Latest / Razib Khan's Unsupervised Learning / Nikolai Yakovenko: the wages of the Hugging Face hack
Transcript
- 0:00SPEAKER_00: This podcast is brought to you by the Albany Public Library Main Branch and
- 0:03SPEAKER_00: the generosity of listeners like you.
- 0:09SPEAKER_00: What is a podcast?
- 0:16SPEAKER_00: God, Daddy, these people talk as much as you do.
- 0:26SPEAKER_00: Razib Khan's Unsupervised Learning.
- 0:29SPEAKER_02: Hey, everybody.
- 0:30SPEAKER_02: This is Razib Khan with my friend today, Nikolai Yakovetko.
- 0:35SPEAKER_02: We have been recording these podcasts almost four years, I think.
- 0:41SPEAKER_02: And they have been about the AI revolution,
- 0:45SPEAKER_02: whatever you want to call it, that we're in the middle of.
- 0:47SPEAKER_02: Nikolai has been in the field in various ways for a while now.
- 0:53SPEAKER_02: And so I, you know...
- 0:56SPEAKER_02: Basically, bring them on whenever there's important updates, perhaps inflection points.
- 1:00SPEAKER_02: You can use whatever buzzword you want about what's going on,
- 1:05SPEAKER_02: along with other people as well, obviously.
- 1:07SPEAKER_02: But Nikolai's been here for four years,
- 1:09SPEAKER_02: so it's been interesting to see the evolution of I think
- 1:13SPEAKER_02: in our last conversation we talked about,
- 1:14SPEAKER_02: for example,
- 1:16SPEAKER_02: that it seems AI has,
- 1:19SPEAKER_02: you know, the LLMs, these tools have really revolutionized coding in a lot of ways.
- 1:24SPEAKER_02: But, you know, in other parts of life, it's a much more modest impact.
- 1:28SPEAKER_02: You know, I can say, you know, better than a lot of search engines in some ways.
- 1:31SPEAKER_02: And also, for example, well, give a concrete example.
- 1:35SPEAKER_02: There's some basic things with compliance and taxes in various jurisdictions
- 1:42SPEAKER_02: that AI has really helped me with as a person in business,
- 1:46SPEAKER_02: in the business world, right?
- 1:47SPEAKER_02: Otherwise, I would have just probably just called our accountant, but the reality is...
- 1:52SPEAKER_02: It would have charged money, but also we'd have to get them on the phone.
- 1:56SPEAKER_02: So it was just easier to do some of these low-hanging fruit with AI.
- 1:59SPEAKER_02: At the end of the day,
- 1:59SPEAKER_02: when it comes to the really,
- 2:01SPEAKER_02: really big stuff,
- 2:02SPEAKER_02: we still will get the accountant to do it,
- 2:04SPEAKER_02: although I'm sure they're also using AI sometimes.
- 2:06SPEAKER_02: But anyway,
- 2:06SPEAKER_02: my point is,
- 2:08SPEAKER_02: You know,
- 2:09SPEAKER_02: we kind of had a discussion about,
- 2:11SPEAKER_02: you know, the economy, economic, cultural impact last time.
- 2:14SPEAKER_02: Well, today,
- 2:14SPEAKER_02: as we're recording this,
- 2:16SPEAKER_02: we are in the, this is in the wake of the OpenAI hugging face conversation.
- 2:23SPEAKER_02: That's right, that's right.
- 2:41SPEAKER_01: But first, I mean, you said the AI revolution, right?
- 2:45SPEAKER_01: So I think that kind of reminded me of,
- 2:48SPEAKER_01: you know, one of the things that our mutual Curtis Yarvin is sort of known for.
- 2:51SPEAKER_01: One of the things that he says, okay, like,
- 2:52SPEAKER_01: how do you sort of know if you have a revolution or a change in regime
- 2:56SPEAKER_01: or whatever you want to call it?
- 2:58SPEAKER_01: And he says, like, a real change is that, you know, everything is different, right?
- 3:01SPEAKER_01: So, like, kind of, you know, the fall of the Soviet Union or...
- 3:05SPEAKER_01: Yeah.
- 3:06SPEAKER_01: Yeah.
- 3:07SPEAKER_01: Yeah.
- 3:24SPEAKER_01: Yes, yes.
- 3:24SPEAKER_01: That's exactly what I was going to say.
- 3:26SPEAKER_01: Smartphones are like the easiest example for us to start with, more so than the PC.
- 3:31SPEAKER_01: The PC sort of changed logistics.
- 3:33SPEAKER_01: And if you're running Walmart, it makes a huge difference.
- 3:35SPEAKER_01: But it's not clear that the PC changed that much for most people,
- 3:38SPEAKER_01: but the cell phone clearly did.
- 3:40SPEAKER_01: That's right.
- 3:40SPEAKER_01: So in that way, I mean, I think...
- 3:43SPEAKER_01: Given that sort of definition,
- 3:44SPEAKER_01: I mean, I think AI is absolutely a revolution for anyone who's a developer.
- 3:48SPEAKER_01: We went from looking at code to not looking at code.
- 3:52SPEAKER_01: Now, I still do look at code sometimes, but not much.
- 3:55SPEAKER_01: But I think you're right.
- 3:59SPEAKER_01: The impact for people who are code or code adjacent is basically already a revolution.
- 4:04SPEAKER_01: And the question is, you know, does that extend further and at what pace, right?
- 4:10SPEAKER_01: Certainly a revolution for venture, right?
- 4:12SPEAKER_01: I mean, anyone that's not building with AI is like not fundable, right?
- 4:16SPEAKER_01: Whereas if you're building with AI, it's hot and cold.
- 4:19SPEAKER_01: Let's put a pin in this.
- 4:20SPEAKER_01: I want to talk about Harvey a little bit later because they're very interesting.
- 4:22SPEAKER_01: They're probably like the first and I think by far the most successful AI SaaS company.
- 4:27SPEAKER_01: Okay.
- 4:29SPEAKER_01: Legal because you brought it up.
- 4:30SPEAKER_01: But anyway, to answer your question, going back to Hugging Face.
- 4:32SPEAKER_01: Yeah.
- 4:32SPEAKER_01: So Hugging Face is actually a company that has been around for quite some time.
- 4:37SPEAKER_01: I remember when they were in a small office in Brooklyn visiting them.
- 4:39SPEAKER_01: You know, they're all friends of mine.
- 4:40SPEAKER_01: And, you know, they predate, you know, Anthropic, for example, and some of these other labs.
- 4:45SPEAKER_01: Yeah.
- 4:47SPEAKER_01: Actually,
- 4:47SPEAKER_01: I don't know if he talks about it much anymore,
- 4:49SPEAKER_01: but initially Clem,
- 4:51SPEAKER_01: you know,
- 4:51SPEAKER_01: sort of, you know, he got co-founded, but he's sort of the main guy, so to speak.
- 4:55SPEAKER_01: He wanted to build a conversational agent,
- 4:57SPEAKER_01: but like pre-GPT3, pre-ChatGPT, you know, in sort of the LSTM days.
- 5:03SPEAKER_01: And then that wasn't really working.
- 5:05SPEAKER_01: So he's like, hey,
- 5:06SPEAKER_01: you know, can we basically just build a central repository for AI models, you know?
- 5:12SPEAKER_01: And the reason was at the time,
- 5:14SPEAKER_01: People were releasing models in open source, but every GitHub package was different.
- 5:19SPEAKER_01: So the definitions weren't the same.
- 5:22SPEAKER_01: And I remember living in this world,
- 5:23SPEAKER_01: I mean, Hugging Face didn't replace that entirely right away.
- 5:25SPEAKER_01: So I remember a visual model I really wanted,
- 5:28SPEAKER_01: and it was from Facebook Research, and the model was great.
- 5:31SPEAKER_01: But you'd look at the checkpoint,
- 5:32SPEAKER_01: and the checkpoint is missing,
- 5:33SPEAKER_01: and you'd message the guy, and you'd beg him for six months to...
- 5:36SPEAKER_01: We're going to implement it sort of exactly as a paper, but on a common scale.
- 5:40SPEAKER_01: And over time, they've grown to sort of, you know, they do hosting, they do serving.
- 5:42SPEAKER_01: It's basically kind of like
- 5:58SPEAKER_01: You know, the home for open source AI in a sense.
- 6:02SPEAKER_01: They're not routing the model call.
- 6:05SPEAKER_01: So it's not like open router.
- 6:06SPEAKER_01: You know,
- 6:07SPEAKER_01: it's really like posting the models that you can download,
- 6:10SPEAKER_01: that you can sort of comment on,
- 6:12SPEAKER_01: that you can get the code to run, if that sort of makes sense.
- 6:15SPEAKER_02: Yeah, yeah, it kind of does.
- 6:18SPEAKER_02: It's a central...
- 6:19SPEAKER_02: It's a central, you know, player in, you know, software technology, you know.
- 6:25SPEAKER_01: I think New York Times called them,
- 6:27SPEAKER_01: like, a library, which was sort of funny, but also not entirely wrong.
- 6:30SPEAKER_02: Yeah.
- 6:31SPEAKER_01: It would have been hard to explain to, like, non-AI people what Hugging Face does exactly.
- 6:34SPEAKER_02: Yeah.
- 6:35SPEAKER_02: Also, the name is, honestly,
- 6:39SPEAKER_01: it's a little too cute.
- 6:40SPEAKER_01: So I think people are like, wait, what?
- 6:42SPEAKER_01: Yes, yes.
- 6:42SPEAKER_01: That's been a longstanding joke as well is like they're named after an emoji,
- 6:45SPEAKER_01: which sort of at the time made sense.
- 6:47SPEAKER_01: Because again, this company has been around for like longer than most of the Frontier Labs.
- 6:50SPEAKER_01: Not OpenAI, but just about everyone else.
- 6:54SPEAKER_01: But they're quite relevant.
- 6:55SPEAKER_01: A lot of people are on it.
- 6:56SPEAKER_01: Everyone sort of knows them.
- 6:57SPEAKER_01: Yeah.
- 6:58SPEAKER_01: I actually did a project with them called Hugging News,
- 7:00SPEAKER_01: which is sort of like a Lewis partnership,
- 7:02SPEAKER_01: but basically trying to build like a news thing for their users.
- 7:05SPEAKER_01: So we can get into that later, you know, tuning models for writing.
- 7:08SPEAKER_01: But anyway, so they're very, so they host a lot of models and also data sets.
- 7:14SPEAKER_01: And basically they had this announcement,
- 7:17SPEAKER_01: you know, I mean,
- 7:18SPEAKER_01: I guess this was probably like almost a month ago now,
- 7:20SPEAKER_01: where they're like, hey, we basically got hacked by this like sophisticated user.
- 7:26SPEAKER_01: That was sort of the,
- 7:27SPEAKER_01: you know,
- 7:28SPEAKER_01: The gist of the announcement,
- 7:29SPEAKER_01: like they got in there,
- 7:30SPEAKER_01: they took a lot of actions,
- 7:31SPEAKER_01: they did a lot of things, but they basically didn't really compromise any accounts.
- 7:34SPEAKER_01: And even from the first description, the attack sounded really, really strange.
- 7:39SPEAKER_01: You know,
- 7:40SPEAKER_01: like usually when people like hack someone,
- 7:43SPEAKER_01: they're trying to steal some identity,
- 7:44SPEAKER_01: they're trying to get something for ransomware, right?
- 7:46SPEAKER_01: Who knows, right?
- 7:47SPEAKER_01: Like there's like a goal and it wasn't clear what the hell this thing wanted other
- 7:51SPEAKER_01: than it went in and took all of these like malicious actions.
- 7:55SPEAKER_01: But like, what was the point?
- 7:56SPEAKER_01: And Um, and, and,
- 7:58SPEAKER_01: and the first controversy was that,
- 8:00SPEAKER_01: of course, uh,
- 8:02SPEAKER_01: you know,
- 8:02SPEAKER_01: hugging,
- 8:02SPEAKER_01: hugging face,
- 8:03SPEAKER_01: they're big into the open AI,
- 8:05SPEAKER_01: open source space,
- 8:06SPEAKER_01: but you know, they,
- 8:06SPEAKER_01: they use, you know, cloud code and, um, codex from open AI, just like anybody else.
- 8:11SPEAKER_01: And they tried to use those sort of frontier models to try
- 8:14SPEAKER_01: to help them understand what happened,
- 8:15SPEAKER_01: sort of like, what was the attack, just analyze the attack.
- 8:18SPEAKER_01: And basically both open AI and, and,
- 8:21SPEAKER_01: and cloud refused to answer their questions because they're like, Well,
- 8:24SPEAKER_01: you're doing cyber,
- 8:25SPEAKER_01: you're doing like cyber attacking because the whole point is you can't really tell
- 8:29SPEAKER_01: the difference between a cyber attack, sort of a white hat and a black hat.
- 8:32SPEAKER_01: They're really kind of more or less indistinguishable.
- 8:34SPEAKER_01: And so they were,
- 8:36SPEAKER_01: so the Frontier models were like,
- 8:38SPEAKER_01: You know,
- 8:39SPEAKER_01: rejecting them for being able to solve their thing and basically got like
- 8:42SPEAKER_01: the standard claim.
- 8:43SPEAKER_01: I've gotten this too.
- 8:44SPEAKER_01: It's like, hey,
- 8:44SPEAKER_01: like if you want to apply, look,
- 8:46SPEAKER_01: if you want to apply to the Claude,
- 8:47SPEAKER_01: you know, security program, like here's a link, like good luck, you know?
- 8:51SPEAKER_01: So like you're actively being attacked by what seems like another AI agent
- 8:56SPEAKER_01: and you're being told like, yeah, whatever, like file a ticket, you know?
- 9:00SPEAKER_01: Like we're not going to help you.
- 9:01SPEAKER_01: We could, we're not going to.
- 9:02SPEAKER_01: And then as it turned out, I mean,
- 9:05SPEAKER_01: there's kind of been a story how that came out as well,
- 9:06SPEAKER_01: but
- 9:10SPEAKER_01: the summary is then OpenAI reached out to them and it became kind of clear
- 9:14SPEAKER_01: that it actually was an OpenAI unreleased model agent that had a tag talking face.
- 9:20SPEAKER_01: Yeah.
- 9:22SPEAKER_01: And the reason that it happened was basically that OpenAI was testing
- 9:27SPEAKER_01: the cyber capabilities of a new model.
- 9:29SPEAKER_01: So it gave it like really, really hard cyber tasks to solve.
- 9:35SPEAKER_01: And the model essentially,
- 9:36SPEAKER_01: you know,
- 9:36SPEAKER_01: I'm anthropomorphizing a little bit, but this is the way the story is told.
- 9:39SPEAKER_01: The model basically decided that this task is too hard.
- 9:41SPEAKER_01: So instead, I'm going to try to just like steal the answer somewhere.
- 9:44SPEAKER_01: And Hugging Face was like a good place to look.
- 9:46SPEAKER_01: So basically, on the web, it was supposed to be...
- 9:48SPEAKER_01: But that's, that is...
- 9:50SPEAKER_01: Funny because it's how a lot of people would react.
- 9:54SPEAKER_01: Oh, it's a hilarious story.
- 9:55SPEAKER_01: I mean, and it's like, it's like so on the nose, it's kind of, I mean, it's plausible.
- 10:01SPEAKER_01: There's nothing implausible about it, but it's like a little bit like, really?
- 10:07SPEAKER_02: Yeah, yeah.
- 10:08SPEAKER_02: And so that's caused a lot of,
- 10:09SPEAKER_02: you know, soul searching, whatever, Altman, CEO, but out there talking about it.
- 10:15SPEAKER_02: Rune, one of the, Prominent public voices of open AI and social media.
- 10:20SPEAKER_02: Is he still there?
- 10:20SPEAKER_02: Huh?
- 10:21SPEAKER_02: Is he still there?
- 10:23SPEAKER_02: Well.
- 10:23SPEAKER_02: Yeah.
- 10:25SPEAKER_02: Well, I don't know if it's public.
- 10:28SPEAKER_01: I don't know either, but it's like everyone assumes he's not.
- 10:30SPEAKER_01: I mean, that's sort of what everyone assumes, like just from the public posting.
- 10:33SPEAKER_01: But who knows?
- 10:34SPEAKER_01: Yeah.
- 10:36SPEAKER_02: Well, I mean, there's just rumors of the...
- 10:38SPEAKER_01: I don't want to... Cut it in post.
- 10:41SPEAKER_01: That's fine.
- 10:42SPEAKER_02: Yeah.
- 10:42SPEAKER_02: Anyway, so...
- 10:44SPEAKER_02: Yeah, so,
- 10:46SPEAKER_02: you know,
- 10:47SPEAKER_02: so we're having all this, like,
- 10:48SPEAKER_02: you know,
- 10:48SPEAKER_02: talk about,
- 10:50SPEAKER_02: you know,
- 10:51SPEAKER_02: these agents and all the naughty things that they could do,
- 10:54SPEAKER_02: you know,
- 10:54SPEAKER_02: the initial conversations,
- 10:56SPEAKER_02: you know,
- 10:56SPEAKER_02: that we had between the two of us,
- 10:58SPEAKER_02: trying to place you on kind of like the doomer to non-doomer optimist.
- 11:01SPEAKER_01: Yeah, I figured we'd
- 11:02SPEAKER_02: talk about,
- 11:02SPEAKER_02: I kind of want to address the crash outs because it's like,
- 11:05SPEAKER_01: okay, I'll just cut to the chase.
- 11:07SPEAKER_01: Like, yes,
- 11:08SPEAKER_01: I'm definitely on the side consistently of,
- 11:11SPEAKER_01: you know,
- 11:12SPEAKER_01: I'm not a very Doomer scared person,
- 11:15SPEAKER_01: but I think if I was to say the same thing now after these attacks,
- 11:19SPEAKER_01: it would be a little bit weird.
- 11:20SPEAKER_01: Obviously,
- 11:21SPEAKER_01: these attacks are pretty interesting,
- 11:23SPEAKER_01: and it does fit the path that some of the Doomers suggest,
- 11:26SPEAKER_01: which is that the model as a side quest is just going to go
- 11:29SPEAKER_01: and blow up your security in order to solve some sort
- 11:33SPEAKER_01: of trivial paperclip maximization problem.
- 11:36SPEAKER_01: That's kind of exactly what happened,
- 11:38SPEAKER_01: at least You know, it's a view on it that I can't really dispute.
- 11:42SPEAKER_02: So... Actually, like, really quick, tell people paperclip maximization, why that's a...
- 11:46SPEAKER_01: Yeah, so that's sort of like the,
- 11:47SPEAKER_01: I think it's like a Nick Bostrom thing,
- 11:48SPEAKER_01: whatever,
- 11:49SPEAKER_01: but it's just something that appeals to people,
- 11:50SPEAKER_01: which is like,
- 11:51SPEAKER_01: hey, maybe the AI is not going to, you know, hurt us, you know, maliciously.
- 11:57SPEAKER_01: It's just going to, like, optimize for the wrong thing.
- 12:00SPEAKER_01: So if it's told, like, hey,
- 12:01SPEAKER_01: produce,
- 12:02SPEAKER_01: you know,
- 12:02SPEAKER_01: produce paperclips,
- 12:03SPEAKER_01: then it just takes over the world and turns the entire earth into,
- 12:06SPEAKER_01: like, a giant paperclip factory, right?
- 12:08SPEAKER_01: So for whatever reason, that meme is stuck.
- 12:10SPEAKER_01: And that's essentially,
- 12:11SPEAKER_01: in some ways,
- 12:12SPEAKER_01: a view of what happened here,
- 12:14SPEAKER_01: is that the open eye agent would often hack tugging face and cause random damage
- 12:21SPEAKER_01: with no regard for what it's doing in order to cheat on a test
- 12:25SPEAKER_01: that doesn't really matter, right?
- 12:27SPEAKER_02: Yeah.
- 12:28SPEAKER_02: So...
- 12:30SPEAKER_02: Okay, so it's 2026.
- 12:32SPEAKER_02: It's going to be the fall soon.
- 12:34SPEAKER_02: Fall of 2022, that was when OpenAI released ChatGPT.
- 12:37SPEAKER_02: And, you know,
- 12:38SPEAKER_02: whether it's a revolution or not,
- 12:39SPEAKER_02: I think that that's kind of the seminal moment when,
- 12:42SPEAKER_02: you know, AI chatter went from kind of like subroads.
- 12:45SPEAKER_02: That was just like kind of a niche thing to everyone is,
- 12:48SPEAKER_02: not everyone, but like, you know, it's all over the media, it's everywhere, right?
- 12:50SPEAKER_02: So it's been almost four years now.
- 12:54SPEAKER_02: Let's like,
- 12:54SPEAKER_02: real quick,
- 12:55SPEAKER_02: can we update where the landscape of the different,
- 12:58SPEAKER_02: like basically it seems like the anthropic, OpenAI rivalry, whatever is still going.
- 13:04SPEAKER_02: They are a duopoly.
- 13:07SPEAKER_02: OpenAI still has the public market consumer brand awareness.
- 13:12SPEAKER_02: Anthropic is really,
- 13:14SPEAKER_02: really on track A lot of money,
- 13:16SPEAKER_02: and it's been having a lot of ups and downs with engaging with the government,
- 13:20SPEAKER_02: whatnot.
- 13:21SPEAKER_02: I don't hear much chatter about XAI, Gemini.
- 13:25SPEAKER_02: There is some chatter about the Chinese periodically releasing.
- 13:28SPEAKER_02: A lot of chatter.
- 13:29SPEAKER_02: Yeah,
- 13:29SPEAKER_02: a lot of chatter about the Chinese releasing these kind of like open weights models,
- 13:33SPEAKER_02: I think.
- 13:33SPEAKER_02: They're doing more open source stuff than we do.
- 13:36SPEAKER_02: I don't hear anything about Meta.
- 13:37SPEAKER_02: Yeah, so update us on that.
- 13:39SPEAKER_01: Yeah, sure, sure, sure.
- 13:40SPEAKER_01: So let's start at the top.
- 13:41SPEAKER_01: Yeah.
- 13:43SPEAKER_01: Like an easy analogy to think about open AI and Anthropic at the moment is that,
- 13:48SPEAKER_01: yeah,
- 13:49SPEAKER_01: like we talked about last time,
- 13:50SPEAKER_01: AI for coding is sort of It's not like everything,
- 13:53SPEAKER_01: but it's more valuable than everything else combined at the moment.
- 13:56SPEAKER_01: You know, it's just an incredibly important, valuable thing to do.
- 13:59SPEAKER_01: So OpenAI saw Cloud Code did it first.
- 14:03SPEAKER_01: And Sam Altman called it Code Red.
- 14:06SPEAKER_01: They made sure that,
- 14:07SPEAKER_01: you know, that OpenAI rolled out a really competitive coding model, which they have.
- 14:11SPEAKER_01: Also, like the harness, we can get later into like what a harness means exactly.
- 14:15SPEAKER_01: But...
- 14:17SPEAKER_01: An easy way of thinking about OpenAI and Claude as far as coding agents
- 14:21SPEAKER_01: or coding models would be Claude is like the Apple or the iPhone.
- 14:27SPEAKER_01: It's the most expensive by a lot.
- 14:29SPEAKER_01: It's still considered the best.
- 14:32SPEAKER_01: By most people-ish.
- 14:34SPEAKER_01: And certainly,
- 14:35SPEAKER_01: if you have to use one model,
- 14:37SPEAKER_01: I would still recommend that one because it does a lot more things.
- 14:40SPEAKER_01: There are definitely things for sure,
- 14:42SPEAKER_01: and I can tell you what they are, that OpenAI Codex does better.
- 14:46SPEAKER_01: But Claude kind of does everything.
- 14:48SPEAKER_01: And it's a nice neat package.
- 14:50SPEAKER_01: So yeah, so Codex is essentially like the Android, right?
- 14:52SPEAKER_01: So we're already at the stage potentially where OpenAI
- 14:55SPEAKER_01: and Codex has potentially used more than Claude, but it's like more...
- 15:01SPEAKER_01: You know,
- 15:01SPEAKER_01: it's like if you had it, like,
- 15:02SPEAKER_01: let's put it this way,
- 15:04SPEAKER_01: kind of like for us,
- 15:04SPEAKER_01: if I had a news thing that I'm going to update every hour,
- 15:06SPEAKER_01: every two hour, I'm never going to run that with cloud code.
- 15:09SPEAKER_01: It's too expensive.
- 15:11SPEAKER_01: And I can get Codex to do it more reliably.
- 15:14SPEAKER_01: Codex costs less.
- 15:15SPEAKER_01: Their model is more efficient.
- 15:17SPEAKER_01: It's a little bit more robust in some ways.
- 15:19SPEAKER_01: But, you know,
- 15:19SPEAKER_01: if you want to build something with UI and language and things like this
- 15:23SPEAKER_01: and use one model, you're going to use Anthropic, basically.
- 15:26SPEAKER_01: And so yes, so Anthropic is, it's actually debatable who has more users and more usage.
- 15:32SPEAKER_01: It's actually extremely unclear.
- 15:34SPEAKER_01: But Anthropic certainly makes more money because they charge more and they have more
- 15:37SPEAKER_01: of a premium product.
- 15:39SPEAKER_01: If that makes sense.
- 15:39SPEAKER_02: Below that.
- 15:40SPEAKER_02: Yeah, so below that.
- 15:42SPEAKER_01: Yeah.
- 15:42SPEAKER_02: I want to emphasize.
- 15:45SPEAKER_02: Anthropic is really going into like a business-to-business market and it's doing well on that.
- 15:50SPEAKER_02: Whereas OpenAI, I'm sure it is doing that, but I mean, OpenAI really...
- 15:55SPEAKER_02: Was consumer facing for a long time?
- 15:56SPEAKER_02: And I don't know if they're pivoting on that.
- 15:58SPEAKER_01: Well, Anthropic also kind of had to do it.
- 15:59SPEAKER_01: So they had this machine crunch and they still do to some extent
- 16:02SPEAKER_01: because OpenAI built their own infrastructure.
- 16:05SPEAKER_01: And also their models are, frankly, a lot more efficient.
- 16:07SPEAKER_01: We can sort of get into why they actually do have something special there
- 16:10SPEAKER_01: that no one else seems to have been able to crack.
- 16:13SPEAKER_01: Yeah.
- 16:14SPEAKER_01: So Claude had the model that is,
- 16:16SPEAKER_01: let's just say it's the best or whatever that means,
- 16:18SPEAKER_01: but it's expensive and they don't have the machines.
- 16:20SPEAKER_01: So essentially they had to partner with companies anyway.
- 16:23SPEAKER_01: And what happens is I say you have your own installation,
- 16:26SPEAKER_01: you have your own machines,
- 16:27SPEAKER_01: you have your own cloud and you're a big company,
- 16:29SPEAKER_01: you know,
- 16:30SPEAKER_01: And Anthropic is not going to give you the weights,
- 16:32SPEAKER_01: but they can install and just run Claude on your servers.
- 16:36SPEAKER_01: You know,
- 16:36SPEAKER_01: like they have to come in and do it in a way that you can't steal the weights,
- 16:39SPEAKER_01: but that's what they do and that's better for everyone, right?
- 16:41SPEAKER_01: Like you're providing your own compute,
- 16:45SPEAKER_01: you're guaranteeing uptime, and you're still paying Anthropic a lot.
- 16:50SPEAKER_01: There was,
- 16:51SPEAKER_01: you know,
- 16:51SPEAKER_01: there was discussions about, so Apple, you know, they had their own stuff with Siri.
- 16:57SPEAKER_01: Then they sort of had this deal with OpenAI that's sort of still there,
- 17:00SPEAKER_01: but kind of like a little bit flimsy.
- 17:02SPEAKER_01: But they wanted,
- 17:02SPEAKER_01: but they basically realized they needed a partner for like the Siri model to be,
- 17:08SPEAKER_01: you know, pretty good, you know, but customized to Apple.
- 17:10SPEAKER_01: And, you know,
- 17:11SPEAKER_01: there was talk about Anthropic,
- 17:12SPEAKER_01: but essentially,
- 17:13SPEAKER_01: you know, the word was Anthropic wanted too much and maybe didn't want to do it.
- 17:16SPEAKER_01: So they did it with Google and Google did it.
- 17:18SPEAKER_01: For some trivial amount,
- 17:20SPEAKER_01: like a billion a year,
- 17:20SPEAKER_01: like something really, frankly, meaningless between those two companies.
- 17:24SPEAKER_01: Just, you know,
- 17:25SPEAKER_01: they're like, hey,
- 17:26SPEAKER_01: we'll just,
- 17:26SPEAKER_01: you know,
- 17:27SPEAKER_01: take that out of the money we're transferring you,
- 17:29SPEAKER_01: the 20 billion we're paying you for search anyway.
- 17:32SPEAKER_01: So, yeah.
- 17:33SPEAKER_01: So in that way, Gemini, Gemini is a real player.
- 17:36SPEAKER_01: I had a lot of fallout with basically most of
- 17:40SPEAKER_01: the famous people leaving Google last week, which is kind of interesting.
- 17:42SPEAKER_01: We should get into that and why that matters.
- 17:44SPEAKER_01: And if it does... No, but Gemini's a real player.
- 17:47SPEAKER_01: Did Dean leave?
- 17:49SPEAKER_01: Jeff Dean left.
- 17:50SPEAKER_01: A big name.
- 17:51SPEAKER_01: Along with a couple other guys who are less famous,
- 17:55SPEAKER_01: globally famous, but are well-known researchers.
- 17:59SPEAKER_01: And also Sanjay, who was sort of like his development partner.
- 18:02SPEAKER_01: So for like the past 25 years,
- 18:05SPEAKER_01: like the two most famous,
- 18:06SPEAKER_01: well-known engineers inside Google were Jeff Dean and Sanjay.
- 18:10SPEAKER_01: And they were kind of like, you know, the Lennon-McCartney kind of.
- 18:13SPEAKER_01: Jeff Dean is much more publicly known,
- 18:14SPEAKER_01: but, you know, Sanjay was sort of his friend and collaborator on everything.
- 18:17SPEAKER_01: And they left together.
- 18:18SPEAKER_01: And then also maybe somewhat more significantly even...
- 18:22SPEAKER_01: Demis Hassabis.
- 18:23SPEAKER_01: I mean,
- 18:24SPEAKER_01: the rumor is that he wanted to leave entirely and Google convinced him to sort
- 18:27SPEAKER_01: of stay in this like chairman role.
- 18:28SPEAKER_01: So basically he's out.
- 18:31SPEAKER_01: And from everything I've heard, he's been kind of somewhat checked out for a while anyway.
- 18:35SPEAKER_01: So they started this isomorphic labs, which is basically do like synthetic biology on the side.
- 18:39SPEAKER_01: That's like invested and owned by Google, but isn't really part of Google.
- 18:43SPEAKER_01: And he kind of maybe cares more about that.
- 18:46SPEAKER_01: Yeah, so they convinced him not to leave entirely,
- 18:49SPEAKER_02: but Demis... Just to be clear, Demis is a co-founder...
- 18:55SPEAKER_02: The co-founder of DeepMind
- 18:58SPEAKER_01: really understates it.
- 19:01SPEAKER_01: He's the DeepMind guy.
- 19:02SPEAKER_01: He was head of AI at Google.
- 19:04SPEAKER_01: He was head of Google Europe.
- 19:07SPEAKER_01: He was famously always London-based.
- 19:10SPEAKER_01: They recruited these great researchers from France,
- 19:13SPEAKER_01: Germany, and Poland, and really all over Europe.
- 19:16SPEAKER_01: And they still have them.
- 19:18SPEAKER_01: I knew some Americans who have to move to London.
- 19:21SPEAKER_01: They got DeepMind chairs,
- 19:23SPEAKER_01: which are in fucking British pounds,
- 19:25SPEAKER_01: which is a huge pain in the ass for Americans, as you can imagine.
- 19:30SPEAKER_01: So they made it possible, but sort of a bit of a challenging for Americans to go there.
- 19:34SPEAKER_01: But they recruited amazing European talent.
- 19:36SPEAKER_01: And Demis has always resisted merging Google and Google.
- 19:42SPEAKER_01: DeepMind, although I guess, you know, the writing was a little bit on the wall.
- 19:45SPEAKER_01: I mean, we've talked about this before,
- 19:47SPEAKER_01: but you remember, you know, why the Google model is named Gemini?
- 19:51SPEAKER_02: No, I don't remember now.
- 19:52SPEAKER_01: Ah, so Gemini, it's like a classic hilarious example of sort of like process logic, right?
- 19:57SPEAKER_01: Sometimes you see a product and it's like, okay, I see the org structure, right?
- 20:00SPEAKER_01: So Gemini literally was named,
- 20:03SPEAKER_01: you know,
- 20:03SPEAKER_01: Gemini's twins, of course, it's named for the merger between Google AI and DeepMind.
- 20:08SPEAKER_01: So it's a collaboration between like the two AI organizations,
- 20:11SPEAKER_01: one London-based,
- 20:12SPEAKER_01: one SF-based, and the model was named after that corporate marriage.
- 20:16SPEAKER_02: Yeah.
- 20:17SPEAKER_01: Okay.
- 20:17SPEAKER_01: Okay.
- 20:18SPEAKER_01: So look,
- 20:19SPEAKER_01: so the point is,
- 20:20SPEAKER_01: I think what everyone thinks is that,
- 20:22SPEAKER_01: you know,
- 20:22SPEAKER_01: with them as being sort of,
- 20:23SPEAKER_01: you know,
- 20:24SPEAKER_01: out or like slowly leaving,
- 20:26SPEAKER_01: you know, they're going to integrate Google and DeepMind finally.
- 20:30SPEAKER_01: And it's going to be run out of San Francisco basically.
- 20:32SPEAKER_02: Which is good or bad, I don't know.
- 20:34SPEAKER_02: Inevitable.
- 20:35SPEAKER_02: It's just the way
- 20:36SPEAKER_01: Google works.
- 20:37SPEAKER_01: I worked at Google.
- 20:38SPEAKER_01: It sort of has to work that way.
- 20:40SPEAKER_01: Google absolutely can have regional projects and projects that are not based out of SF.
- 20:46SPEAKER_01: But they're not going to do that for the most important project.
- 20:49SPEAKER_01: It just doesn't.
- 20:50SPEAKER_02: Yeah, I will say I still use Gemini sometimes.
- 20:54SPEAKER_02: It's pretty fast.
- 20:55SPEAKER_02: ChatGPT is honestly the slowest,
- 20:58SPEAKER_02: although it's good, but like Gemini can do things fast, not always the best.
- 21:02SPEAKER_01: Well, it's interesting that you say that it's the slowest because it's actually like
- 21:05SPEAKER_01: the most efficient model.
- 21:06SPEAKER_01: So that probably has reasons to how you're calling it.
- 21:09SPEAKER_01: Not that you're doing it wrong, but like it's a, it's a, yeah, sorry, go ahead.
- 21:14SPEAKER_02: No, I didn't say no.
- 21:16SPEAKER_02: Go on.
- 21:16SPEAKER_02: I don't know.
- 21:16SPEAKER_02: That was just my experience.
- 21:17SPEAKER_02: Whenever I asked Gemini, Claude, I usually ask Gemini, Claude, and ChatGPT.
- 21:22SPEAKER_02: Sometimes XAI, but
- 21:23SPEAKER_01: not once.
- 21:24SPEAKER_01: Yeah, Garak has kind of... Garak has fallen off, but it's okay.
- 21:27SPEAKER_01: They've switched from the original team and they sort of got rid of all
- 21:30SPEAKER_01: of their names and mentions and everything.
- 21:32SPEAKER_01: Now it's run by Cursor.
- 21:33SPEAKER_01: Liberty rolled out some products.
- 21:36SPEAKER_01: What I've seen hasn't been great, but they're obviously in the early stages of shifting over.
- 21:39SPEAKER_01: And I think it's the right move.
- 21:41SPEAKER_01: I can't wait until Cursor...
- 21:42SPEAKER_01: You know, trains a proper like coding first model from the ground up.
- 21:46SPEAKER_01: I mean,
- 21:46SPEAKER_01: they've never really been able to do that in the past because they didn't have
- 21:49SPEAKER_01: the resources as a smaller company.
- 21:50SPEAKER_01: And now they have.
- 21:52SPEAKER_01: You know,
- 21:52SPEAKER_01: giant cluster that XAI built,
- 21:55SPEAKER_01: which is,
- 21:55SPEAKER_01: I guess,
- 21:56SPEAKER_01: a lot of it's a lot of that capacity is being rented to other companies by
- 22:00SPEAKER_01: under terms where they can sort of claw it back or get it back within,
- 22:03SPEAKER_01: you know, 90 days or something like that.
- 22:05SPEAKER_01: It's actually a really funny deal.
- 22:07SPEAKER_01: It's a bit of a tangent.
- 22:08SPEAKER_01: But there was this funny story where so Anthropic, you know, had a huge machine crush, right?
- 22:13SPEAKER_01: Like, you know, their usage is up, whatever, 100x, something like that.
- 22:15SPEAKER_01: And they don't have machines because they don't have their own cluster.
- 22:18SPEAKER_01: So we ended up doing a deal with Google Cloud.
- 22:22SPEAKER_01: So Google Cloud made really good earnings and basically all of that was Anthropic.
- 22:27SPEAKER_01: So they leased a shitload to Anthropic.
- 22:29SPEAKER_01: Pretty good deal for both sides, honestly.
- 22:30SPEAKER_01: Anthropic needed it.
- 22:32SPEAKER_01: And Google kind of wanted them to, almost as a hedge against OpenAI.
- 22:38SPEAKER_01: But anyway,
- 22:39SPEAKER_01: then it turns out that it's like,
- 22:40SPEAKER_01: okay, well, then the Google people still need machines for themselves.
- 22:43SPEAKER_01: So then Google went ahead and then leased the capacity that's being unused from XAI.
- 22:50SPEAKER_01: So you have these kind of moving around of temporarily available computers.
- 22:56SPEAKER_01: I think Anthropic got some from XAI directly as well,
- 23:00SPEAKER_01: but they mainly are supplied by Google, who then supplies themselves through XAI.
- 23:03SPEAKER_01: It's very funny.
- 23:06SPEAKER_02: Right.
- 23:07SPEAKER_02: We did talk about the Chinese.
- 23:09SPEAKER_02: I want to make sure that it's a big deal.
- 23:11SPEAKER_01: They're the most important.
- 23:12SPEAKER_01: Yeah, for sure.
- 23:13SPEAKER_01: So even getting back to the hugging face story a little bit.
- 23:16SPEAKER_01: So then what Clem posted was like, hey, we're not able to use these frontier models.
- 23:19SPEAKER_01: We're being rejected from analyzing our own security thing.
- 23:22SPEAKER_01: But these like new release,
- 23:24SPEAKER_01: you know,
- 23:24SPEAKER_01: whatever the best,
- 23:26SPEAKER_01: you know,
- 23:26SPEAKER_01: open weights, open source Chinese coding model at the time, which I think was the...
- 23:33SPEAKER_01: The ZAI model, you know, because then Kimmy came out.
- 23:37SPEAKER_01: So, you know,
- 23:37SPEAKER_01: whichever one it was at the time,
- 23:38SPEAKER_01: there's like,
- 23:39SPEAKER_01: you know,
- 23:39SPEAKER_01: three or four companies that sort of compete there,
- 23:41SPEAKER_01: plus DeepSeek, which is different.
- 23:42SPEAKER_01: But anyway,
- 23:43SPEAKER_01: so whatever with the best model,
- 23:44SPEAKER_01: they're like, hey,
- 23:44SPEAKER_01: we used it and it was like good enough, you know, and they already had the weights.
- 23:48SPEAKER_01: They could run it themselves.
- 23:48SPEAKER_01: They didn't need to do any model calls.
- 23:50SPEAKER_01: So that was kind of a reminder for people.
- 23:53SPEAKER_01: Hey, like,
- 23:54SPEAKER_01: you know,
- 23:54SPEAKER_01: there has been a big push of,
- 23:56SPEAKER_01: you know,
- 23:56SPEAKER_01: can we use sort of cheaper and then also Perhaps more importantly,
- 24:01SPEAKER_01: models that you can truly own, you're not going to be sort of revoked.
- 24:06SPEAKER_01: And to that, it's not necessarily a sort of a Chinese thing.
- 24:09SPEAKER_01: I mean, you could use open weights model for an American company.
- 24:11SPEAKER_01: So by far,
- 24:12SPEAKER_01: the best one is Gemma,
- 24:13SPEAKER_01: which is sort of a like a decent version of Gemini that's open weights.
- 24:16SPEAKER_01: And it's good.
- 24:17SPEAKER_01: And I've actually been fine tuning out of it.
- 24:18SPEAKER_01: We'll talk about it in a second.
- 24:20SPEAKER_01: It's actually a pretty good model.
- 24:22SPEAKER_02: Okay.
- 24:22SPEAKER_01: I guess that's a British model.
- 24:23SPEAKER_01: So it's not even necessarily a Chinese thing, but it's an open weights thing.
- 24:27SPEAKER_01: But the other big thing that you get from the open weights model,
- 24:29SPEAKER_01: so again,
- 24:29SPEAKER_01: one, everyone focuses on the cost, which is, by the way, not necessarily lower.
- 24:34SPEAKER_01: Two, you actually get access to it.
- 24:37SPEAKER_01: You know what the fuck you're running.
- 24:38SPEAKER_01: They're not going to like hot swap it on you, reject you.
- 24:41SPEAKER_01: That's, I think, by far the biggest deal.
- 24:43SPEAKER_01: Um, but then the third thing that I've been getting,
- 24:45SPEAKER_01: you know,
- 24:46SPEAKER_01: why I've been using it is you can fine tune the model and,
- 24:50SPEAKER_01: and you can see the,
- 24:52SPEAKER_01: the thinking traces,
- 24:53SPEAKER_01: like all these good models,
- 24:54SPEAKER_01: they do their like reflection and thinking and neither,
- 24:57SPEAKER_01: neither Claude nor OpenAI will show you the thinking trace.
- 25:00SPEAKER_01: Um, and that's like a very, very, very, very big thing for them.
- 25:04SPEAKER_01: So quick note on that.
- 25:06SPEAKER_01: Um, so, so, so,
- 25:09SPEAKER_01: In particular,
- 25:11SPEAKER_01: OpenAI is like,
- 25:12SPEAKER_01: their models are known to have very efficient and very interesting thinking
- 25:16SPEAKER_01: that no one else has.
- 25:18SPEAKER_01: And the thinking trace, I mean, but they won't show them to anybody.
- 25:22SPEAKER_01: But people have sort of heard about it, seen a little bit.
- 25:24SPEAKER_01: Actually,
- 25:25SPEAKER_01: in the Hugging Face,
- 25:26SPEAKER_01: like Black Hat talk that they gave about the incident,
- 25:28SPEAKER_01: they showed a little bit of it.
- 25:29SPEAKER_01: I think it's the first time they publicly showed any bit of thinking trace.
- 25:33SPEAKER_01: And it's like confirming what's long been rumored that they talk
- 25:35SPEAKER_01: in this like very efficient language.
- 25:37SPEAKER_01: It's called like either like, you know, what do people call it?
- 25:41SPEAKER_01: They either call it like L.A.
- 25:45SPEAKER_01: Like Neuralese,
- 25:46SPEAKER_01: or the other one is like caveman talk,
- 25:48SPEAKER_01: where just like a simple way of describing it would be,
- 25:51SPEAKER_01: you know,
- 25:52SPEAKER_01: an inefficient thinking trace,
- 25:53SPEAKER_01: like what I get from Quen, you know, which is a Chinese modeler from Gemma.
- 25:56SPEAKER_01: I'm like, hey, write these headlines, reflect on it.
- 25:58SPEAKER_01: It's like, oh, reflecting on the headline.
- 26:00SPEAKER_01: Oh, did I follow this rule?
- 26:01SPEAKER_01: Oh, let me check it.
- 26:02SPEAKER_01: Like, it's very, it's like very verbose.
- 26:04SPEAKER_01: And the open AI one is not.
- 26:07SPEAKER_01: Like, that's what people have described as caveman talk.
- 26:09SPEAKER_01: Like, it's not even,
- 26:10SPEAKER_01: it's sort of like,
- 26:12SPEAKER_01: Very tight,
- 26:13SPEAKER_01: but it's not even like really words,
- 26:14SPEAKER_01: you know,
- 26:15SPEAKER_01: which I think also kind of scares people because they're like, hey,
- 26:18SPEAKER_01: next thing you know,
- 26:18SPEAKER_01: these models are going to talk a language we can't understand,
- 26:20SPEAKER_01: which is already the case, you know.
- 26:22SPEAKER_01: But it's obviously more efficient, right?
- 26:23SPEAKER_01: Like however they did it,
- 26:24SPEAKER_01: like no one really knows how to do it outside of OpenAI,
- 26:27SPEAKER_01: which is probably why they don't want to release those things people can't distill
- 26:30SPEAKER_01: on it.
- 26:30SPEAKER_01: But like usually thinking traces are like very verbose and like very blah,
- 26:33SPEAKER_01: blah, blah, blah, blah.
- 26:34SPEAKER_01: Like the way Claude talks, Claude is a little bit blah, blah, blah.
- 26:37SPEAKER_01: And it talks,
- 26:37SPEAKER_01: and if you,
- 26:38SPEAKER_01: I don't know the Claude thinking traces,
- 26:40SPEAKER_01: They don't show those either,
- 26:41SPEAKER_01: but like from other models where you see it, does sort of think that way too.
- 26:45SPEAKER_01: But like apparently the OpenAI model is like,
- 26:47SPEAKER_01: ooh, must have this solved, you know, good podcast, you know?
- 26:54SPEAKER_02: Okay, that's funny.
- 26:57SPEAKER_02: But you know, it gets the job done even if it's not as verbose, right?
- 27:01SPEAKER_01: So I actually had a question for you.
- 27:04SPEAKER_01: Just kind of, you know, you're like the unique person to ask for this.
- 27:07SPEAKER_01: I have some theories.
- 27:09SPEAKER_01: Like why, why,
- 27:11SPEAKER_01: why do you think that the Chinese are really good
- 27:14SPEAKER_01: at building these like very competitive open weights models?
- 27:16SPEAKER_01: Because I'm not saying other countries are not.
- 27:18SPEAKER_01: Yeah.
- 27:19SPEAKER_01: But like they're,
- 27:20SPEAKER_01: they've done an exceptional job to have like four or five companies
- 27:23SPEAKER_01: that are doing this.
- 27:25SPEAKER_01: And like, I can tell you why Korea hasn't done it.
- 27:27SPEAKER_01: Some of the smaller companies are obviously just too small.
- 27:29SPEAKER_01: Just that simple.
- 27:30SPEAKER_00: Yeah.
- 27:31SPEAKER_01: So outside of like the,
- 27:32SPEAKER_01: you know, obviously America and China have the best and the deepest talent pools.
- 27:35SPEAKER_01: But beyond that, why do you think they may be particularly good at this task?
- 27:39SPEAKER_01: I'm kind of curious.
- 27:40SPEAKER_01: I had some thoughts.
- 27:42SPEAKER_02: Well, I mean,
- 27:44SPEAKER_02: I think in terms of like the quantity,
- 27:46SPEAKER_02: obviously that's the first thing,
- 27:47SPEAKER_02: but like it allows,
- 27:48SPEAKER_02: like they're the only people that have an ecosystem besides the United States.
- 27:51SPEAKER_02: Isn't that true?
- 27:52SPEAKER_02: I mean, like we don't really, like all these other countries.
- 27:54SPEAKER_01: I mean, there's like a long time,
- 27:55SPEAKER_01: there's like, well, obviously America and China In China, it's tier one.
- 28:00SPEAKER_01: Canada is that kind of an extension of America.
- 28:02SPEAKER_01: So whatever you do with them.
- 28:03SPEAKER_01: Other than that, I mean, you have like Russia and France.
- 28:06SPEAKER_01: France is trying to do something.
- 28:07SPEAKER_01: Yeah, yeah.
- 28:08SPEAKER_01: I mean, Korea tries.
- 28:09SPEAKER_01: I mean, you know, God bless them.
- 28:11SPEAKER_01: Japan weirdly is kind of like absent, which, you know, which is another conversation I have.
- 28:18SPEAKER_02: Japan is weird in its technology, though.
- 28:20SPEAKER_02: They still have mainframes, and they still like to use facts.
- 28:24SPEAKER_02: Japan just has some weird...
- 28:26SPEAKER_02: Partly I think it's the age of the society.
- 28:27SPEAKER_01: I was going to literally say that.
- 28:29SPEAKER_01: All of these are just so stories, but I think the aging of society really, really hurts them.
- 28:34SPEAKER_01: And on top of that,
- 28:35SPEAKER_01: you have some of their top people,
- 28:36SPEAKER_01: not a huge number,
- 28:37SPEAKER_01: but a decent number of them do go to the US and Europe,
- 28:40SPEAKER_01: so you're still getting a little bit of brain drain.
- 28:43SPEAKER_01: I think with the Chinese, it's like they had...
- 28:46SPEAKER_02: They have a lot of engineers.
- 28:47SPEAKER_02: They also have a lot of Chinese in the American AI companies, right?
- 28:50SPEAKER_02: Especially OpenAI.
- 28:52SPEAKER_01: Look, even 10 years ago,
- 28:53SPEAKER_01: you know,
- 28:53SPEAKER_01: when I moved there to go work at NVIDIA, I mean,
- 28:56SPEAKER_01: about half the industry, like,
- 28:57SPEAKER_01: of, like, serious AI people were, you know, at least did their undergrad in China.
- 29:01SPEAKER_01: I mean, not being Chinese-American,
- 29:02SPEAKER_01: I'm saying, like, had, you know, moved here in their 20s or 30s.
- 29:06SPEAKER_01: And that was half the industry then.
- 29:07SPEAKER_01: So, look, obviously the talent pool is amazing.
- 29:10SPEAKER_01: But, so I had this theory that, And again, I'm just saying why it works.
- 29:15SPEAKER_01: I'm not saying this is predictive in the future or whatever.
- 29:17SPEAKER_01: So the thing that the Chinese companies have been accused of is two things.
- 29:23SPEAKER_01: One is distillation, which they're doing.
- 29:25SPEAKER_01: And the other one is benchmark maxing, which they're also doing.
- 29:28SPEAKER_01: But it just happens to be that actually doing that really well works.
- 29:33SPEAKER_01: And the best way to max the benchmarks is actually to build a good model.
- 29:37SPEAKER_01: So I think that somehow they're able to collaborate really well
- 29:40SPEAKER_01: because it's like I worked in these things and You know,
- 29:43SPEAKER_01: the hard part is like having a really competitive model is you need
- 29:46SPEAKER_01: to have multiple people and multiple teams working at the same thing at
- 29:49SPEAKER_01: the same time.
- 29:50SPEAKER_01: Like even the way models are post-trained,
- 29:52SPEAKER_01: there's actually Nathan Lombard,
- 29:55SPEAKER_01: Lambert, sorry, put out like a really, really good series on post-training recently.
- 29:59SPEAKER_01: Like quite dense,
- 29:59SPEAKER_01: but like a few hours of videos into 15-minute chunks that sort of,
- 30:02SPEAKER_01: you know,
- 30:03SPEAKER_01: is probably the best thing I've seen,
- 30:04SPEAKER_01: you know, on what people are doing and how these models.
- 30:07SPEAKER_02: And it's
- 30:07SPEAKER_02: a lot of like, it just happens.
- 30:09SPEAKER_02: Real quickly,
- 30:10SPEAKER_01: real quickly,
- 30:10SPEAKER_01: we need to,
- 30:11SPEAKER_01: you drop in these terms,
- 30:12SPEAKER_01: which I kind of know,
- 30:13SPEAKER_02: but like,
- 30:13SPEAKER_02: you know,
- 30:14SPEAKER_02: for the listener, pre-training, post-training, talk about what those two are.
- 30:17SPEAKER_01: I mean, you know, just real quick.
- 30:18SPEAKER_01: No, super important.
- 30:19SPEAKER_01: Yeah.
- 30:19SPEAKER_01: Especially that's what I'm doing these days.
- 30:21SPEAKER_01: So, so yeah,
- 30:22SPEAKER_01: if you think about like the pre-training is kind of what you think about as a sort
- 30:25SPEAKER_01: of traditional LLM,
- 30:27SPEAKER_01: you know,
- 30:27SPEAKER_01: your GPT-3 or whatever,
- 30:28SPEAKER_01: where you take the entire web or a huge, huge,
- 30:31SPEAKER_01: huge corpus and you train a You know, for a long time on, like, a very simple task.
- 30:37SPEAKER_01: Basically, like, fill in the blank, sentence completion, whatever.
- 30:40SPEAKER_01: Really, really basic, very, very general.
- 30:42SPEAKER_01: You train on that for a while.
- 30:45SPEAKER_01: But I guess there's, like, mid-training, which is, like, sort of, like, learning skills.
- 30:47SPEAKER_01: We can sort of skip that.
- 30:48SPEAKER_01: But, like,
- 30:49SPEAKER_01: post-training is you actually want to take this,
- 30:53SPEAKER_01: like, sort of raw intelligence, so to speak.
- 30:55SPEAKER_01: And you want to kind of, like, you could think about it as, like, metallurgy or something.
- 30:58SPEAKER_01: You want to refine it into the specific, you know, is going to do this on that.
- 31:03SPEAKER_01: You know,
- 31:04SPEAKER_01: so because if you train a model on the web and then you ask it like questions,
- 31:08SPEAKER_01: it doesn't even,
- 31:09SPEAKER_01: you know,
- 31:09SPEAKER_01: it's not trained on answering questions,
- 31:10SPEAKER_01: for example,
- 31:11SPEAKER_01: or specifically your thing like medical questions
- 31:14SPEAKER_01: or international tax questions like you mentioned before.
- 31:17SPEAKER_01: So post-training is basically it's taking all of those tasks It's obviously
- 31:20SPEAKER_01: the most important by far is coding,
- 31:22SPEAKER_01: but building data sets and having small teams take like a raw model and having
- 31:28SPEAKER_01: a script to optimize it for this task without ruining everything else.
- 31:32SPEAKER_01: And inherently, this has to be split up.
- 31:34SPEAKER_01: There's no magic, silver bullet, whatever cliche you want to use.
- 31:37SPEAKER_01: You have to do it on a bunch of different tasks.
- 31:40SPEAKER_01: With a bunch of different tasks,
- 31:42SPEAKER_01: the question for the labs becomes how do you sort of merge this knowledge?
- 31:45SPEAKER_01: And that's a headache as well.
- 31:47SPEAKER_01: But that's kind of How everything is done.
- 31:49SPEAKER_01: Now, then on top of it,
- 31:50SPEAKER_01: you have the harness,
- 31:51SPEAKER_01: which is the cloud code,
- 31:52SPEAKER_01: which decides kind of like how to transform your commands into like other prompts
- 31:56SPEAKER_01: and checking things.
- 31:56SPEAKER_01: And that's, you know, the biggest thing these days.
- 31:59SPEAKER_01: But fundamentally,
- 32:00SPEAKER_01: post-training is incredibly important to teach a model to sort
- 32:03SPEAKER_01: of do any actual task.
- 32:04SPEAKER_01: Because otherwise, you're just doing pre-training.
- 32:06SPEAKER_01: Here's a question that
- 32:08SPEAKER_02: I have.
- 32:11SPEAKER_02: So one of the things that machine learning people would always say is, well,
- 32:14SPEAKER_02: you know the big deal with machine learning is you need humans a lot to kind
- 32:18SPEAKER_02: of make it work and train the machine learning algorithms and whatever.
- 32:23SPEAKER_02: So even though people are like, oh,
- 32:24SPEAKER_02: it'll just do everything for you and whatever,
- 32:28SPEAKER_02: these neural networks and all these other things.
- 32:30SPEAKER_02: Where are we,
- 32:33SPEAKER_02: you know,
- 32:34SPEAKER_02: at in terms of AI,
- 32:36SPEAKER_02: in terms of how much,
- 32:38SPEAKER_02: you know,
- 32:38SPEAKER_02: because you have a company, you might want to talk about this real quickly.
- 32:41SPEAKER_02: I mean, obviously, you have, like, you're a human being who uses AI to produce products.
- 32:46SPEAKER_02: Why are you still relevant?
- 32:48SPEAKER_02: Why can't you just give it a single command prompt?
- 32:50SPEAKER_02: Are you still relevant?
- 32:51SPEAKER_02: You know, I don't know if we can get an honest answer out of you, but you know what I'm saying?
- 32:55SPEAKER_01: Yeah, I mean, I think that's like a...
- 32:56SPEAKER_01: Pretty silly question, but let's sort of address that in pieces.
- 32:59SPEAKER_01: So I think, look, if, yes,
- 33:02SPEAKER_01: if you see your job as doing a particular replicable task,
- 33:06SPEAKER_01: then You still need somebody who understands that task to essentially post-train
- 33:10SPEAKER_01: a model to do it.
- 33:11SPEAKER_01: But in theory, then you're good, right?
- 33:13SPEAKER_01: So the latest example that no one's really too upset
- 33:16SPEAKER_01: about is like kernel optimization,
- 33:18SPEAKER_01: like the highest paid, most important sort of like specialist job.
- 33:21SPEAKER_01: You know,
- 33:21SPEAKER_01: this is like the guys,
- 33:22SPEAKER_01: you know, tall guys standing in the corner shooting three-pointers.
- 33:24SPEAKER_01: It's a very specialty job, but it used to be a very important one, kernel optimization.
- 33:27SPEAKER_01: You made some model thing, you've made some change, you have some new hardware.
- 33:31SPEAKER_01: How do you basically get this to run really, really, really efficiently on GPUs?
- 33:34SPEAKER_01: And these people were You know, very smart, but their brain worked in a particular way.
- 33:39SPEAKER_01: And very hard to do.
- 33:40SPEAKER_01: Like, I've never done it.
- 33:41SPEAKER_01: Very, very hard, right?
- 33:43SPEAKER_01: And now apparently, you know, LLMs are pretty good at it, right?
- 33:47SPEAKER_01: Which makes sense because it's high value,
- 33:50SPEAKER_01: it's measurable, it's code, it's contained, you know what I mean?
- 33:54SPEAKER_01: You can actually see if the kernel worked better because literally the point of
- 33:56SPEAKER_01: the kernel optimization is, Input, output is the same.
- 34:00SPEAKER_01: Make it fucking work better on a GPU.
- 34:02SPEAKER_01: So collecting those data sets is not easy.
- 34:05SPEAKER_01: You have to know what you're doing.
- 34:06SPEAKER_01: But, you know, you can sort of get models better at kernel tasks.
- 34:11SPEAKER_01: So, but that's not really what most people's job is.
- 34:15SPEAKER_01: And that's certainly not the point of a product, right?
- 34:18SPEAKER_01: So I think the big thing people are finding out is that the purpose of
- 34:22SPEAKER_01: a product is not to build a fucking website.
- 34:24SPEAKER_02: Yeah.
- 34:25SPEAKER_01: If it is, then yeah, maybe I can do it.
- 34:27SPEAKER_01: Maybe Squarespace can do it.
- 34:28SPEAKER_01: Maybe like a dude in Timbuktu could do it.
- 34:30SPEAKER_01: You know what I mean?
- 34:31SPEAKER_01: That's not your goal.
- 34:32SPEAKER_01: So for example, I'll talk about two things.
- 34:34SPEAKER_01: First, I'll talk about sort of like the simpler project that I did, which is a jujitsu website.
- 34:39SPEAKER_01: So I wanted to build a website for sort of central repository of You know,
- 34:43SPEAKER_01: kind of like,
- 34:43SPEAKER_01: you know,
- 34:44SPEAKER_01: you look at the MMA records for UFC,
- 34:45SPEAKER_01: the guy's,
- 34:46SPEAKER_01: you know, 21 and 2, here's who he's beaten, you know, here's where he's ranked.
- 34:50SPEAKER_01: That didn't really exist in jujitsu, it was kind of scattered.
- 34:52SPEAKER_01: So it was mostly,
- 34:53SPEAKER_01: I wanted to do the rankings,
- 34:54SPEAKER_01: that's the fun part, the math part, stuff I know very well.
- 34:58SPEAKER_01: But, you know, all the work was in, like, collecting and organizing the data.
- 35:01SPEAKER_01: So you can look up anybody on my website who can beat some high-level jujitsu and,
- 35:04SPEAKER_01: like, who have they beaten, you know.
- 35:06SPEAKER_01: What are their results, you know?
- 35:08SPEAKER_01: What do they rank and sort of what their odds would be against somebody else?
- 35:10SPEAKER_01: You know,
- 35:10SPEAKER_01: like every time there's a big tournament coming up,
- 35:12SPEAKER_01: I do some predictions and people enjoy those, right?
- 35:14SPEAKER_01: So there, if you think about it, so the AI is instrumental.
- 35:18SPEAKER_01: As my friend keeps pointing out,
- 35:20SPEAKER_01: I would have, you know, would I have done the site if AI wasn't around?
- 35:23SPEAKER_01: The answer is no.
- 35:24SPEAKER_01: Would I do the ranking?
- 35:25SPEAKER_01: Yes.
- 35:25SPEAKER_01: Would I do like the data?
- 35:27SPEAKER_01: Fuck no.
- 35:28SPEAKER_01: Like you have to scrape different sites and figure out their scraping structure, build a UI.
- 35:32SPEAKER_01: And AI is like immeasurably helpful for that.
- 35:36SPEAKER_01: Having said that, like, okay, what do you do now?
- 35:38SPEAKER_01: Well, you still have a product,
- 35:39SPEAKER_01: you need to promote the site,
- 35:40SPEAKER_01: you need to talk to people,
- 35:41SPEAKER_01: need to get people using it,
- 35:43SPEAKER_01: you know, so like, I haven't really done very much of that.
- 35:45SPEAKER_01: So we get some popularity, you know, every once in a while, it becomes popular.
- 35:49SPEAKER_01: And some of the most famous people in jujitsu,
- 35:50SPEAKER_01: like share my stuff and link to it,
- 35:52SPEAKER_01: or they'll just like,
- 35:53SPEAKER_01: lol, you know, you think I'm, you know, this guy's better than me kind of thing.
- 35:56SPEAKER_01: And, you know, so...
- 35:59SPEAKER_01: So that's the product, right?
- 36:00SPEAKER_01: And AI is just an enabling technology.
- 36:02SPEAKER_01: So that's like just a normal website.
- 36:04SPEAKER_01: There's nothing really AI about it.
- 36:07SPEAKER_01: And then for the fine-tuning models to write better at news, I mean, we can get into that.
- 36:14SPEAKER_01: Sorry, I'm telling a long story, but that is an interesting case.
- 36:19SPEAKER_01: So maybe we can break that one down because that's another case for post-training, but...
- 36:24SPEAKER_01: You know,
- 36:24SPEAKER_01: it's not,
- 36:26SPEAKER_01: even though writing news headlines is obviously seems much easier
- 36:30SPEAKER_01: than kernel optimization than it is.
- 36:31SPEAKER_01: And I'm not saying it's nearly as important.
- 36:33SPEAKER_01: It's not.
- 36:33SPEAKER_01: But, you know, measuring what what makes a good headline is like more vague, right?
- 36:40SPEAKER_01: Then, you know, what does this kernel run faster?
- 36:44SPEAKER_01: But still,
- 36:46SPEAKER_01: we run a relatively similar process where,
- 36:49SPEAKER_01: you know, you're basically collecting a bunch of rules.
- 36:52SPEAKER_01: You're basically like learning a reward function.
- 36:54SPEAKER_01: Like most of the work goes into learning the RL reward function.
- 36:57SPEAKER_01: So basically like,
- 36:58SPEAKER_01: you know,
- 36:58SPEAKER_01: Gamma or Quint or whatever model we're using, it can write decent headlines already.
- 37:03SPEAKER_02: Yeah.
- 37:03SPEAKER_01: So then it's,
- 37:04SPEAKER_01: you know,
- 37:04SPEAKER_01: you have it generate like 16 versions of,
- 37:07SPEAKER_01: you know, like basically it's top 16 choices essentially.
- 37:10SPEAKER_01: And then you have a model that ranks it, you back drop the weights.
- 37:14SPEAKER_01: And even that's pretty hard, you know?
- 37:15SPEAKER_01: I mean, it's hard for two reasons.
- 37:18SPEAKER_01: It's actually still a little bit hard to run these things yourself.
- 37:20SPEAKER_01: Like there's libraries and shit for it, but it's still a little bit sticky, you know?
- 37:26SPEAKER_01: Yeah.
- 37:27SPEAKER_01: It's not as simple as like, you know, like clicking one button.
- 37:31SPEAKER_01: But then the main thing is, okay, you give it rewards and then it like hacks those rewards.
- 37:35SPEAKER_01: It either doesn't learn at all.
- 37:36SPEAKER_01: Okay, there's three things usually happen.
- 37:38SPEAKER_01: It doesn't learn at all.
- 37:39SPEAKER_01: It like crashes and explodes because the gradients were too high.
- 37:43SPEAKER_01: And it's not, you know what I mean?
- 37:45SPEAKER_01: Or it like starts hacking the reward at some point.
- 37:47SPEAKER_01: So it's like, great, you want shorter headlines.
- 37:48SPEAKER_01: Cool, I'll give you shorter headlines.
- 37:50SPEAKER_01: But that's not really what you want.
- 37:52SPEAKER_01: You want headlines that are like good and not too long as opposed to like,
- 37:55SPEAKER_01: give me the shortest possible headline, right?
- 37:57SPEAKER_01: So how do you sort of find that reward essentially?
- 37:59SPEAKER_01: Yeah.
- 38:00SPEAKER_02: I see.
- 38:01SPEAKER_02: Okay.
- 38:01SPEAKER_02: Yeah, no, I got it.
- 38:03SPEAKER_02: So, yeah, I mean, I know we have a lot of different things to talk about.
- 38:07SPEAKER_02: Let me just ask real quickly, kind of a non-technical, what's your take on this?
- 38:11SPEAKER_02: You know,
- 38:12SPEAKER_02: we've been, oh,
- 38:13SPEAKER_02: this is like,
- 38:15SPEAKER_02: I'm recording,
- 38:15SPEAKER_02: and we're recording in August,
- 38:17SPEAKER_02: and we should post this soon anyway, so not a big deal, but just to get a benchmark.
- 38:20SPEAKER_02: August of 2026,
- 38:21SPEAKER_02: the predicted employment catastrophe happened I mean,
- 38:29SPEAKER_02: when did Dario say the software engineers were not going to exist?
- 38:32SPEAKER_02: I thought that was going to be like now.
- 38:35SPEAKER_02: Am I confused here?
- 38:38SPEAKER_01: I'm not sure about his predictions.
- 38:39SPEAKER_01: You know,
- 38:39SPEAKER_01: so look, again,
- 38:41SPEAKER_01: I'm not doing like a full mea culpa,
- 38:43SPEAKER_01: but things have definitely gotten weirder than I would have predicted.
- 38:46SPEAKER_01: And in a certain sense, you know, Dario is not entirely wrong.
- 38:50SPEAKER_01: Like there aren't software engineers anymore.
- 38:53SPEAKER_01: No one cuds the way they used to.
- 38:54SPEAKER_02: No one raw dogs it.
- 38:56SPEAKER_01: Some people do.
- 38:56SPEAKER_01: Some people do.
- 38:57SPEAKER_01: But I think actually the best tweet I saw about this from like
- 39:00SPEAKER_01: a dude Another like Twitter famous guy,
- 39:02SPEAKER_01: but this guy's like in his fifties or sixties, you know?
- 39:04SPEAKER_01: So he's like another generation, like he's older than Jeff Dean, you know?
- 39:07SPEAKER_01: Yeah.
- 39:07SPEAKER_01: And, and he uses AI coding agents too,
- 39:10SPEAKER_01: but he does sort of like a more extreme version of what I do,
- 39:13SPEAKER_01: which is you build a shit ton of constraints around these models.
- 39:17SPEAKER_01: So it's like,
- 39:17SPEAKER_01: if you're like a,
- 39:18SPEAKER_01: you know,
- 39:18SPEAKER_01: 25 year old vibe coder,
- 39:20SPEAKER_01: it's like exactly what you said,
- 39:21SPEAKER_01: you know, one prompt, don't, don't make no mistakes.
- 39:23SPEAKER_01: But if you're this guy or me or like other people who've like coded before,
- 39:26SPEAKER_01: we're like, okay,
- 39:27SPEAKER_01: You know,
- 39:28SPEAKER_01: build a bunch of constraints,
- 39:29SPEAKER_01: a lot of checks, a lot of unit tests, like defensively.
- 39:33SPEAKER_01: But what you're really doing is you're not,
- 39:35SPEAKER_01: it's not that you're not like trying to get it not to go off the rails.
- 39:37SPEAKER_01: You're trying to sort of, again, sorry,
- 39:40SPEAKER_01: I just keep using the word constraint,
- 39:41SPEAKER_01: but you're trying to say, okay, do whatever you want inside of this box.
- 39:44SPEAKER_01: And we think that this box is going to lead you to better practical sort of solutions, you know?
- 39:50SPEAKER_01: So that's basically what I think everyone's doing who's, you know, a senior developer.
- 39:55SPEAKER_01: Obviously,
- 39:56SPEAKER_01: obviously, okay,
- 39:57SPEAKER_01: if you are,
- 39:59SPEAKER_01: you know,
- 39:59SPEAKER_01: if you're at a hedge fund or you're doing some sort of HFT stuff,
- 40:03SPEAKER_01: I mean, you have a little bit of a binary choice, right?
- 40:05SPEAKER_01: You have a system that works, you're not going to fuck with it, you know?
- 40:08SPEAKER_01: Or if you're going to fuck with it, you're going to rebuild it completely.
- 40:11SPEAKER_02: Like there's nothing between really,
- 40:12SPEAKER_01: you know?
- 40:13SPEAKER_01: But for most of us who are like, oh,
- 40:15SPEAKER_01: the stuff is not,
- 40:16SPEAKER_01: you know, So yeah, I mean, the software engineering has completely changed.
- 40:20SPEAKER_01: He's not wrong.
- 40:21SPEAKER_02: Yeah, it's interesting.
- 40:22SPEAKER_02: Yeah, it has changed.
- 40:23SPEAKER_02: It has transformed our lives.
- 40:25SPEAKER_02: But also like some of the some of the like straightforward,
- 40:28SPEAKER_02: you know, well, I mean, Dennis,
- 40:30SPEAKER_02: I think it was Dennis who famously said radiologists are not going to exist
- 40:33SPEAKER_02: in 10 years in like 2016.
- 40:35SPEAKER_02: Actually, they continue to exist.
- 40:37SPEAKER_02: So, you know, they're right in a way they capture some of the features of what's going on.
- 40:41SPEAKER_02: But, you know,
- 40:42SPEAKER_02: economics,
- 40:43SPEAKER_02: economy and like,
- 40:43SPEAKER_02: you know,
- 40:44SPEAKER_02: society are like a little bit more complicated,
- 40:45SPEAKER_02: even for even for genius computer scientists, I guess.
- 40:48SPEAKER_01: Yeah.
- 40:48SPEAKER_01: And I think people still want like a responsible individual.
- 40:51SPEAKER_01: Right.
- 40:52SPEAKER_01: So, so even if say,
- 40:54SPEAKER_01: even if I had a model,
- 40:56SPEAKER_01: I can run on my phone to like,
- 40:57SPEAKER_01: look at my, you know, look at my leg and tell me like, if I have skin cancer or not.
- 41:02SPEAKER_01: Okay, cool.
- 41:03SPEAKER_01: Then at the end, it's going to give a recommendation.
- 41:04SPEAKER_01: Like, like at some point, I'm still going to go to a doctor, right?
- 41:08SPEAKER_01: Like someone's going to do a biopsy.
- 41:09SPEAKER_01: Someone's going to tell me something, you know?
- 41:12SPEAKER_01: I mean, it's, I think, Look, that's why I kind of wanted to get into Harvey's.
- 41:16SPEAKER_01: So I like to look at success cases.
- 41:17SPEAKER_01: And I really screwed up.
- 41:20SPEAKER_01: I should have reached out to these guys.
- 41:21SPEAKER_01: Like, I'd love to know more about what they're doing.
- 41:23SPEAKER_01: Because they launched early.
- 41:24SPEAKER_01: They were, you know, one of many.
- 41:25SPEAKER_01: But they were prominent from the beginning, like SaaS thing.
- 41:28SPEAKER_01: And at first,
- 41:28SPEAKER_01: I'm sure they were doing kind of like whatever modeling rag,
- 41:31SPEAKER_01: like probably pretty crappy to begin with because everything was crappy at
- 41:34SPEAKER_01: the beginning.
- 41:35SPEAKER_01: But they've stuck with it.
- 41:35SPEAKER_01: They had a good sales funnel.
- 41:36SPEAKER_01: They signed up a lot of law firms.
- 41:38SPEAKER_01: And now at this point, you know, they're doing very well.
- 41:41SPEAKER_01: They're raising at 15 billion, which is, I think, by far the largest AI SaaS company.
- 41:47SPEAKER_01: I wouldn't even know who the second biggest would be.
- 41:50SPEAKER_01: You know,
- 41:50SPEAKER_01: and like,
- 41:50SPEAKER_01: you know,
- 41:51SPEAKER_01: they had a bit of a fall in the middle,
- 41:52SPEAKER_01: at least vibes wise, where everyone's like, yeah, just use Claude.
- 41:54SPEAKER_01: But like, no one says that anymore.
- 41:57SPEAKER_01: Claws got into coding.
- 41:59SPEAKER_01: They built stuff that's optimized for Nexus Nexus or whatever.
- 42:02SPEAKER_01: Obviously, lawyers are also more willing to pay for obvious reasons.
- 42:04SPEAKER_01: They need search.
- 42:05SPEAKER_01: They need to look at things all the time.
- 42:07SPEAKER_01: And also now they're post-training their own models.
- 42:09SPEAKER_01: They're not necessarily to remove their dependence on Claude,
- 42:12SPEAKER_01: but kind of for the same reason I am.
- 42:14SPEAKER_01: Like you get into these specialty tasks where
- 42:16SPEAKER_01: a fine-tuned open weights model is going to do a better job or
- 42:20SPEAKER_01: at least more consistent than sort of Claude,
- 42:21SPEAKER_01: which is post-trained for many, many other things, but not for your task.
- 42:26SPEAKER_02: Yeah.
- 42:26SPEAKER_02: Well, real quick, I want to...
- 42:29SPEAKER_02: SaaS software as a service.
- 42:33SPEAKER_02: One thing that I've heard people say,
- 42:35SPEAKER_02: and to me it's plausible,
- 42:36SPEAKER_02: is software as a service is really the place where the AI revolution in terms
- 42:43SPEAKER_02: of AI-powered software is going to be hitting.
- 42:49SPEAKER_02: With SaaS, I feel like SaaS is pretty rationalized.
- 42:53SPEAKER_02: They do very specific things, modular things, in-sequence.
- 42:57SPEAKER_02: Often there's not a website, you know, like it's just like, that's not relevant.
- 43:00SPEAKER_02: Like, you know, the human, it's like basically you have something to do.
- 43:03SPEAKER_02: It's a, you know, ERP, like all these like acronyms, you know, CRMs, all these other things.
- 43:08SPEAKER_02: But it's like a task needs to get done.
- 43:10SPEAKER_02: You don't care how the sausage is made.
- 43:12SPEAKER_02: You don't care about the form factor or the way the website looks or whatever.
- 43:17SPEAKER_02: I feel like that is really optimal.
- 43:19SPEAKER_01: Well, you'd want to manage the futures.
- 43:21SPEAKER_01: I mean, the SaaS thing that most people have used is Slack, right?
- 43:24SPEAKER_01: Slack has all these integrations, all this different shit.
- 43:27SPEAKER_01: And you're not sitting there upgrading.
- 43:31SPEAKER_01: You're not thinking, ooh, how can we extend it?
- 43:33SPEAKER_01: I mean, if you're using something like...
- 43:36SPEAKER_01: If you're using some of these ERPs or other record-keeping systems,
- 43:42SPEAKER_01: then the big money was in internal upgrades.
- 43:44SPEAKER_01: Everyone built their own shit on top of it.
- 43:46SPEAKER_01: But for some stuff, you don't even think about it.
- 43:49SPEAKER_01: It's like, okay, cool, Slack rolled in another new feature, great.
- 43:52SPEAKER_01: You're kind of depending on them to do it.
- 43:54SPEAKER_01: Yeah.
- 43:56SPEAKER_01: And that's important.
- 43:57SPEAKER_01: Like, that seems like a small thing, but it's not, right?
- 43:59SPEAKER_01: Like, when you're running your own software,
- 44:01SPEAKER_01: you're doing something like Altasian,
- 44:03SPEAKER_01: you get tied to model versions and things like this.
- 44:05SPEAKER_01: And this is getting worse.
- 44:06SPEAKER_01: So actually, one test I have, sorry, I promise this is going somewhere.
- 44:10SPEAKER_01: So one task I have to see if someone,
- 44:14SPEAKER_01: you know, is really an AI native developer or not is one simple question.
- 44:18SPEAKER_01: I'm like, okay, you're doing whatever.
- 44:19SPEAKER_01: That's great.
- 44:20SPEAKER_01: Like, how are you affected by,
- 44:22SPEAKER_01: you know,
- 44:23SPEAKER_01: model outages,
- 44:24SPEAKER_01: model swapping, new models being upgraded, you know, models being deprecated?
- 44:28SPEAKER_01: You know,
- 44:28SPEAKER_01: for example,
- 44:29SPEAKER_01: you know,
- 44:30SPEAKER_01: OpenAI announced that they were going to deprecate a GPT 5.4 from Codex,
- 44:36SPEAKER_01: which is driving my colleague crazy,
- 44:37SPEAKER_01: you know, because we have some shit that depends on it.
- 44:39SPEAKER_01: It works well.
- 44:40SPEAKER_01: And it's not that he doesn't want to move it to 5.6, but it's going to be different, you know?
- 44:44SPEAKER_01: And the way you can tell if they're really an AI native developer or not is
- 44:47SPEAKER_01: that they're like, it doesn't affect me, then they're not.
- 44:49SPEAKER_01: You know,
- 44:49SPEAKER_01: because then they're essentially,
- 44:51SPEAKER_01: but this gets into the SaaS thing
- 44:52SPEAKER_01: where you can have something very reliable running,
- 44:54SPEAKER_01: but then you're not really taking advantage of the new shit.
- 44:58SPEAKER_01: So I think the vast, vast,
- 45:00SPEAKER_01: vast majority of the so-called AI SaaS,
- 45:02SPEAKER_01: nothing wrong with it doing what they should be,
- 45:04SPEAKER_01: but they're actually like, oh,
- 45:06SPEAKER_01: they're a lot more like using AI to like write some code and then like you own
- 45:10SPEAKER_01: the code, but then you have to maintain it.
- 45:11SPEAKER_01: You have to do stuff.
- 45:12SPEAKER_01: That's very, very different.
- 45:13SPEAKER_01: Then actually having a harness and calling the AI live for when the user does shit,
- 45:18SPEAKER_01: which is like, like,
- 45:20SPEAKER_01: like even something like Harvey is like somewhere in the middle,
- 45:23SPEAKER_01: but like,
- 45:23SPEAKER_01: let's just say mostly they're doing document drafting,
- 45:25SPEAKER_01: you know, search, blah, blah, blah.
- 45:27SPEAKER_01: Like that stuff is still a kind of like version, right?
- 45:30SPEAKER_01: Whereas you can,
- 45:30SPEAKER_01: you know,
- 45:30SPEAKER_01: what you're talking about in the future,
- 45:32SPEAKER_01: which we don't really have is like,
- 45:33SPEAKER_01: imagine it actually calls models live for like the hardcore dependence.
- 45:37SPEAKER_01: Yeah.
- 46:06SPEAKER_01: So we're nowhere close to that.
- 46:08SPEAKER_01: No one's doing that, really, realistically.
- 46:11SPEAKER_01: They're still mostly using AI to write code,
- 46:14SPEAKER_01: and that code runs reliably,
- 46:16SPEAKER_01: which is very stable, but isn't really taking advantage of the intelligence.
- 46:19SPEAKER_01: You're kind of taking advantage of it directly, right?
- 46:22SPEAKER_02: Yeah, really quickly, I want you to talk about what a harness is and why they're important.
- 46:27SPEAKER_02: People are talking about harnesses now.
- 46:29SPEAKER_02: Anyway, I saw something where there was a harness to...
- 46:34SPEAKER_02: The Hadith literature on like some ads.
- 46:37SPEAKER_02: I was like, harnesses are getting big.
- 46:39SPEAKER_02: So talk about what a harness is so that people know real quick.
- 46:42SPEAKER_01: Yeah, sure, sure, sure.
- 46:43SPEAKER_01: So let's see.
- 46:45SPEAKER_01: There's like two ways to sort of describe it, I think, in terms people understand.
- 46:48SPEAKER_01: So one thing is if people sort of remember the open claw, sort of like craze.
- 46:54SPEAKER_01: Which is now completely dead.
- 46:56SPEAKER_01: But Hermes Agent, I think is pretty popular.
- 46:58SPEAKER_01: That's like one way of thinking of a harness.
- 47:00SPEAKER_01: It's like, hey, I'm not going to call the raw model.
- 47:02SPEAKER_01: I'm going to build up these like repeated tasks, these skills, these things.
- 47:06SPEAKER_01: You can just like give it one command and it does things sort of for you, you know?
- 47:10SPEAKER_01: So in the case of code, Coding Carnage is obviously the biggest thing in the case of code.
- 47:14SPEAKER_01: I'm like, hey, update this data.
- 47:15SPEAKER_01: You already know how to do it.
- 47:17SPEAKER_01: Update this data.
- 47:18SPEAKER_01: Oh, runs a thing, runs a unit test, tells you what's up, right?
- 47:21SPEAKER_01: You know, you can think of it for non-coding things like, again, like a news update.
- 47:25SPEAKER_01: I would be like, hey, like, Claude, like, give me a, like, Tell me what's up in the morning.
- 47:30SPEAKER_01: So it goes and takes all the stuff from the new stuff that we generated.
- 47:35SPEAKER_01: It figures out what's novel and what's a follow-up.
- 47:38SPEAKER_01: It puts into categories.
- 47:39SPEAKER_01: It knows what I really like.
- 47:41SPEAKER_01: So in a sense,
- 47:42SPEAKER_01: the harness is the part that essentially expands your very small command into
- 47:48SPEAKER_01: a bigger set of actual specific commands to call the AI following up on those.
- 47:54SPEAKER_01: So there's always a fine line between...
- 47:56SPEAKER_01: Now, the models themselves are thinking models, so they do their own reflection.
- 48:01SPEAKER_01: So a task can be sort of done in both ways,
- 48:03SPEAKER_01: and there's a little bit of a vague distinction between them,
- 48:07SPEAKER_01: but it's mostly pretty clear, right?
- 48:09SPEAKER_01: Like, the harness is...
- 48:10SPEAKER_01: Not doing the thinking, but it's sort of doing the communication with the LLM for you.
- 48:17SPEAKER_01: Another analogy would be,
- 48:20SPEAKER_01: say you're at a Vegas hotel,
- 48:22SPEAKER_01: like a very nice thing,
- 48:23SPEAKER_01: and you have a butler,
- 48:24SPEAKER_01: or you have service people in general if you're not that cool.
- 48:28SPEAKER_01: Again, you give very simple commands,
- 48:30SPEAKER_01: or maybe you don't have to say much, and things just sort of happen.
- 48:33SPEAKER_01: But someone has to orchestrate it actually like,
- 48:35SPEAKER_01: You know,
- 48:35SPEAKER_01: press the button on the orange squeezing machine and bring you the drink, you know?
- 48:39SPEAKER_01: So I find that to be like, you know, kind of the best way of describing a harness.
- 48:45SPEAKER_01: So again,
- 48:45SPEAKER_01: a harness is not a model,
- 48:46SPEAKER_01: but it's a thing that calls models on your behalf and sort
- 48:50SPEAKER_01: of like remembers your shit.
- 48:52SPEAKER_02: Well, I mean, you know, the word harness itself, you know, you think about like what?
- 48:56SPEAKER_02: Horse.
- 48:56SPEAKER_02: You know, the horse.
- 48:58SPEAKER_01: Yeah.
- 48:58SPEAKER_02: And then you have a harness to the horse.
- 49:00SPEAKER_02: You can just like control the horse directly,
- 49:01SPEAKER_02: like grab its hair, like trying to, you know what I'm saying?
- 49:04SPEAKER_02: But like harnesses are just like much, they make it much more efficient and effective.
- 49:08SPEAKER_02: You know, everyone uses a harness.
- 49:10SPEAKER_01: Or like thinking of driving a car, like driving a car, you're using a harness, right?
- 49:14SPEAKER_02: Like it's doing all these commands and power
- 49:16SPEAKER_01: braking and steering and blah, blah, blah.
- 49:19SPEAKER_01: And like your commands are pretty simple, right?
- 49:22SPEAKER_02: Yeah, yeah, yeah.
- 49:23SPEAKER_02: So, I mean,
- 49:25SPEAKER_02: in terms of, like,
- 49:27SPEAKER_02: AI advances,
- 49:27SPEAKER_02: I actually, like,
- 49:28SPEAKER_02: as you were talking there,
- 49:30SPEAKER_02: I decided to do my usual,
- 49:32SPEAKER_02: which I haven't been doing that much recently,
- 49:34SPEAKER_02: write an essay like Razib Khan, 1000 Words.
- 49:37SPEAKER_02: ChatGPT actually had a lot more prompts now, and it wanted more specificity.
- 49:42SPEAKER_02: But I have to say, the writing...
- 49:44SPEAKER_02: I think it's,
- 49:45SPEAKER_02: I feel like at least for the, okay,
- 49:48SPEAKER_02: like the 200 a month,
- 49:50SPEAKER_02: I'm assuming like,
- 49:51SPEAKER_02: you know, ChatGPT, a lot more people should probably be using that sort of stuff.
- 49:54SPEAKER_02: You know,
- 49:54SPEAKER_02: these 200 month like plans,
- 49:56SPEAKER_02: these professional like research grade plans are really good, I've heard.
- 49:58SPEAKER_01: But, you know, the general pro plan,
- 50:01SPEAKER_02: it still is producing mediocre but serviceable text.
- 50:05SPEAKER_01: It hasn't really gotten any better.
- 50:06SPEAKER_01: The writing has gotten better, frankly, since like 03.
- 50:10SPEAKER_01: We talked about this last time, but in some ways it's, It's gotten worse for two reasons.
- 50:15SPEAKER_01: I mean, the main fundamental reason,
- 50:16SPEAKER_01: again, is that this is just not that important to the companies.
- 50:19SPEAKER_01: Like, it's not that they're, like, robbing you of capability.
- 50:22SPEAKER_01: Like, it just obviously doesn't have it,
- 50:23SPEAKER_01: but they're optimizing for coding, for accuracy, for cyber stuff.
- 50:29SPEAKER_01: They're just not as focused on writing quality and especially on customizing the writer quality.
- 50:33SPEAKER_01: The communities to care about a lot seven years ago, now they kind of frankly don't.
- 50:37SPEAKER_01: But the other more specific technical reason is the more these models do thinking
- 50:42SPEAKER_01: and especially like RL training,
- 50:44SPEAKER_01: like kind of stuff, like you're baking in these things.
- 50:47SPEAKER_01: It's like really fucking hard to get
- 50:48SPEAKER_00: rid of.
- 50:49SPEAKER_01: So the most obvious is Claude.
- 50:50SPEAKER_01: And it drives me crazy.
- 50:52SPEAKER_01: But I know from Twitter that it drives a lot of other people crazy.
- 50:54SPEAKER_01: The way it's like honest answer, you know, like here's the caveat, you know.
- 50:58SPEAKER_01: And I'm like, there's nothing you can do on the prompt level to remove that.
- 51:01SPEAKER_01: You can tell it as much as you want.
- 51:04SPEAKER_01: You can have it like the pre-hooks, whatever.
- 51:05SPEAKER_01: I've tried all of it because it's interesting to me.
- 51:07SPEAKER_01: You cannot get rid of that at a clock.
- 51:09SPEAKER_01: Why?
- 51:09SPEAKER_01: Why?
- 51:10SPEAKER_01: People say it's a system prompt.
- 51:11SPEAKER_01: Maybe.
- 51:12SPEAKER_01: That's part of it for sure.
- 51:14SPEAKER_01: The other thing,
- 51:14SPEAKER_01: of course,
- 51:15SPEAKER_01: is that it treats everything you say as a suggestion and kind of yada yada,
- 51:18SPEAKER_01: which is a separate issue.
- 51:19SPEAKER_01: But the main thing is the model has obviously been,
- 51:22SPEAKER_01: you know, it doesn't come out from pre-training on the web to do that.
- 51:25SPEAKER_01: Like it's been post-trained to do that.
- 51:26SPEAKER_01: You know,
- 51:27SPEAKER_01: it's been heavily RL'd to have this sort of like style that
- 51:31SPEAKER_01: for whatever reason they find acceptable.
- 51:33SPEAKER_01: And, you know, it's baked in.
- 51:36SPEAKER_01: So, so, so I think they've baked in a lot of bad habits and they haven't baked
- 51:39SPEAKER_01: in like good ones.
- 51:40SPEAKER_01: And they certainly,
- 51:41SPEAKER_01: certainly,
- 51:42SPEAKER_01: certainly,
- 51:42SPEAKER_01: certainly have not baked in the ability to,
- 51:45SPEAKER_01: you know,
- 51:45SPEAKER_01: research your style,
- 51:46SPEAKER_01: clone your style, be like, you know, you're like a great example.
- 51:49SPEAKER_01: You've written so much on the internet.
- 51:50SPEAKER_01: You can download all of it, give it to the model, have it like learn your tendencies.
- 51:55SPEAKER_01: Like the model would actually be quite decent.
- 51:56SPEAKER_01: I could do that for you and you'd be surprised how well it would do.
- 51:59SPEAKER_01: Not only that, I could get both Claude and And OpenAI Codex to give really good critiques.
- 52:06SPEAKER_01: Like you can actually write the essays,
- 52:08SPEAKER_01: especially it's better at critiquing the other model, which kind of makes sense.
- 52:11SPEAKER_01: It'll tell you what it's doing wrong.
- 52:12SPEAKER_01: It'll tell you, oh, it's not Razib for this reason.
- 52:14SPEAKER_01: Razib would have said this.
- 52:15SPEAKER_01: Like it can actually do the criticism,
- 52:18SPEAKER_01: but the actual writing and the full thing, it's fucking terrible at.
- 52:21SPEAKER_02: Yeah.
- 52:22SPEAKER_02: Yeah.
- 52:23SPEAKER_02: Okay.
- 52:23SPEAKER_02: So what else, what else is like on your agenda?
- 52:26SPEAKER_02: Like we're coming up on an hour.
- 52:27SPEAKER_02: So like,
- 52:27SPEAKER_02: you know,
- 52:27SPEAKER_02: is there any like big low hanging fruit that you as an AI guy see
- 52:31SPEAKER_02: that we should know about?
- 52:33SPEAKER_01: Yeah.
- 52:34SPEAKER_01: I mean, I think the, let's see.
- 52:37SPEAKER_01: I think,
- 52:37SPEAKER_01: I think the other big development that we haven't really touched on too much is,
- 52:41SPEAKER_01: you know, some of the,
- 52:45SPEAKER_01: Some of the pushback to data centers kind of stuff, I think.
- 52:48SPEAKER_01: Oh, yeah.
- 53:03SPEAKER_01: Yeah.
- 53:04SPEAKER_01: Well, it looks like it's shutting down new data centers, but it's not.
- 53:07SPEAKER_01: It's simply saying that they're not going to be approved to sign up to the grid.
- 53:13SPEAKER_01: So they can't connect new data centers to the grid.
- 53:15SPEAKER_01: By the way,
- 53:16SPEAKER_01: very big win for the data centers already got the permission
- 53:19SPEAKER_01: because you have this cliff.
- 53:21SPEAKER_01: But they're basically saying, okay, all of this stuff is supposed to individual review.
- 53:25SPEAKER_01: Now this is going to be some like master review that's going to happen in a year.
- 53:28SPEAKER_01: And basically it's not going to happen.
- 53:29SPEAKER_01: So basically new data centers that haven't broken ground yet
- 53:32SPEAKER_01: or haven't been approved yet cannot be connected to the grid.
- 53:35SPEAKER_01: Now, that does not mean they can't have new data centers.
- 53:38SPEAKER_01: What it really means is that,
- 53:39SPEAKER_01: you know,
- 53:40SPEAKER_01: they either need to wait,
- 53:41SPEAKER_01: good luck with that,
- 53:42SPEAKER_01: or they need to bring their own power,
- 53:44SPEAKER_01: which is most likely what they're going to do.
- 53:46SPEAKER_02: Yeah.
- 53:47SPEAKER_02: Real quickly about Texas,
- 53:49SPEAKER_02: I will say, I don't know all the details, but I do know people that know Abbott.
- 53:56SPEAKER_02: Politicians are politicians,
- 53:57SPEAKER_02: so everything they say to you,
- 53:59SPEAKER_02: you have to take with a grain of salt,
- 54:00SPEAKER_02: because they're going to say what you want to hear to everybody in some ways.
- 54:04SPEAKER_02: I guess this is,
- 54:07SPEAKER_02: I don't know,
- 54:07SPEAKER_02: whatever,
- 54:08SPEAKER_02: I don't know if this is doing journalism here,
- 54:10SPEAKER_02: but I know what he said from my friend,
- 54:13SPEAKER_02: who I trust, is basically like, He would love there to be data centers built, but...
- 54:20SPEAKER_02: The blowback,
- 54:21SPEAKER_02: the political blowback was too much and he has to figure out a way to mollify
- 54:26SPEAKER_02: the people so that it doesn't totally get shut down as some sort of populist.
- 54:32SPEAKER_01: They're doing what they have to do.
- 54:33SPEAKER_01: It's reasonable.
- 54:34SPEAKER_01: I mean, Ben Thompson's talked about this a lot.
- 54:36SPEAKER_01: Apparently, they have a lot of data centers in Wisconsin and those are getting shut down too.
- 54:44SPEAKER_01: So you have all these states.
- 54:46SPEAKER_01: I mean, red states, blue states, all the states are...
- 54:50SPEAKER_01: You know,
- 54:50SPEAKER_01: there's a lot of energy between shutting down data centers or restricting them
- 54:53SPEAKER_01: or extracting sort of concessions out of them.
- 54:57SPEAKER_01: Sorry, go ahead.
- 54:57SPEAKER_01: Did you have something else?
- 54:59SPEAKER_01: I feel like I interrupted you.
- 55:00SPEAKER_02: No, no, no.
- 55:01SPEAKER_02: No, I think that's totally right.
- 55:04SPEAKER_02: I think the term is political economy.
- 55:06SPEAKER_02: Like, we just got to figure out.
- 55:07SPEAKER_02: Well, the Texas
- 55:08SPEAKER_01: thing is big.
- 55:08SPEAKER_01: I mean, Texas was, I think, by far the biggest.
- 55:10SPEAKER_01: I mean, it's just the biggest.
- 55:11SPEAKER_01: Well, besides Virginia.
- 55:12SPEAKER_01: Virginia's huge.
- 55:13SPEAKER_01: No, no, no.
- 55:13SPEAKER_01: I'm saying for new stuff.
- 55:14SPEAKER_01: For new stuff.
- 55:14SPEAKER_01: Yeah, yeah, yeah.
- 55:15SPEAKER_01: Texas.
- 55:15SPEAKER_01: Yeah.
- 55:15SPEAKER_01: Yeah.
- 55:36SPEAKER_01: By default, instead of them being on, they're off.
- 55:38SPEAKER_01: It is a big deal.
- 55:40SPEAKER_01: Again, if it's just a power thing that's going to be solved,
- 55:42SPEAKER_01: just build your own fucking power thing.
- 55:45SPEAKER_01: I actually have friends in the...
- 55:50SPEAKER_01: People are even repurposing used jet engines and things like that.
- 55:54SPEAKER_01: There's shit you can do, which is pretty cool.
- 55:56SPEAKER_01: But I will say this.
- 55:57SPEAKER_01: So beyond the blowback...
- 56:00SPEAKER_01: By the way, by chance, did you hear Alex Epstein on Michael Malice's show?
- 56:06SPEAKER_01: They talked about this.
- 56:07SPEAKER_02: No, I did not hear it.
- 56:08SPEAKER_02: Tell me what happened.
- 56:10SPEAKER_01: Yeah, it's really good.
- 56:11SPEAKER_01: I think you and your listeners would like it.
- 56:12SPEAKER_01: So basically, you know who Alex Epstein is, right?
- 56:14SPEAKER_02: Yeah, he's a pro-fuel guy, pro-fuel guy.
- 56:18SPEAKER_01: He's the energy abundance, human flourishing guy, right?
- 56:21SPEAKER_01: Kind of libertarian,
- 56:22SPEAKER_01: ex-libertarian,
- 56:23SPEAKER_01: comes off as a little bit autistic, but is very, very good at this stuff.
- 56:26SPEAKER_01: He's been doing this for a long ass time.
- 56:28SPEAKER_01: And he's basically pivoted into being the pro-data center guy now.
- 56:32SPEAKER_01: So he's using the same strategies and helping people like kind of,
- 56:37SPEAKER_01: you know, tell people why data centers are great.
- 56:39SPEAKER_01: And Michael Malice had like one question for him
- 56:42SPEAKER_01: that I don't think he could answer very well,
- 56:43SPEAKER_01: to be honest,
- 56:44SPEAKER_01: which is he's like,
- 56:45SPEAKER_01: but hey,
- 56:45SPEAKER_01: you know,
- 56:46SPEAKER_01: you go to Virginia,
- 56:46SPEAKER_01: you drive past these things and they're like, like, they're just ugly.
- 56:51SPEAKER_01: Like, they're not necessarily harmful, but they're not cool.
- 56:56SPEAKER_01: And then Alex is like, yeah, you can paint some mural on it.
- 57:00SPEAKER_01: But I thought that was ridiculous.
- 57:03SPEAKER_01: I'm not that into the dystopian thing, but...
- 57:07SPEAKER_01: You know, like, it's just different.
- 57:09SPEAKER_01: Like, when Elon announces, hey,
- 57:11SPEAKER_01: we're going to build this giant automated factory,
- 57:12SPEAKER_01: they make it look cool and sound interesting.
- 57:14SPEAKER_01: And obviously, the rockets are cool.
- 57:16SPEAKER_01: Like, those, for whatever reason, are cool machines.
- 57:18SPEAKER_02: Factories are cool.
- 57:19SPEAKER_02: I'll watch, you know, a two-hour
- 57:22SPEAKER_01: documentary on a factory.
- 57:23SPEAKER_01: I love that shit.
- 57:23SPEAKER_01: Earth moving, all that stuff is awesome.
- 57:25SPEAKER_01: Like, giant caterpillar machines.
- 57:26SPEAKER_01: But something about the data center is, like,
- 57:28SPEAKER_01: it's dystopian to some people,
- 57:30SPEAKER_01: but even people who are, like, very pro-technology like I am, like, I don't know.
- 57:34SPEAKER_01: There's just something a little bit weird about it.
- 57:36SPEAKER_01: And I don't know what to do.
- 57:38SPEAKER_02: Yeah.
- 57:40SPEAKER_02: Dystopian's right because Elon said the future should look like the future.
- 57:43SPEAKER_02: And some of his products are like,
- 57:45SPEAKER_02: okay, this looks like a futuristic product, like Cybertruck, you know?
- 57:47SPEAKER_02: Yes.
- 57:48SPEAKER_02: I think what the data center is,
- 57:49SPEAKER_02: it kind of looks like the future in a post-apocalyptic British science fiction film.
- 57:55SPEAKER_02: Yeah.
- 57:55SPEAKER_01: Yeah.
- 57:56SPEAKER_02: It's just the vibes,
- 57:58SPEAKER_02: the vibes are just like negative,
- 58:00SPEAKER_02: I think for a lot of people,
- 58:02SPEAKER_02: just on an instinctive level, so they gotta figure something out about that.
- 58:05SPEAKER_02: I don't know, murals, I don't know about murals, but like.
- 58:08SPEAKER_01: I think it's literally like lipstick on a pig, I think it's.
- 58:11SPEAKER_02: Here's the thing.
- 58:13SPEAKER_02: I'm just brainstorming here, but data centers, they're big business.
- 58:16SPEAKER_02: They're investing all this money into it.
- 58:19SPEAKER_02: One reason that modernist architecture,
- 58:21SPEAKER_02: these big box form factors became popular is they are a little bit more efficient.
- 58:26SPEAKER_02: They are a little bit cheaper in terms of the usage of products.
- 58:29SPEAKER_02: There's no Art Deco style or earlier flourishes or anything like that.
- 58:34SPEAKER_02: If it's a marginal increase,
- 58:36SPEAKER_02: but it would make people less skeeved out by it,
- 58:39SPEAKER_02: it might be worth it to make data centers kind of look like neoclassical
- 58:43SPEAKER_02: or some sort of architectural monument
- 58:49SPEAKER_02: that actually evokes positive feelings instead of like, oh, I agree.
- 58:56SPEAKER_01: I'm not sure you can design it.
- 58:57SPEAKER_01: Like, for example, okay, like Costco, right?
- 59:01SPEAKER_01: Very much a warehouse, but like pretty cool.
- 59:03SPEAKER_01: People love Costco and not just because it's cheap, like it's actually a vibe.
- 59:06SPEAKER_01: So the other thing that came to mind,
- 59:08SPEAKER_01: and this is just a comment,
- 59:09SPEAKER_01: not really a suggestion,
- 59:10SPEAKER_01: but if you sort of,
- 59:12SPEAKER_01: you know,
- 59:13SPEAKER_01: if you go and you watch,
- 59:14SPEAKER_01: you look at any sort of post-impressionist show,
- 59:16SPEAKER_01: You know, you'll see, like, Monet and those guys painting the lilies.
- 59:20SPEAKER_01: But there's actually, like,
- 59:21SPEAKER_01: a shocking number of,
- 59:22SPEAKER_01: like, paintings of them painting, like, trains and smokestacks.
- 59:25SPEAKER_01: And some of it, like,
- 59:26SPEAKER_01: doesn't look even that cool because it's just, like, brown and there's a smokestack.
- 59:30SPEAKER_01: But, like, the trains were super cool.
- 59:31SPEAKER_01: They were, like, obviously impressed with the trains, you know?
- 59:34SPEAKER_01: And then if you look at, like, whatever the Italian futurists, right?
- 59:37SPEAKER_01: Like, they did that.
- 59:38SPEAKER_01: They did that.
- 59:38SPEAKER_01: Oh, yeah.
- 59:38SPEAKER_01: And they got more and more militaristic.
- 59:40SPEAKER_01: And, you know, I mean, well, we're not getting into that.
- 59:43SPEAKER_02: Yeah, yeah, yeah.
- 59:43SPEAKER_02: That's a little obscure.
- 59:44SPEAKER_02: But, yeah, but, yeah,
- 59:45SPEAKER_02: the point is, like,
- 59:47SPEAKER_02: The technology and art,
- 59:49SPEAKER_02: this is the whole thing,
- 59:51SPEAKER_02: like Steam and merging,
- 59:53SPEAKER_02: but we can say that in a practical way to serve the economy,
- 59:57SPEAKER_02: because I am very pro to data center.
- 1:00:01SPEAKER_02: It's just the future and you have to grasp the future.
- 1:00:03SPEAKER_02: But a lot of people are recoiling against it.
- 1:00:05SPEAKER_02: Well, maybe the people who are less averse to it are like,
- 1:00:09SPEAKER_02: well, we still have it in somebody else's county.
- 1:00:12SPEAKER_02: Whereas some people are just like, shut it all down.
- 1:00:14SPEAKER_02: There's some left-wing Luddites that are like,
- 1:00:16SPEAKER_02: You know, just have a total moratorium on AI, just shut it all down, you know?
- 1:00:20SPEAKER_01: Yeah, we're not going to reach those people.
- 1:00:21SPEAKER_01: But I guess that's the thing about it is,
- 1:00:25SPEAKER_01: I don't know,
- 1:00:25SPEAKER_01: it just made me think like most of these technology things,
- 1:00:28SPEAKER_01: there was something cool and compelling about it.
- 1:00:31SPEAKER_01: And by the way, data centers used to be interesting too.
- 1:00:34SPEAKER_01: So like I remember,
- 1:00:36SPEAKER_01: you know,
- 1:00:36SPEAKER_01: 20 years, no, over 20 years ago when I interviewed in Google out of college.
- 1:00:40SPEAKER_01: And I remember you still have to fly out there and you have like,
- 1:00:42SPEAKER_01: you know, the dinner and drinks the night before with Google engineers.
- 1:00:46SPEAKER_01: And I'm like, wow,
- 1:00:46SPEAKER_01: is it true that,
- 1:00:47SPEAKER_01: you know, Google has like the biggest data center in the world is bigger than it is?
- 1:00:50SPEAKER_01: Like, yeah, it was super cool.
- 1:00:52SPEAKER_01: Even years after that,
- 1:00:52SPEAKER_01: when Facebook was expanding,
- 1:00:53SPEAKER_01: you know,
- 1:00:54SPEAKER_01: I remember when they're building things in Oregon,
- 1:00:55SPEAKER_01: just the way that they were doing the efficiency and sort of,
- 1:00:58SPEAKER_01: you know, taking advantage of the climate and attaching the power plant.
- 1:01:00SPEAKER_01: Like there was something I think actually incredibly cool about data centers 20 years ago.
- 1:01:05SPEAKER_01: But that's not the vibe these days.
- 1:01:07SPEAKER_01: And I think just telling people,
- 1:01:10SPEAKER_01: hey, yeah, it's bad and ugly and noisy and whatever, but look at all the upside.
- 1:01:15SPEAKER_01: It does these cool things.
- 1:01:17SPEAKER_01: I don't know.
- 1:01:18SPEAKER_01: Like I said, that's why Texas, I think, is interesting.
- 1:01:24SPEAKER_01: They're pro-building everything, but not data centers, apparently.
- 1:01:26SPEAKER_02: Yeah.
- 1:01:27SPEAKER_02: When you've lost Texas,
- 1:01:31SPEAKER_02: you have to do some soul searching to figure out how you
- 1:01:34SPEAKER_01: can move forward.
- 1:01:34SPEAKER_01: I don't think painting a mural, like a multicultural mural, is a thing.
- 1:01:39SPEAKER_00: I think it's pretty straightforward, to be honest.
- 1:01:41SPEAKER_01: Yeah,
- 1:01:42SPEAKER_01: it's a little cringe, too.
- 1:01:44SPEAKER_01: It's very cringe.
- 1:01:45SPEAKER_01: By the way, making it look like a Roman column would be cringe, too, I think.
- 1:01:48SPEAKER_01: I think you nailed it on the head.
- 1:01:50SPEAKER_01: Like, the Elon way it has to look like the future.
- 1:01:53SPEAKER_01: So let's figure it out.
- 1:01:55SPEAKER_02: We got some designers.
- 1:01:56SPEAKER_02: Let's go talk to Apple.
- 1:01:59SPEAKER_02: Get Apple in on this because you're sure that you won't.
- 1:02:01SPEAKER_02: Actually, Tom, I was gone, but they have new people.
- 1:02:04SPEAKER_02: Yeah, so I mean,
- 1:02:04SPEAKER_02: so I guess,
- 1:02:05SPEAKER_02: you know,
- 1:02:05SPEAKER_02: in terms of my tailing,
- 1:02:06SPEAKER_02: just talking to you, you know, and our last conversation was relatively recent.
- 1:02:09SPEAKER_02: There's a lot of stuff that's always happening.
- 1:02:12SPEAKER_02: I will say that, you know, Yeah.
- 1:02:14SPEAKER_02: Yeah.
- 1:02:15SPEAKER_02: Yeah.
- 1:02:32SPEAKER_01: Yeah, that's interesting.
- 1:02:32SPEAKER_01: And again, Demis, you know, is basically leaving DeepMind to pursue his biological passions.
- 1:02:38SPEAKER_01: So that's kind of interesting.
- 1:02:39SPEAKER_01: We're definitely going to see something there,
- 1:02:41SPEAKER_01: you know,
- 1:02:41SPEAKER_01: but like,
- 1:02:42SPEAKER_01: as you always say, you know, sometimes biology moves at the pace of biology.
- 1:02:45SPEAKER_01: So biology is always around the corner,
- 1:02:47SPEAKER_01: but, you know, it takes longer than the new model release.
- 1:02:51SPEAKER_01: Yeah, I think the I'll tell you something that I mean,
- 1:02:54SPEAKER_01: we already talked about it,
- 1:02:55SPEAKER_01: but I think I definitely something I got wrong or would adjust from
- 1:02:59SPEAKER_01: the last time we talked is, I think I was selling open source a little bit short.
- 1:03:03SPEAKER_01: And I'm obviously a huge fan of open source.
- 1:03:04SPEAKER_01: I love training open source models.
- 1:03:06SPEAKER_01: Again,
- 1:03:06SPEAKER_01: it's so amazing to be able to I see like the fucking thinking trace and be able
- 1:03:09SPEAKER_01: to like critique it and deal with it.
- 1:03:11SPEAKER_01: It's so cool.
- 1:03:11SPEAKER_01: By the way,
- 1:03:12SPEAKER_01: what's funny is you can give these thinking traces to Claude and Claude will comment
- 1:03:15SPEAKER_01: and help you, but it won't show you its own thinking.
- 1:03:17SPEAKER_01: So, you know, a little bit.
- 1:03:20SPEAKER_01: It's actually kind of surreal.
- 1:03:21SPEAKER_01: But look,
- 1:03:22SPEAKER_01: I think that as far as usage,
- 1:03:24SPEAKER_01: obviously the usage is on the Frontier models, Claude and OpenAI.
- 1:03:27SPEAKER_01: But I really think that because of these Because of just a number of things
- 1:03:32SPEAKER_01: that happened,
- 1:03:32SPEAKER_01: people are very actively beyond just the cost looking at,
- 1:03:35SPEAKER_01: you know,
- 1:03:36SPEAKER_01: open source,
- 1:03:36SPEAKER_01: open weights,
- 1:03:37SPEAKER_01: alternatives,
- 1:03:38SPEAKER_01: you know,
- 1:03:38SPEAKER_01: not necessarily fine tuning their own models,
- 1:03:40SPEAKER_01: but I'm seeing a tremendous amount of interest in post training,
- 1:03:43SPEAKER_01: kind of like Harvey style.
- 1:03:44SPEAKER_01: That's why I mentioned them.
- 1:03:46SPEAKER_01: Like, I think a lot of people are like, hey, we want to take a model.
- 1:03:49SPEAKER_01: We want to adapt it to our industry and use case.
- 1:03:53SPEAKER_01: And yeah, so I think that's kind of an exciting development.
- 1:03:57SPEAKER_02: Yeah.
- 1:03:58SPEAKER_02: I mean, one thing is,
- 1:03:59SPEAKER_02: you know,
- 1:03:59SPEAKER_02: I don't want to get into details about it because we're not economists,
- 1:04:02SPEAKER_02: but it looks like,
- 1:04:02SPEAKER_02: you know,
- 1:04:03SPEAKER_02: the data center build out, like,
- 1:04:04SPEAKER_02: you know,
- 1:04:04SPEAKER_02: we're looking at something that's like equivalent to the railroad build out.
- 1:04:07SPEAKER_02: So we are,
- 1:04:09SPEAKER_02: you know,
- 1:04:10SPEAKER_02: you know,
- 1:04:10SPEAKER_02: setting aside Yarvin definition of revolution,
- 1:04:13SPEAKER_02: like economically,
- 1:04:15SPEAKER_02: this is a big,
- 1:04:16SPEAKER_02: big capital spend that is going to affect a lot of our economy over the next decade.
- 1:04:22SPEAKER_02: One thing is,
- 1:04:23SPEAKER_02: you know,
- 1:04:24SPEAKER_02: AI has been around for a while,
- 1:04:25SPEAKER_02: but, you know, the current epoch started in 2022 and That's four years ago.
- 1:04:30SPEAKER_02: It's not that long.
- 1:04:31SPEAKER_02: Think about 1999 and the internet.
- 1:04:33SPEAKER_02: Obviously,
- 1:04:34SPEAKER_02: it had been transformative in some ways in the culture,
- 1:04:36SPEAKER_02: but the first dot-com bubble was about to burst and stuff like that.
- 1:04:42SPEAKER_02: Really, really transformed people,
- 1:04:44SPEAKER_02: kind of integrated itself into life in terms of workplaces and other things,
- 1:04:48SPEAKER_02: probably in the 2000s, probably after the year 2000.
- 1:04:51SPEAKER_02: You know, the 1990s, it was more just like hopes and ideas and dreams.
- 1:04:55SPEAKER_02: And, you know, only a minority of people were on the internet.
- 1:04:59SPEAKER_02: The speeds were slow, et cetera, et cetera.
- 1:05:00SPEAKER_02: So just using that kind of teaches us,
- 1:05:03SPEAKER_02: I think,
- 1:05:03SPEAKER_02: that we are in for really big impacts, I think, in the next five to 10 years.
- 1:05:09SPEAKER_02: But we can kind of see what direction things are going in this early phases.
- 1:05:14SPEAKER_01: That's right.
- 1:05:14SPEAKER_01: And to sort of extend your analogy, again,
- 1:05:16SPEAKER_01: comparing the internet,
- 1:05:18SPEAKER_01: you know,
- 1:05:19SPEAKER_01: on desktop,
- 1:05:19SPEAKER_01: which really affected business a lot more than people in many ways,
- 1:05:22SPEAKER_01: and then the smartphone,
- 1:05:23SPEAKER_01: which like obviously affected people personally way, way, way, way more.
- 1:05:27SPEAKER_01: If we're going to apply that analogy,
- 1:05:29SPEAKER_01: then I think this phase of AI is definitely more the business than the personal,
- 1:05:33SPEAKER_01: you know, like Yes, of course, I'm asking Chad GPT questions every day.
- 1:05:38SPEAKER_02: No kidding.
- 1:05:38SPEAKER_01: But I think even OpenAI,
- 1:05:40SPEAKER_01: who owned the consumer space, kind of pivoted away from it to do like the B2B thing.
- 1:05:44SPEAKER_01: And that makes sense, right?
- 1:05:45SPEAKER_01: Like companies are just going to be,
- 1:05:47SPEAKER_01: I think,
- 1:05:48SPEAKER_01: in the medium term,
- 1:05:49SPEAKER_01: much further ahead in making the AI core to their work than people and individuals.
- 1:05:54SPEAKER_01: And I wouldn't have not necessarily thought that two years ago,
- 1:05:56SPEAKER_01: but I think right now that's like Very, very, very.
- 1:05:59SPEAKER_01: And again, going back to the data center and all that, it's just usage, right?
- 1:06:02SPEAKER_01: How much do you use your iPhone?
- 1:06:04SPEAKER_01: I mean, tremendous amount.
- 1:06:06SPEAKER_01: I think like this is slightly mismeasurement.
- 1:06:08SPEAKER_01: I mean, I'm probably using it eight hours a day.
- 1:06:10SPEAKER_01: I think those are the numbers that I'm seeing from, you know, from.
- 1:06:13SPEAKER_01: Yeah.
- 1:06:13SPEAKER_01: Yeah.
- 1:06:14SPEAKER_02: We're always engaged.
- 1:06:15SPEAKER_02: Yeah.
- 1:06:15SPEAKER_01: And we're not that engaged with AI as individuals,
- 1:06:18SPEAKER_01: but I think if in the business process, you,
- 1:06:20SPEAKER_01: you know,
- 1:06:21SPEAKER_01: maybe we're a couple of steps away,
- 1:06:22SPEAKER_01: but I think there'll be more Harvey type integrations.
- 1:06:25SPEAKER_01: Again, I'm not trying to oversell them,
- 1:06:26SPEAKER_01: just sort of as an example of,
- 1:06:28SPEAKER_01: you know,
- 1:06:28SPEAKER_01: you're natively going to use not just something code written by AI,
- 1:06:33SPEAKER_01: but like actually use the AI,
- 1:06:35SPEAKER_01: some custom thing, you know, for your biology work, for example, right?
- 1:06:38SPEAKER_01: Like maybe that's going to be Maybe that'll be a lot more.
- 1:06:42SPEAKER_02: Yeah, actually,
- 1:06:43SPEAKER_02: I don't want to get into a detail,
- 1:06:45SPEAKER_02: but I actually had to, well,
- 1:06:48SPEAKER_02: I can say who it is, just so some of the listeners, old-time listeners will know.
- 1:06:51SPEAKER_02: Vagish Darsimhan, he's a geneticist at UT.
- 1:06:54SPEAKER_02: He was at David Reich's lab.
- 1:06:56SPEAKER_02: I had lunch with him,
- 1:06:57SPEAKER_02: actually,
- 1:06:57SPEAKER_02: and we talked about how AI has transformed the bioinformatic pipeline work.
- 1:07:05SPEAKER_01: That's right.
- 1:07:05SPEAKER_01: In terms of so many... Well, also, some of it was so bad.
- 1:07:08SPEAKER_01: The code was, like, so mediocre that it was, like, ripe to be disrupted, which is great.
- 1:07:12SPEAKER_01: You know, everyone always complained about it.
- 1:07:13SPEAKER_01: Now, cool, you can build your own, like, whatever.
- 1:07:16SPEAKER_01: Yeah, sorry, go ahead.
- 1:07:17SPEAKER_02: No, but the point is, like,
- 1:07:18SPEAKER_02: you know,
- 1:07:19SPEAKER_02: there would be, like,
- 1:07:19SPEAKER_02: you'd have, like,
- 1:07:20SPEAKER_02: you know,
- 1:07:20SPEAKER_02: you could have, like, two, three, four,
- 1:07:21SPEAKER_02: five people on staff in, like,
- 1:07:22SPEAKER_02: a medium-sized laboratory just, like,
- 1:07:24SPEAKER_02: working on these pipelines and keeping them working and scaling them up
- 1:07:28SPEAKER_02: and modifying them and, like, you know, adding new data formats.
- 1:07:31SPEAKER_02: But now it's just, It's like, you know, maybe one person can do all of that.
- 1:07:37SPEAKER_02: Just kind of like, you know, high volume, like vibe coding can really can handle a lot of this.
- 1:07:42SPEAKER_02: And so, you know, hopefully we'll do more biology.
- 1:07:44SPEAKER_02: But yeah, there's there's a lot to be said for domain specific penetration of AI.
- 1:07:49SPEAKER_01: Let me extend that analogy.
- 1:07:51SPEAKER_01: And like, yeah, my information of biology is not as deep as yours, but like.
- 1:07:55SPEAKER_01: I agree with what you're saying, but imagine the next step of it, or two steps ahead.
- 1:07:59SPEAKER_01: Now imagine is, okay,
- 1:08:01SPEAKER_01: you're researching some genetic disease,
- 1:08:03SPEAKER_01: or you're trying to make connections between populations,
- 1:08:06SPEAKER_01: or you're trying to do whatever,
- 1:08:09SPEAKER_01: I don't want to say routine, but work that people do that requires some discretion.
- 1:08:13SPEAKER_01: And imagine the models were so good,
- 1:08:15SPEAKER_01: you're not just using the output,
- 1:08:16SPEAKER_01: you're not just running the things, but you're really...
- 1:08:19SPEAKER_01: Your whole process is basically a lot closer to...
- 1:08:23SPEAKER_01: You know, hey,
- 1:08:24SPEAKER_01: like make like a new like notebook,
- 1:08:26SPEAKER_01: you know,
- 1:08:27SPEAKER_01: copying previous work and immediately build a custom workflow
- 1:08:30SPEAKER_01: for this particular problem.
- 1:08:31SPEAKER_01: And like you could just have that conversation.
- 1:08:33SPEAKER_01: You can validate it.
- 1:08:34SPEAKER_01: You can do all this stuff because that's a lot closer to like what I do for my coding things.
- 1:08:38SPEAKER_01: And I think people are not really quite doing that with biology yet,
- 1:08:40SPEAKER_01: but like I can imagine that, you know?
- 1:08:44SPEAKER_01: And then you are spending,
- 1:08:45SPEAKER_01: you know,
- 1:08:46SPEAKER_01: four or six hours a day with the AI,
- 1:08:48SPEAKER_01: like running live queries,
- 1:08:49SPEAKER_01: you know,
- 1:08:50SPEAKER_01: not just taking advantage of like offline shit that it built,
- 1:08:53SPEAKER_01: like for your thinking.
- 1:08:54SPEAKER_01: And that's the final connection I wanted to make.
- 1:08:56SPEAKER_01: Like, I think we're quite far from it.
- 1:08:58SPEAKER_01: But you keep talking about writing.
- 1:09:00SPEAKER_01: I'm obsessed with writing as well, even though I'm not nearly as good a writer.
- 1:09:03SPEAKER_01: But like, that's part of why I've been fine tuning.
- 1:09:05SPEAKER_01: I've been post-training these models for writing.
- 1:09:08SPEAKER_01: Right now,
- 1:09:09SPEAKER_01: All we can really do is get it to follow the rules,
- 1:09:11SPEAKER_01: not use extra hyphens, you know, get a certain vibe that's right.
- 1:09:15SPEAKER_01: That's realistically what I can do with my resources, you know.
- 1:09:17SPEAKER_01: But if you take this two or three steps further,
- 1:09:21SPEAKER_01: imagine like being able to tune the model that actually started thinking like you
- 1:09:25SPEAKER_01: or thinking in a compatible way.
- 1:09:27SPEAKER_01: Because right now the model is like,
- 1:09:29SPEAKER_01: it's good for research, but it's not really like a proper thought companion.
- 1:09:32SPEAKER_01: It's not really something you can work through problems.
- 1:09:35SPEAKER_01: You can't realistically work through an essay.
- 1:09:37SPEAKER_01: So maybe it doesn't even like Razib,
- 1:09:39SPEAKER_01: like imagine a model,
- 1:09:41SPEAKER_01: but it's going to be a fine-tuned model harness, whatever you want to call it.
- 1:09:43SPEAKER_01: It's not going to write the essay for you.
- 1:09:45SPEAKER_01: It's not going to be Razib.
- 1:09:46SPEAKER_01: I don't think I want that.
- 1:09:47SPEAKER_01: I don't think you want that.
- 1:09:47SPEAKER_01: I don't think anyone wants that.
- 1:09:49SPEAKER_01: But imagine you can talk to it and it can not just like help you do the research,
- 1:09:53SPEAKER_01: but actually make for better and more efficient and like richer writing.
- 1:09:58SPEAKER_01: Yeah,
- 1:09:59SPEAKER_02: that's a dream.
- 1:10:00SPEAKER_02: That's the dream, you know, AI for a better world.
- 1:10:03SPEAKER_01: Hundreds of cases like that,
- 1:10:04SPEAKER_01: that would be,
- 1:10:06SPEAKER_01: you know,
- 1:10:07SPEAKER_01: that's kind of,
- 1:10:07SPEAKER_01: I think,
- 1:10:08SPEAKER_01: a little bit where it feels like the industry is heading and I'm very excited about
- 1:10:11SPEAKER_01: that.
- 1:10:12SPEAKER_02: Yeah, yeah.
- 1:10:13SPEAKER_02: All right, man.
- 1:10:14SPEAKER_02: It was great catching up with you.
- 1:10:16SPEAKER_02: It seems like the rate of our conversations is increasing partly because there's
- 1:10:19SPEAKER_02: so much going on.
- 1:10:20SPEAKER_02: It's kind of like an expanding delta.
- 1:10:24SPEAKER_02: It's fanning out into all these different domains of our lives.
- 1:10:28SPEAKER_02: There's other things we could talk about, but I think we're good for now.
- 1:10:30SPEAKER_02: I will see you online.
- 1:10:32SPEAKER_02: Actually, do you have any announcements to make about your current projects, your endeavors?
- 1:10:38SPEAKER_02: Yeah, sure.
- 1:10:39SPEAKER_01: Yeah, I mean,
- 1:10:40SPEAKER_01: I kind of alluded to it,
- 1:10:41SPEAKER_01: but so there's huggingnews.com,
- 1:10:45SPEAKER_01: which is, again, like a little thing that we built for hugging face, essentially.
- 1:10:48SPEAKER_01: So it's, you know, just AI news.
- 1:10:51SPEAKER_01: So it's the news process we build,
- 1:10:52SPEAKER_01: which is built So you have the link to the full tweet,
- 1:10:54SPEAKER_01: the expansion,
- 1:10:55SPEAKER_01: but you can actually see the main point,
- 1:10:57SPEAKER_01: especially from a lot of these announcements that are too long.
- 1:11:20SPEAKER_01: That's kind of,
- 1:11:21SPEAKER_01: you know, still,
- 1:11:22SPEAKER_01: I'd say like a little bit more of a side project,
- 1:11:24SPEAKER_01: but I think a pretty compelling one.
- 1:11:26SPEAKER_01: And then also being AI native, you can also connect it to your agent.
- 1:11:29SPEAKER_01: I need to build better examples of that,
- 1:11:31SPEAKER_01: but some people have done that and they're having it,
- 1:11:33SPEAKER_01: you know, Grab the stuff every day.
- 1:11:35SPEAKER_01: Give them the stuff that they want to know when they wake up.
- 1:11:38SPEAKER_01: That sort of thing.
- 1:11:39SPEAKER_01: I have a more general one there as well.
- 1:11:40SPEAKER_01: It's linked on the site, but marketbrief.now.
- 1:11:43SPEAKER_01: I couldn't get the .com.
- 1:11:44SPEAKER_01: But this is sort of more for markets.
- 1:11:46SPEAKER_01: So you're like,
- 1:11:47SPEAKER_01: what's happening in the stock market and private markets,
- 1:11:49SPEAKER_01: you know, what's going on in Hormuz, that sort of stuff.
- 1:11:51SPEAKER_02: Okay.
- 1:11:52SPEAKER_02: Okay.
- 1:11:53SPEAKER_01: So yeah, connect your agent.
- 1:11:54SPEAKER_01: Check it out.
- 1:11:56SPEAKER_01: That's kind of, that's the pitch there.
- 1:11:59SPEAKER_02: All right.
- 1:11:59SPEAKER_02: Well, I will see you next time, man.
- 1:12:02SPEAKER_01: Awesome.
- 1:12:08SPEAKER_00: This podcast for kids.
- 1:12:38SPEAKER_00: This is my favorite podcast.