER_01: And then we kind of got into it over the summer.
- 1:03:44SPEAKER_01: And, you know, the first...
- 1:03:46SPEAKER_01: very hacky version we probably released in in august or september so it's been it's
- 1:03:50SPEAKER_01: been a bit less than a year um you know but but but actually the concept has kind
- 1:03:55SPEAKER_01: of remained the same is you take tens of thousands of sources most of it from
- 1:04:01SPEAKER_01: twitter iex and you know you have like a at a high level relatively straightforward
- 1:04:07SPEAKER_01: process
- 1:04:08SPEAKER_01: You look at all the recent information.
- 1:04:10SPEAKER_01: You try to generate story candidates.
- 1:04:12SPEAKER_01: Here's a cluster of tweets or articles or whatever talking about a particular thing.
- 1:04:18SPEAKER_01: And then you try to see, okay, you're doing initial version.
- 1:04:23SPEAKER_01: Is this newsworthy or not?
- 1:04:24SPEAKER_01: Is it interesting?
- 1:04:26SPEAKER_01: And if it is, you try harder to write the best version of it and then you rank it.
- 1:04:31SPEAKER_01: So kind of as an example,
- 1:04:33SPEAKER_01: I would say sort of an old example by,
- 1:04:35SPEAKER_01: you know,
- 1:04:36SPEAKER_01: say a bunch of VCs go to vacation together in Cabo and talk about it.
- 1:04:42SPEAKER_01: It's a thing.
- 1:04:43SPEAKER_01: It's probably not newsworthy.
- 1:04:44SPEAKER_01: Nobody really cares.
- 1:04:45SPEAKER_01: But a bunch of VCs go to Burning Man and get stuck in the mud.
- 1:04:49SPEAKER_01: That is interesting.
- 1:04:50SPEAKER_01: So like it's a lot of it is things like that, like common sense.
- 1:04:55SPEAKER_01: So you're not you're not declaring what is or isn't news.
- 1:04:57SPEAKER_01: You're really just trying to,
- 1:04:58SPEAKER_01: you know,
- 1:04:59SPEAKER_01: you can train even frankly,
- 1:05:01SPEAKER_01: relatively small models to figure out,
- 1:05:02SPEAKER_01: hey,
- 1:05:03SPEAKER_01: is this sort of like is this cluster of information interesting?
- 1:05:06SPEAKER_01: And if it is, then you try more to figure out like, what are the facts?
- 1:05:10SPEAKER_01: What are the interesting pieces?
- 1:05:11SPEAKER_01: Do we pull in previous information?
- 1:05:13SPEAKER_01: That sort of thing.
- 1:05:13SPEAKER_01: So it is fundamentally kind of a little bit of a,
- 1:05:16SPEAKER_01: of a prompt chain agent work,
- 1:05:18SPEAKER_01: however you want to describe it.
- 1:05:19SPEAKER_01: Um, it's not as simple as take these tweets and summarize it.
- 1:05:22SPEAKER_01: That would not work as well.
- 1:05:23SPEAKER_01: But, um,
- 1:05:25SPEAKER_01: You know, and that's basically the idea.
- 1:05:28SPEAKER_01: And then, you know, the hard part gets into, okay, what things should you show first?
- 1:05:32SPEAKER_01: What is actually worthy of a notification?
- 1:05:33SPEAKER_01: What topics is it?
- 1:05:34SPEAKER_01: Hey, like Razeeb, you're very interested in biology.
- 1:05:37SPEAKER_01: You're not at all interested in sports.
- 1:05:39SPEAKER_01: Maybe you even hate sports.
- 1:05:40SPEAKER_01: You know,
- 1:05:41SPEAKER_01: how do we set up a view where like you're not ever going to see sports unless maybe
- 1:05:46SPEAKER_01: it like crosses over into,
- 1:05:47SPEAKER_01: you know,
- 1:05:48SPEAKER_01: like,
- 1:05:48SPEAKER_01: you know,
- 1:05:49SPEAKER_01: if there's an earthquake of the world series,
- 1:05:50SPEAKER_01: maybe you're interested.
- 1:05:51SPEAKER_01: I don't know.
- 1:05:52SPEAKER_01: to use a sort of a bad example.
- 1:05:53SPEAKER_00: Yeah.
- 1:05:54SPEAKER_00: Well, so, you know, Google news has been around for a long time.
- 1:05:58SPEAKER_00: Uh, what, I mean, I don't actually know how does Google, I, I'm not gonna lie.
- 1:06:02SPEAKER_00: I don't use Google news too much anymore.
- 1:06:04SPEAKER_00: I used to use it more, but I feel like it's a little too generic if that ever, if that makes sense.
- 1:06:09SPEAKER_00: But yeah, so, so Google news is, I, yeah, I mean, it's a standalone product.
- 1:06:14SPEAKER_01: I know almost nobody who uses, obviously we all get exposed to it through Google search.
- 1:06:18SPEAKER_01: Um,
- 1:06:19SPEAKER_01: The fundamental difference between what we're doing and something like Google News
- 1:06:23SPEAKER_01: is that we're writing our own stories,
- 1:06:26SPEAKER_01: right?
- 1:06:27SPEAKER_01: Google News at core says, hey, I'm going to show you the right news article.
- 1:06:33SPEAKER_01: And then it does some aggregation step.
- 1:06:35SPEAKER_01: It says, okay, like these seven articles are related, right?
- 1:06:38SPEAKER_01: And these are just fundamentally different things.
- 1:06:41SPEAKER_01: I think being recommended articles to read is perfectly fine.
- 1:06:45SPEAKER_01: Apple News does it very well.
- 1:06:48SPEAKER_01: People were super into this company called Artifact that then sort of got shut down.
- 1:06:52SPEAKER_01: Now the UI and IP got bought by Yahoo while the founder went to Anthropic, but whatever.
- 1:07:00SPEAKER_01: I think article aggregation is perfectly reasonable, but...
- 1:07:04SPEAKER_01: Actually,
- 1:07:05SPEAKER_01: I think what we're doing is fundamentally different in that we just say what's
- 1:07:09SPEAKER_01: going on,
- 1:07:10SPEAKER_01: you know,
- 1:07:10SPEAKER_01: which Google News never really tried to do and still doesn't.
- 1:07:14SPEAKER_01: And I think it's actually quite hard.
- 1:07:16SPEAKER_01: You know, like, for example, I've tested many versions of Gemini.
- 1:07:19SPEAKER_01: I'm very interested, of course.
- 1:07:20SPEAKER_01: And my standard question was, what's going on in the UFC?
- 1:07:23SPEAKER_01: Just tell me.
- 1:07:24SPEAKER_01: I'm a big fan of martial arts, jiu-jitsu, whatever.
- 1:07:28SPEAKER_01: What's happening in the UFC?
- 1:07:30SPEAKER_01: And somebody who's informed, who's knowledgeable can answer that question.
- 1:07:32SPEAKER_01: Like, well, this fight is coming up.
- 1:07:34SPEAKER_01: People are excited about this.
- 1:07:36SPEAKER_01: This fight just happened.
- 1:07:37SPEAKER_01: This guy got injured.
- 1:07:38SPEAKER_01: This guy got banned.
- 1:07:39SPEAKER_01: This guy talked shit about this other guy.
- 1:07:42SPEAKER_01: And a really informed fan or whomever could tell you this stuff.
- 1:07:47SPEAKER_01: But Gemini has no ability to do this.
- 1:07:50SPEAKER_01: Like it'll tell you facts, but it's, but they're not, they're not organized in an interesting way.
- 1:07:57SPEAKER_01: There's no priority.
- 1:07:58SPEAKER_01: You know, it doesn't have a concept of this fight's important.
- 1:08:01SPEAKER_01: That fight's not important because you know, there's 30 fights every week.
- 1:08:06SPEAKER_01: And actually,
- 1:08:07SPEAKER_01: my favorite version of Gemini,
- 1:08:08SPEAKER_01: which I guess was called Bard then,
- 1:08:10SPEAKER_01: there was this quick one-week period where they focused on making it interesting at
- 1:08:15SPEAKER_01: the expense of accuracy.
- 1:08:16SPEAKER_01: So for a while, you could ask Gemini, hey, what's going on in the UFC?
- 1:08:19SPEAKER_01: And it would invent these interesting fights that weren't real.
- 1:08:21SPEAKER_01: And that was actually fun.
- 1:08:24SPEAKER_01: That was actually exciting.
- 1:08:26SPEAKER_01: And this is what I think is something that is cool about what we're doing is we're
- 1:08:30SPEAKER_01: actually seeing,
- 1:08:31SPEAKER_01: hey,
- 1:08:32SPEAKER_01: here's tens of thousands of important people and official accounts,
- 1:08:37SPEAKER_01: everything like this.
- 1:08:39SPEAKER_01: And we've actually been able to train on what do people care about.
- 1:08:42SPEAKER_01: Kind of democratizing it out.
- 1:08:44SPEAKER_01: Yeah.
- 1:08:45SPEAKER_01: So like I said, I mentioned the Supreme Court stuff before.
- 1:08:48SPEAKER_01: There's cases all the time.
- 1:08:50SPEAKER_01: But which cases are actually interesting to people and why?
- 1:08:52SPEAKER_01: I think you're essentially this is sort of a little bit mostly Twitter,
- 1:08:58SPEAKER_01: but you kind of think of the hive mind,
- 1:09:00SPEAKER_01: right?
- 1:09:00SPEAKER_01: Is what are people talking about?
- 1:09:02SPEAKER_01: But not just random people.
- 1:09:03SPEAKER_01: Obviously, you give more weight to people who have consistently written good things in the past.
- 1:09:10SPEAKER_01: So you have a pretty good idea what is going on in biology, what is going on in genomics.
- 1:09:18SPEAKER_01: And some of that's your own opinion,
- 1:09:19SPEAKER_01: but a lot of it is you just have a notion of what is interesting to people.
- 1:09:23SPEAKER_01: And I think the AI is now good enough to do that, which is amazing.
- 1:09:27SPEAKER_00: Yeah, yeah, that is amazing.
- 1:09:30SPEAKER_00: You know, and that's utility.
- 1:09:33SPEAKER_00: You know,
- 1:09:34SPEAKER_00: it's...
- 1:09:37SPEAKER_00: You know,
- 1:09:37SPEAKER_00: there's this extreme...
- 1:09:41SPEAKER_00: So it's like,
- 1:09:42SPEAKER_00: I talk to you about AI like every six months to a year,
- 1:09:45SPEAKER_00: and,
- 1:09:46SPEAKER_00: you know,
- 1:09:46SPEAKER_00: we have these same discussions,
- 1:09:47SPEAKER_00: or I bring up the issues,
- 1:09:49SPEAKER_00: like,
- 1:09:49SPEAKER_00: are we going to be able to have this discussion in a year,
- 1:09:51SPEAKER_00: or will the world be swallowed up and,
- 1:09:53SPEAKER_00: like,
- 1:09:54SPEAKER_00: turned into paperclips and stupid stuff like that?
- 1:09:56SPEAKER_00: Ha ha!
- 1:09:58SPEAKER_00: But then the reality is there are things that AI can do for us that can help us out
- 1:10:04SPEAKER_00: in terms of designing software or organizing a party,
- 1:10:09SPEAKER_00: finding the news you want,
- 1:10:12SPEAKER_00: surfacing.
- 1:10:12SPEAKER_00: I mean, what I would love –
- 1:10:16SPEAKER_00: What I would love – and I think people already have forms as AI personal assistant,
- 1:10:21SPEAKER_00: can look at your emails and everything like that,
- 1:10:23SPEAKER_00: organize the calendar in a way for you.
- 1:10:27SPEAKER_00: There's a lot of low-level task rabbit category things.
- 1:10:35SPEAKER_00: where you just need someone to do it.
- 1:10:37SPEAKER_00: But it's not really a high-value task,
- 1:10:41SPEAKER_00: but you push all of these low-value tasks that are just busy work onto the AI,
- 1:10:48SPEAKER_00: and that's great,
- 1:10:48SPEAKER_00: in my opinion.
- 1:10:50SPEAKER_00: That increases your productivity.
- 1:10:52SPEAKER_00: It allows you to focus on things that you want to focus on.
- 1:10:56SPEAKER_00: The main thing I think that I would worry about
- 1:10:59SPEAKER_00: is over-optimization, right?
- 1:11:00SPEAKER_00: Yes, yes.
- 1:11:01SPEAKER_00: Which we've talked about with deep news and other things, you know?
- 1:11:05SPEAKER_00: Where it's like,
- 1:11:05SPEAKER_00: sometimes there are stories I should know and I would find interesting,
- 1:11:10SPEAKER_00: but I don't necessarily know it until I see it.
- 1:11:13SPEAKER_00: And how do we figure that out?
- 1:11:14SPEAKER_01: And that's my favorite part, exactly.
- 1:11:16SPEAKER_01: Like, there's... I mean, no...
- 1:11:20SPEAKER_01: No offense to the people doing the optimization,
- 1:11:22SPEAKER_01: like organizing my email and all that and all that shit.
- 1:11:25SPEAKER_01: But like, I think that's fine.
- 1:11:27SPEAKER_01: But to be honest, I think there's too many people thinking about it.
- 1:11:29SPEAKER_01: I find that on some level to be quite boring.
- 1:11:32SPEAKER_01: You know, I think I think building a new experience with AI is a lot more interesting.
- 1:11:37SPEAKER_01: You know, you know, it's it's.
- 1:11:41SPEAKER_01: And I think being informed, like some people will say, why would you want to know?
- 1:11:45SPEAKER_01: But the truth is, people do want to know.
- 1:11:46SPEAKER_01: And obviously, some people are more interested in the news than others.
- 1:11:48SPEAKER_01: But like,
- 1:11:48SPEAKER_01: for example,
- 1:11:50SPEAKER_01: you know,
- 1:11:50SPEAKER_01: a story that I saw right before we were speaking was,
- 1:11:53SPEAKER_01: hey,
- 1:11:53SPEAKER_01: you know,
- 1:11:54SPEAKER_01: Afghanistan beat Australia in the T20 cricket.
- 1:11:58SPEAKER_01: And that's something that is probably extremely meaningless to the majority of your audience,
- 1:12:03SPEAKER_01: but it's actually very interesting to a small percentage of your audience because
- 1:12:06SPEAKER_01: it's actually a big deal.
- 1:12:07SPEAKER_01: Afghanistan didn't even really play cricket until some years ago,
- 1:12:10SPEAKER_01: and now they're beating Australia,
- 1:12:12SPEAKER_01: which is a legendary team.
- 1:12:13SPEAKER_01: And there's tournaments going on.
- 1:12:15SPEAKER_01: I would like to watch it, but it's not really on American TV.
- 1:12:18SPEAKER_01: We can get into the utility of it, but the truth is people are interested in these things.
- 1:12:22SPEAKER_01: They want to know.
- 1:12:22SPEAKER_01: And I think it's cool that the AI can route this information to me.
- 1:12:25SPEAKER_01: It's succinct.
- 1:12:26SPEAKER_01: It's interesting.
- 1:12:28SPEAKER_01: And, you know, in theory, this can be done without LLMs, but I've never seen anyone do this.
- 1:12:32SPEAKER_01: I've never seen an app that actually can figure out my interests and tell me,
- 1:12:36SPEAKER_01: or even if I declare my interest and tell me in real time all the things that I
- 1:12:39SPEAKER_01: want to know.
- 1:12:41SPEAKER_01: That's Newsy.
- 1:12:41SPEAKER_01: And obviously, most of the stuff we want to know is not related to news.
- 1:12:44SPEAKER_01: Like, I love your Substack.
- 1:12:46SPEAKER_01: Your Substack has no news in it at all.
- 1:12:48SPEAKER_01: It's talking about, you know, things from tons of thousands of years ago, and that's also interesting.
- 1:12:53SPEAKER_01: Yeah.
- 1:12:54SPEAKER_01: But I think there's an element of...
- 1:12:56SPEAKER_01: That everybody really does want to know new information and sort of like the hot
- 1:13:00SPEAKER_01: thing that people are talking about.
- 1:13:02SPEAKER_01: And the new thing that just happened is just a very human thing.
- 1:13:06SPEAKER_01: Some people are more into it than others.
- 1:13:08SPEAKER_01: But I just think it's a really nice vector where the AI is,
- 1:13:11SPEAKER_01: frankly,
- 1:13:12SPEAKER_01: makes this possible in a way that was not possible before.
- 1:13:14SPEAKER_01: And I think it's amazing.
- 1:13:15SPEAKER_01: And there'll be other things like that.
- 1:13:17SPEAKER_01: And frankly, not to be too biased, but I wish there was more startups trying to do things like that.
- 1:13:22SPEAKER_01: I think there are just in my view,
- 1:13:24SPEAKER_01: too many startups that are either a,
- 1:13:26SPEAKER_01: Hey,
- 1:13:26SPEAKER_01: I'm going to try to like make online shopping better or the better assistant.
- 1:13:31SPEAKER_01: And I'm like, okay, a lot of people are doing that, you know, or on the other end is just too generic.
- 1:13:36SPEAKER_01: They're like, we're going to make your agents better, but what are the agents going to do?
- 1:13:40SPEAKER_01: You know?
- 1:13:42SPEAKER_01: it's a weird space where I think a lot of the people,
- 1:13:44SPEAKER_01: the most interested in the technology are like really into the thing,
- 1:13:49SPEAKER_01: you know,
- 1:13:49SPEAKER_01: and I would like to see more like practical,
- 1:13:52SPEAKER_01: but new applications,
- 1:13:53SPEAKER_01: like things that weren't possible before,
- 1:13:56SPEAKER_01: you know,
- 1:13:56SPEAKER_01: like before we had an iPhone,
- 1:13:57SPEAKER_01: Uber made no sense.
- 1:13:59SPEAKER_01: Yeah.
- 1:14:00SPEAKER_01: Like I think things on that level with AI will be, will be quite interesting.
- 1:14:03SPEAKER_01: And I want to know what those are going to be.
- 1:14:06SPEAKER_00: For sure, for sure.
- 1:14:08SPEAKER_00: All right.
- 1:14:09SPEAKER_00: So Nikolai, it's been great to have you on for the umpteenth time.
- 1:14:13SPEAKER_00: I'm sure we'll have you on again in the future.
- 1:14:15SPEAKER_00: The site is deepnews.com, news with a Z. So check that out.
- 1:14:22SPEAKER_00: It's in the links.
- 1:14:23SPEAKER_00: And, you know, I hope you guys are also using the various AI tools out there.
- 1:14:29SPEAKER_00: Not everyone needs to do it, but I think it's super interesting.
- 1:14:33SPEAKER_00: You learn a lot.
- 1:14:36SPEAKER_00: It's going to be a part of – kind of like the superstructure,
- 1:14:40SPEAKER_00: the framework of our worlds in a way is going to be AI-powered.
- 1:14:44SPEAKER_00: So I think it's a little bit – it behooves us to kind of know what's going on under
- 1:14:49SPEAKER_00: the hood of civilization.
- 1:14:52SPEAKER_00: You know what I'm saying?
- 1:14:54SPEAKER_00: Because I think that's one of the things with AI –
- 1:14:58SPEAKER_00: it's not going to be like some humanoid robot that's,
- 1:15:02SPEAKER_00: you know,
- 1:15:03SPEAKER_00: going to be our companion or something anytime soon.
- 1:15:06SPEAKER_00: I think there's gonna be a lot of business to business,
- 1:15:07SPEAKER_00: a lot of under the hood operations that will be accelerated.
- 1:15:10SPEAKER_00: I think I am optimistic that we will see the productivity numbers within the next decade.
- 1:15:14SPEAKER_00: And, you know, I'm not just optimistic.
- 1:15:18SPEAKER_00: I think the feds are optimistic because they're hoping that productivity growth
- 1:15:22SPEAKER_00: helps to save social security and other,
- 1:15:25SPEAKER_00: you know,
- 1:15:26SPEAKER_00: other benefits programs.
- 1:15:28SPEAKER_00: So we'll see.
- 1:15:29SPEAKER_00: Nikolai, it was great talking to you.
- 1:15:31SPEAKER_00: I will see you online.
- 1:15:33SPEAKER_01: Thank you, Razeeb.
- 1:15:35SPEAKER_00: Embryos have a 30% to 60% difference in the predisposition to severe diseases.
- 1:15:39SPEAKER_00: With ORCID, parents can identify their embryo at lowest genetic risk rather than just a random outcome.
- 1:15:44SPEAKER_00: Our genetic calculator can show the impact of embryo selection.
- 1:15:47SPEAKER_00: Discuss embryo screening and IVF with a genetics expert.
- 1:15:56SPEAKER_02: This podcast for kids.
- 1:16:26SPEAKER_02: This is my favorite podcast.