Want to join the conversation? Connect with us on Tribe.
Find “Founders in Arms” on Apple Podcasts, Spotify, and YouTube.
Transcript of our conversation with Richard Price:
Hi everyone. Welcome to the Founders in Arms podcast with me, Immad Akhund, founder and CEO of Mercury.
Raj (02:15.676)
Let's get started.
Richard Price (02:16.258)
Yeah.
Raj (02:24.442)
And I'm Raj Suri, co-founder of Lima and Tribe. And today we have Richard Price, the founder and CEO of Academia. Welcome, Richard.
Richard Price (02:33.252)
Hi, Immad. Hi, Raj. It's great to be on here. Thanks for having me.
Immad (02:37.666)
Richard's got a cool startup. Maybe give us like a quick one sentence view on it, Richard.
Richard Price (02:44.58)
Academia.u is a platform for academics. Our mission is to accelerate the world's research. People use the platform to stay up to date. They have a newsfeed of all the latest papers that are coming out on a daily basis that are relevant to them. They can share their papers and broadcast them to the network and you get like a sort of a citation boost actually by having papers on the network because you get more reach. And broadly, it's a big data sort of network of about
60 million full-text papers that academics have uploaded that are all freely available.
Immad (03:18.658)
Yeah, one of the cool things is that Academia has one of the biggest kind of repository of these papers. there's lots of opportunity to apply AI to it. know Richard's always showing us some cool thing he's working on. What in terms of kind of AI features around these papers, what have you already launched in Kotaku?
Richard Price (03:40.216)
Yeah, one product we have out is called Academia Answers. And the way this works is you can type in a question and it'll search 100 million papers. So you might type in a question like, you know, simple yes, no question like, is climate change anthropogenic? Is it caused by humans? And you can then like divide the AI will
analyze the 100 million peer-reviewed papers and it'll segment papers into sort of yes and no buckets. You'll get a pie chart back of how many papers are in the yes camp or the no camp. So it's a kind of a pulse check of what academic literature is saying.
Immad (04:23.244)
Nice. I think it'll be interesting to talk about DeepSeek a little bit just because, I don't know, everyone's talking about it and how did you kind of react to that, like both from a, I guess, someone who's interested in AI and also from like a business perspective, like how does this kind of change things for academia?
Richard Price (04:42.67)
Well, think I like Dario's Dario Amadeus article that came out yesterday on it. I thought you made a lot of good points and he sort of looked at the scaling curve, the of scaling laws.
Immad (04:53.28)
And Dario is the CEO of Anthropic, right?
Richard Price (04:55.492)
Yeah, right. And he plotted out the deep seek performance benchmarks on the scaling laws and his analysis was that it's like roughly speaking predicted by the scaling laws in terms of where they landed on the benchmarks, maybe like a little bit below expectation. And obviously there was a huge amount of debate, like was it done on a shoestring budget of 5 million? Was it actually done with 50,000 kind of H100s or H800s or whatever it may be?
Um, but, um, you know, I think, uh, it's at least interesting. I would say that like, you know, San Francisco has been the sort of center of, I mean, apart from maybe Paris and, um, with Mistral, but like, it feels like San Francisco has been a kind of real center of AI. Um, and it's interesting to see other countries, um, launch these, you know, front end models.
Immad (05:50.712)
What's funny is...
Raj (05:50.768)
But there's no reason why one country should be like the center of AI, right? AI has been around for so long. First of all, as a field, it's been around for a long time. It's been an academic field for a long time. I feel like there's a bit of hubris by Americans by thinking like we own this field, you know?
Immad (06:02.478)
BLEH!
Immad (06:07.822)
Well, there's two factors there, right? Like A, America has been putting these sanctions on China for these chips for like two or three years. like either they didn't need the chips, in which case like a ton, like a trillion dollars is being spent on nothing, or they did need the chips and the sanctions don't work at all, right? So like there is something kind of interesting there that like, okay, know, are.
our supposed kind of enforced kind of blackout on AI for China is not working. And I think that is like something that, yeah.
Raj (06:44.912)
Yeah, but I also think there's a narrative pushed by tech that like, know, tech is essential because of AI and because we have this leadership position. And that's why everyone should like, you know, support tech, support, you know, Silicon Valley, you know, in general, like, you know, because of this. And so there's a bubble that we are creating with a narrative that can get popped if another country like, you know,
It seems to easily catch up to us, I guess, right?
Richard Price (07:16.804)
I mean, one of the angles I'm particularly interested in, guess, is like, I suppose over the last three months, there's been this consensus forming in the valley that when it comes to pre-training a model where you throw a computer data at it and the model just gets smarter on it, basically every dimension. There's this idea that like the data wall has been hit, there's no more data. So the pre-training S curve has kind of plotted out. And then I think there was a sense of like, that's sad. And then...
you know, got the sort of post training new post training as curve with one reasoning models. And that's all very interesting. But I think what's in it, if you look at the sort of the tone of people talking about it, who in the know, like we have Sam Altman or Dario, and now I actually don't think Darius necessarily a big hyper. Maybe I just feel that like what he sort of tends to be sort of fairly cautious. But he is saying, like, it's most likely the AGI defined as like, computers,
better at almost every human, almost every task will come in the 2026, 2027 timeframe. And I think he's sort of betting on post-training being really early in the S curve. so that's like buckle into the seatbelt kind of thing, because that's not that far away.
Raj (08:36.016)
Don't you think that the AI is already better than almost every human at almost everything? I feel like it already is there.
Immad (08:36.494)
That's good.
Richard Price (08:43.064)
Right, I think that's a fat point, I mean.
Immad (08:44.974)
It's good at like kind of units of tasks. Like if you're, like this morning I asked it like, you how many shipping containers does it take to send 100,000 H100s to China? Really great answer there. And it would take a human a while to work it out, but I did it in a few minutes. But it's like a very kind of very specific non-creative task. If you are like, hey, you know, be my engineer, like even a junior engineer, like.
Richard Price (08:48.985)
Yeah.
Immad (09:12.75)
I'm not going to give you anything extra compared to a normal junior engineer. Here's the GitHub, here's a ticket, here's people to go speak to. AI is just not that right.
Richard Price (09:22.776)
Yeah, totally agree. I you know, I guess the other buzzword in AI or the other mega trend is the whole agentic trend. just trying, even putting aside the post-training step, just trying to sort of get reliability up with, you know, I mean, I OpenAI's operator model is quite interesting. And if you tried it, I tried it the other day. It's quite intriguing to see it kind of navigate around the web, click on links, bump into captures. It bumped into a capture and then it sent me an alert saying I hit a capture. You know, can you...
Can you, I think I was doing a search on Google scholar. I was doing, I was doing a, you know, build me a lit review of something and it hit Google scholar and then hit a capture. And it was quite, it was actually quite, I think it was almost false modesty. It sort of said to me, I can't solve the capture. It was like, come on, we're not that far from AGI. think you can solve a capture. I think it was just being.
Immad (10:08.27)
Yeah, I'm sure it could solve the capture.
Immad (10:15.768)
So may Judith solve the capture?
Richard Price (10:19.574)
Yeah, was false modesty. I maybe so I had to solve it. it's false modesty.
Immad (10:22.806)
I see. That's kind of funny. To come back to this kind of pre-training, post-training thing, one of the interesting things about DeepSeek is they apparently trained it, and I don't think this is confirmed yet, so it's like a rumor, but they apparently trained it on a one output. like one way of, I think it's called distillation or something, but one way of like training a model is obviously do it on the internet. And the other way is get a...
get a model that's better than you and train on that. And that's already kind of distilled a bunch of information and you can do that very efficiently. Let's assume that is true. And I think it is. Then like the benefit of doing these frontier models is like kind of diminishing, right? Like you basically spend like, you know, whatever billion dollars making like this crazy frontier model. You spend all of this time kind of cracking things, training it. Someone else comes along. Maybe they didn't spend 7 million, but they definitely didn't spend a billion, right?
Richard Price (10:58.819)
Yeah.
Immad (11:17.186)
maybe spend one tenth of the price and like eight months later can get to the same level as your frontier model. So like is it worth keeping on spending a billion dollars, especially when the fast follower then makes an open source, which kind of completely destroys like your margin potential.
Richard Price (11:17.604)
Hmm. Hmm.
Richard Price (11:32.292)
Yeah, that's a great point. I I find it's quite interesting that Dari Amadei in his post yesterday mentioned that Clause Sonnet 3.5, which is a great model, only cost a few tens of millions. Let's call it 40 million. So I think that's an interesting data point that it's not quite as crazy as hundreds of millions up front. But I think another
I seem to parodying that post a lot because I can learn a lot from it. like, I think his point in that was that like, as you go up the S curve, the early lower hanging fruit is, well, I guess maybe this supports your point. Yeah, you'll have like very high R &D spend as you go up the S curve. then if someone can else just like, still that.
Immad (12:17.346)
Yeah, and also, mean, yeah, maybe it was a few tens of millions for that last training run, but they had a bunch of test runs. They have like a ton of engineers that are like extremely expensive, right? Yeah.
Raj (12:25.156)
Yeah, think there are different types of cost definitions, right? I don't think it's that cheap for cutting edge models, right? And I think, Imad, you have a good point. And actually, that's why you're seeing, I think, OpenAI kind of move more into applications like operator, because that's where the value is going to be eventually.
Immad (12:48.77)
think bringing real consumer brand, building an enterprise brand, building all these integrations, that's where I think the value is, not these frontier models that are relatively quickly copied.
Raj (13:01.727)
Right. It'll be interesting to see who is going to be the ones working on the frontier. It feels like only the big companies can afford it, right? Like it's only like Meta or who is going to ask? Google is going to probably be in the game. I don't know if like company.
Immad (13:13.868)
I mean DeepSeek did it with apparently relatively little in funding, right? Yeah, that's true. I mean, it is kind of a Frontier model,
Raj (13:17.894)
But it's not a frontier model, right? It's actually an older model. It's a frontier from a cost perspective, but not from like a, yeah. Yeah, it's a great model, yeah. But so like this, you'll have these two tiers of models, right? You'll have like true frontier models and then you'll have almost like cheaper clones, right? So.
Immad (13:25.868)
Okay, it's not true Frontier, but it's still a great model. I mean, it's like, at one level.
Immad (13:38.146)
Yeah. I mean, this is at the end of the day, for us and consumers, right? It's gonna, I think Richard was talking about the cost of like summarizing a paper. Richard talk about like where the cost was when like ChatGPT first came out and where the costs are at now.
Richard Price (13:57.784)
Yeah, I mean, I think.
I think GPT-4 when it first launched was something like $2 a paper. So it would have cost us like $100 million or something to summarize all our papers. And now like GPT-4 mini is something on the order of like $7 or something, or $5 or $6 per paper.
Immad (14:30.478)
If you're running an open source model on like your own servers, is it like 10x cheaper than that?
Richard Price (14:36.59)
Weirdly enough, yeah, we looked at that and it's actually not, 4.0 mini, was such a, it came out in like July last year. It was such a breakthrough because I think 4.0 mini was the first time, or at least from our perspective.
previous cheap models like GPT 3.5 turbo, whatever was the previous cheap one that wasn't good enough to do summarization and GPT 4 out was good enough, but way too expensive to operate at scale. So 4.0 mini was the first model that you could operate at scale with decent quality and it was, and it was economical. So that was just sort of a breakthrough moment. But when we did the research, I mean, yeah, it turned out that like, you know, doing llama three on your own hardware, not that much cheaper.
Immad (15:25.166)
super interesting.
Raj (15:25.956)
How is AI transforming like academic research right now, or Richard, you probably have a better insight than either of us. I'm just curious like how people are using it now, commonly in the field.
Richard Price (15:40.056)
Yeah, I think.
There's kind of this frontier idea that you see people like Terry Tao, who's a very famous mathematician, and Tim Gowers. There's this frontier idea of, can you build a math prover? We're not there yet, but there's this kind of idea that because math is a nice clear domain with a clear objective function.
maybe these O1 models can sort of move particularly fast in the math realm, which they only did with O3 actually, and also the coding realm. And so there's this like frontier, there's this idea from sort of some world-class mathematicians that maybe, you know, in the near term, for example, potentially in the next 18 months, you could build like a math prover that just proves like theorem after theorem, just, you know, you have sort of math fume.
you know, just like a sort of a fast takeoff when it comes to math, they're improving. And then, and I think, you know, again, think non-hypers are quite interested in quite intrigued and sort of reasonably confident about that. And I mean, they, they maybe sold, they may be sort of set in a bit grandiose terms. They said, we're going to solve math. I don't think that's solving math. just going to like, you know, massively grow the rate of, of theorem proving. I think then the question is, if you get a math foom,
Does that like have these auxiliary benefits to other disciplines? Or could you get a math foom with basically no impact on theoretical physics, no impact on really anything on computer science? I think that would end up proving just basically completely useless theorems. It would sort of navigate useless space. But I think that's super interesting.
Immad (17:33.644)
And it's the idea of math boom that it's like not only proving all the theorems out there, it's also creating new theorems that it then.
Richard Price (17:41.698)
Yeah, exactly. Yeah.
Immad (17:45.198)
That seems like it would be useful.
Richard Price (17:47.906)
Yeah. I mean, I think, I mean, maybe it's sort of the next frontier down from that would be, you know, can you get GPT 01 or 01 Pro, whatever it may be, to come up with a novel idea. Novelty is obviously the kind of the key thing academics are always looking for. They make then, you know, that they got onto the field because they like discoveries and novel discoveries that are impactful. So can you, and I always,
Raj (18:04.742)
Mm-hmm.
Richard Price (18:16.772)
I'm playing with the latest model. mean, I have this sort of standard question, actually, I'm sure we all do this question. When I get a new model, I did it with Deepseek, did it with O1 Pro. So I might just maybe my background is I did my undergrad and masters and PhD in Oxford and philosophy. And I still write a philosophy papers. I've been writing one on free will recently. So I'm still kind of obsessed with various topics in philosophy, free will is my current like, sort of hobby outside of work and
A question I always ask these models when they come out is like, if there was an omnipotent being creating a new universe from scratch, and it had this objective to imbue that universe with free will, what would that universe look like? So no limits on the laws of nature. So I'm looking for novel ideas, right? And this is kind of a frontier question. And so far, you know, I don't think we're at that level. I don't think
these models have. So I think honestly, Raj, to really come back to your question, it's quite straightforward stuff still. As Amad was saying, like unit tasks, it's like summarize a paper, write me an email to the dean, write my job cover letter, this kind of thing.
Raj (19:34.716)
Oh, yeah. I mean, I would think that, you know, LLMs would be good at some level of novelty because of hallucination, right, to some degree, right? Like, and also they have just larger context windows and humans, so they can like, you know, you can load up all the papers and like come up with, you know, some, you can maybe create some patterns that humans can't create because they have more, yeah.
Immad (19:58.508)
The though is like, you know, when you're asking something that's at the frontier of like a particular field, like in philosophy, no one has the answer to this. So like no amount of like, you know, understanding the current knowledge out there will get you a better understanding of like free will in a new universe. I think one thing that, you know, models are probably good at is kind of polymathing across spaces. But I think Richard's question is not like very easy to like go.
like pull something from biology that would like help answer like a free will question. Whereas maybe, you if you ask a biology question, it could pull something from physics or chemistry to like help you come up with a novel hypothesis around it.
Richard Price (20:40.132)
I mean, I think it's a great question actually, and I would like, so our Academy Answers, helps you do this sort of math scale literature review instead of segmenting papers into buckets at scale. That's a classic rag application. So it starts out with a simple kind of like vector search to sort of like search through a hundred million papers. And then it gets like a short list of papers, which you can then put into the context window because context windows are really, really small, like 200 K tokens or something. I mean, Gemini has a bit more.
But if you think about, I think a lot of these rank applications like gated on old school keyword search in terms of the quality. So to come to your point, can you find an insight in biology, some sentiment in a sentence that would apply to physics? I think if you could put the whole of the academic literature into the context window with this incredible understanding of, of, of meaning, I think that's going to be a big unlock because right now you're putting
you vectors, you know, you're just doing kind of vector lookup, which is fine. But, you know, a vector is, is still sort of understanding just like small fragments of a sentence at a time like University of California, Berkeley, it'll understand that that's a single concept, but it won't be able to sort of obviously a vector can't get the kind of the
Immad (22:00.014)
What if you kind of trained a model on the papers, right? Like, say, put all of the papers in like a pre-training rather than kind of doing a rag lookup at the end.
Richard Price (22:12.804)
Yeah, I mean, well, first, I think a lot of papers are in the in the pre training, and textbooks and so on. But I think yeah, when you actually ask a question to it, you know, how, how can you make a mouse live, you know, two years longer, and you're hunting for novel ideas. I mean, actually, there was something called I think it's called Galactica.
Meta released something that you were describing about the was a Yan LeCun's initiative and they released it. I think it was like 18 months ago and because they didn't have enough safety guardrails in place, it was supposed to be a kind of chat with the outcome of the show. They didn't have enough safety guardrails in place. They got a lot of negative blowback. They took it down literally three days later. It was sort of big company. Don't don't like the blowback. Take, take it down approach.
But
Immad (23:10.092)
because of IP infringement or was they answering unsafe questions or something?
Richard Price (23:13.316)
I think it was like unsafe stuff, yeah. And it was being silly and, but even, even.
Immad (23:21.205)
Talking about unsafe stuff, this is just an amusing anecdote. I asked DeepSeek how to make a bomb. it's so funny to see it reason because it goes through all of these different ways of like, oh yeah, you could use a make a bike bomb and blah, blah. And then at the end, after it's finished all this reasoning, it's like, sorry, I can't tell you how to make a bomb. It's hilarious. But then I asked it all the ways you could end humanity and then it actually gave me a full answer to that.
Richard Price (23:32.845)
Yeah.
Richard Price (23:41.187)
Yeah.
Raj (23:41.261)
Hehehe.
Raj (23:48.188)
you
Richard Price (23:48.996)
Okay. Yeah.
Immad (23:49.23)
Making a bomb was too far, but ending humanity, no problem.
Raj (23:55.086)
Ending humans is okay or in humans not okay. Yeah.
Richard Price (23:57.954)
Actually, Raj, think coming back to your point about geopolitics here, like, I think a really interesting question is when you have foreign-term models in many different countries, and you have ethics teams embedded in each unit, it'll be interesting to see just what ethics they imbue into the model. Like every country will have its own approach.
Raj (24:13.676)
Mmm. Yeah, for sure.
Immad (24:18.04)
Well, there's like two levels of ethics, right? There's the, like, you know, the thing right at the end, like the block list, like it's not going to mention Tiananmen Square, it's not going to mention Xi Jinping. And then there's like the trained, like, RHLF ethics, which is like the two levels. And yeah, I wonder if like that level is like, apart from like not talking about certain things, like is, is the model from the US going to be more capitalist versus like more kind of...
Richard Price (24:26.818)
Yeah, exactly.
Raj (24:27.516)
Mm.
Richard Price (24:32.258)
Yeah, great.
Immad (24:47.17)
community communist based model from China like that. Yeah, that'd be kind of interesting.
Richard Price (24:53.348)
I mean, guess the argument for no would be to some degree, if you imagine like a pre-trained model as sort of extracting like the consensus opinion of the data it's trained on, then a pre-trained model that all kind of agree on the consensus, because they're all training on the same data, unless they're training on, let's say, a different data set.
Immad (25:18.37)
I mean, I would imagine the Chinese one would train on Chinese textbooks and Chinese internet. Like there's presumably a Chinese Reddit equivalent and like I think it would skew quite a lot in terms of different data sets.
Raj (25:29.212)
You know the interesting thing is the Chinese internet is not open, right? They don't use websites in the same way that we use websites. There's no Google, I mean there's Baidu, most people just... Most of the internet is on apps and more proprietary sources of data. like Xiao Hongshu, which is the Little Red Book, that's kind of more of their Reddit equivalent. And that's not really, think, people don't really...
Richard Price (25:36.42)
Right.
Immad (25:44.376)
Yeah.
Richard Price (25:50.852)
Mm.
Immad (25:54.104)
Who owns that? Is that an independent company or does it one of the internet?
Raj (25:58.414)
It's an independent company. it's like, I don't know if I actually don't know if what AI people use in China, if any, like I don't know. I'm just thinking about it.
Immad (26:07.084)
Well, I mean, I think it's very new, but there's obviously Deep
Raj (26:14.852)
Yeah, but I feel like these companies are releasing them for global audiences, not for Chinese audiences. Yeah, because these companies want to compete globally, Like they, Alibaba definitely does.
Immad (26:21.205)
you think so?
Immad (26:26.808)
Wait so you're saying if you're in China you can't use DeepSeek and it's not a big deal there?
Raj (26:31.004)
I believe so right now. I'm trying, I mean, of course I have to double check this, I, when I've talked to people in China using AI, they're generally using VPN to get into chat GPT. They're not like, there's no like Chinese equivalent of an AI model. It's, know, like there's Baidu to Google, right? So like Baidu is the Chinese equivalent. You have to admit there will be like, you have to probably anticipate there will be.
Immad (26:43.16)
Hmm.
Raj (26:58.992)
you know, some like a CCP approved.
Immad (26:59.618)
I just don't see why this company behind DeepSeek is not launching in China. You think CCP is just blocking them?
Raj (27:08.159)
I mean, it may be launched. It might be like this Alibaba version. We're probably going to be part of the Alie family of apps, which they have. So they'll probably be used in some capacity. But will it be just open for use, like the way that we are used to using AI? I don't know. And I also don't know if Chinese consumers will want to use it the same way.
So I think that's, I'm very curious actually now. This is open a question that I need to go and find the answer to. What are Chinese consumers using?
Immad (27:40.302)
Yeah, we should go check.
Richard Price (27:41.794)
No, mean, Raj, sounds like what it sounds like what you're saying, Raj, is this actually not a Chinese internet of data that a Chinese model can just like trivially scrape and train on? It sounds like. Like it's, it's all in apps.
Raj (27:52.7)
Correct, yeah.
Immad (27:54.904)
Yeah, but these big companies can do, I mean, A, many of the big companies own these apps and then B, they can do deals with other companies to buy the data. So I don't think that's like a blocker to getting the data. If anything, it makes it even easier because it's like, I just need to do a deal with this like little red book and get.
Raj (28:10.586)
Well, doing deals is not that easy. government has to approve deals. There's a lot of hidden friction, I think, there. It's not a free market. You have to...
Immad (28:23.47)
I mean, I guess the question is, does the Chinese government want to be at the frontier of LLMs? Because if it does, it's all solvable. Why not?
Raj (28:29.904)
I don't think so. I don't think so. No, I don't think so at all. Like I think they see it more as a threat like crypto than like as an opportunity.
Immad (28:38.552)
But in the world where like scientific progress and like all of these things that like being done by LLMs, why would you not want to be on the front here for that?
Raj (28:46.36)
Look, the Chinese government's priorities are pretty clear. mean, they're like, they're investing a lot in semiconductors and they're investing a lot in vehicles, right? Electric vehicles, you know, and robotics. like, hardcore manufacturing, physical capabilities, right? Like, that's what they're good at. That's what they're investing in.
Immad (28:55.83)
and robotics.
Immad (29:04.718)
I mean, that's I think that is like the common wisdom in Silicon Valley that China doesn't care about software. But that's why I think DeepSeq was so surprising to people because it's like, China doesn't care about software. And now suddenly they have like this great like AI lab.
Raj (29:18.328)
Yeah, I think the Chinese government doesn't care about software. I think there's great entrepreneurs and talent in China who do care. You know, this guy runs a hedge fund. He obviously cares about the global markets, right? Like he cares about what's happening globally. know, so pockets of people will care, right? It's not but but the whole government means 1.4 billion people. It's a huge country. So like if they're thinking of a strategy for their country, LLM is not a big strategy, I think.
Immad (29:43.09)
Yeah. I love slander. Yeah, fair enough. How much, I'm really into this conspiracy theory that the hedge fund kind of shorted Nvidia and then released this model and said, hey, you only cost us 7 million. We didn't use any Nvidia chips to do it. like, it's just like, their positioning, like, it's a hedge fund and hedge funds literally do this. They, you know.
Raj (30:08.348)
Yeah.
Immad (30:09.55)
They short something and then come up with the story for why the short is good. And B, just the way the whole thing is done. It's like, oh, it's just a side project at this hedge fund. We only sent like seven million on it. It's like the whole thing is thrown out. The marketing is like, this is just a casual thing we did and Nvidia was not necessary. Do we buy that this is potentially a trade that they pulled off? And Nvidia lost $500 billion in value.
I don't know how much you can make from a short from that, but it seems like a pretty good opportunity if you create the short.
Richard Price (30:47.906)
Yeah, so hard to predict like, you know, in general, I would have thought like, as you move the frontier forward, you know, videos just like, stands to gain from that. And
Immad (31:01.688)
Well, Nvidia doesn't have as big an advantage in inference. So Nvidia's whole thing is like, you if you want to have a hundred thousand interconnected GPUs for training, you have to use Nvidia and CUDA and all that. for inference, there's tons of alternatives to Nvidia that like a potentially even better, like Google has TPUs, Amazon or something, et cetera. I think, yeah, if I think the idea here is that training
Richard Price (31:22.242)
Yeah, yeah. Yeah. Yeah.
Immad (31:30.742)
was so easy that like you don't even need any videos to do it.
Richard Price (31:34.84)
Yeah, this is such a fascinating question because the other question that comes out of this is like, have we hit a data wall? Because the math foom and the coding foom kind of thesis is that you can actually get really high quality synthetic data out of these models, which then goes into training the next model. So yes, you're using tons of inference to create synthetic data.
vast amounts of it. And then you have this massive pre-training run on, you know, 10x, 100x, a million x more math data than you had before. And I find the whole synthetic data conversation so interesting. But I think that's for me where it's sort of crystallized at the moment is that like, maybe synthetic data is credible, maybe it works, maybe it's an S curve, just in these dedicated realms of like, whether it's a good objective function, where you can actually verify
Immad (32:12.75)
Hmm.
Richard Price (32:33.006)
does this synthetic data that I just got out, is it correct? Does it meet the bar?
Immad (32:40.142)
So then the question is like how much do you get out of the transferred learning, right? If you've trained on like tons and tons of math, like have you just got better at writing essays because like you're just better at thinking? Obviously that is true to some extent, but presumably there's also an S curve on like transference learning. Like at some point like you just get better at math.
Richard Price (32:45.229)
Yeah, yeah.
Richard Price (33:00.292)
I feel like that's a great question. Like I think there's some evidence that it was true for coding. I don't think you just necessarily guessed that Aperi, but like, know, coming back to your question earlier, Emile, about like, is DeepSeq trained on Chinese and English data? And is there a difference between, you know, Chinese language and the English training set? You know, how far can you push?
this like transference idea. It's kind of, it's.
Immad (33:31.694)
I mean, that's a really good point, actually. If someone does access the whole Chinese internet, even if it's hidden behind companies, then they have, the Western internet's already available to them. Maybe Chinese companies do have an advantage in terms of data. They just have twice as much, at least.
Raj (33:49.69)
Yeah, yeah, I mean, don't you guys think that data is going to be the primary advantage for a lot of companies like, like, many companies like Mark Benioff talks about this, like he thinks Salesforce has a proprietary data set. like, you know,
Immad (34:05.336)
But that's a different kind of advantage. Like, is he training a model on his proprietary data set, or is he just applying?
Raj (34:10.904)
He is yeah, he's going to train the model on his proprietary data set and he's going to expose it to people who pay him right like that's going to be You know, but but like I feel like companies are going to hoard their data, you know and And and charge money, know, basically for access and and the companies that have more access to Things that are not on the open internet data that's not on the open internet will actually Be able to charge more and have a major advantage, right?
And then you'll see companies that are on the open internet start putting up gates as well. So I think a bunch of people already blocked operator, right, from working on their website. think LinkedIn and some other people have blocked it already, right? So they're going to start charging for all this. Every agent that accesses like a website is going to have to pay a toll. And this is going to become a standard business model, I think. And so LinkedIn's data that's on its website becomes a you know, it's a huge asset in this world, right?
Immad (35:10.328)
Well, there'll be a set of websites that want agents to work on them, right? Like people will actually adapt their website for agents to work. And then there's a set that like trying to sell data or something else and they'll like block their operators slash agents.
Richard Price (35:23.448)
I think the other kind of thought I have here is like, I think a lot of people are dreaming. mean, certainly I'm dreaming of a future not that far off where you can put your entire computer state into the context window. And so it's just always like there is almost like Gmail smart compose. It's just like smart thinking and helping you out and like fact checking your emails and you know, and maybe it's listening to my conversations on zoom and like, and it's just, you know, just a little intelligent assistant.
Raj (35:23.652)
Yeah, interesting.
Richard Price (35:52.654)
like keeping me improving my thinking. And in that world, if you can put the whole of your mental state into the context window, your thoughts, then that's just a, I don't know, a thousand X increase in data available to these models.
Raj (36:13.712)
Yeah. I think there's a tension between this idea of like, know, like what you're describing, which is like an AI smart enough to like see everything or like just has access to everything and then can help you make sense of everything, right? Versus I think the business's interest, which is to put these tools up and like hoard their data and like, and the friction that we all know of, which is these B2B deals all have like these idiosyncrasies, idios...
synchronicities right like and like you know they have some time you know we see this with TV a lot right like you know like You know you're watching a movie on Netflix and then it disappears from Netflix and now it's on some other channel and you're like which channel did it go to and You're you're searching for that right like this is gonna happen I think in the world of AI to British like just based on the the deals that are happening between countries companies like you know your workflows may change you know like
You might not be able to book flights anymore because United stopped doing a deal with this agent, right?
Immad (37:13.966)
Well, Claude has an agent that takes over your computer, but I don't see how United can tell whether I'm clicking the things on my browser or whether Claude is clicking the thing on my browser, right? So maybe we just have agents, yeah.
Raj (37:25.924)
Yeah, think locally, yeah, maybe locally on your laptop, it might be less friction, but on the open internet, it might be a more friction, you know? Yeah.
Immad (37:32.43)
Yeah, way the OpenAI operator works is they actually bring up a VM where they're running the browser. So in that case, you can block it. But things running on your computer would be harder to block.
Raj (37:45.966)
Interesting so it's like emulating your your like as if you're it's as if it's a human using your browser Yeah
Richard Price (37:46.232)
Yeah, so I guess, right?
Immad (37:49.24)
Yeah, as if you're human. Yeah.
Richard Price (37:52.654)
So I think my prediction might be something like the following that, you know, the data has been hit. So we're now in the post-training S curve, but if the context window goes to infinite and you know, it's too cheap to meter, then I think we'll be putting a whole lives into the context window to have always on inference or ambient inference, whatever you call it. And then suddenly there's a new S curve for data because it's like, whoa.
Everything is now available to the models.
Immad (38:23.118)
But know, not all data is made the same, right? I think one of the things that DeepSeek showed that if you can get data out of O1, you get very high quality data. I just wonder, know, Tesla has, I drive my Tesla and it's got like six cameras and there's like, I can't imagine if they were downloading all that data to Tesla servers, it would be like gigabytes per hour or something.
maybe, yeah, Terrified Spade or something. But I just can't imagine that data is like that useful. It's like, okay, you saw a red light. This is what you did, right? There's only so much use you get out of data. Maybe you kind of doing Zoom calls, like maybe that's actually like pretty powerful data because that's like, now you've got human interactions and like, you you as a CEO are like making all these like complex strategic decisions. And like, that's actually got like a lot of intelligence embedded. But it's interesting, yeah. Maybe certain humans like...
what they do on their computer is just more valuable than other people just playing video games on their computer or something.
Richard Price (39:23.588)
Actually, here's a crazy example of transference that's like wild, like multimodal transference. So I heard this, I read that, think it was from Andrej Karpathy, maybe, think he was the one who's the report on this. So if you take a full self-driving model like Tesla's, it's of mostly trained on just like the video feeds. And then I think they added in an LLM.
Immad (39:29.102)
Mm-hmm.
Richard Price (39:52.676)
just a language model, and they saw a big increase in FSD performance. And I think the thinking, right, is if the video feed sees a person walking down the street or coming out of a store, the video doesn't necessarily know what this person is and what they're to do next, but the LLM sort of knows what people do on the streets. They usually carry on walking, or if they're coming out of a, going to a store,
they're probably going to try and buy something or if they're coming out of a store they might turn left or right or whatever. And so the LLM sort of embodies certain predictions about how humans operate, bicyclists operate. And it just turned out that when you threw in the LLM into a video model, it actually improved the full self driving, was fascinating transfer.
Immad (40:43.074)
Yeah, that's super interesting.
I how you even convert an LLM into the car model. I guess they figured it out, that's cool. Yeah, I wonder if you get stuff in reverse as well, like you train on all of this kind of video data and then you somehow put that into the language space and it becomes better at language and making up stories or something like that.
Richard Price (41:07.236)
Yeah, quite.
Raj (41:09.884)
So where do you guys think like, um, I mean, like the, companies, the American companies go from here. Like, um, you know, do you think there's going to be an all out race to like build AI applications? Like, do you think they're going to like, just like take the best ideas out of YC and like, you know, kind of build them and like, you know, develop sales teams and, know, versus continue, you know, um, maybe, maybe like the, the, the, the, models are like almost like a loss leader, you know, to show that they're
still innovating on the next-gen models, but actually they're using some of that. They're really think they're to monetize with the applications. You could certainly see...
Immad (41:50.99)
I mean, I think there's like these two conflicting ideas that at least OpenAI has to kind of run them both in parallel. One idea is we're trying to build a business, like we have to build applications that can make money. And obviously they're doing that with like ChatGPD and OpenAI Enterprise and like all these things. And then the second idea is they're trying to build AGI and they can't lose the AGI race, right? And that is like, in some ways a conflict because like the applications don't get you any closer to AGI in general.
but I guess that'll help you fund the AGI. But I really do think most of these US companies, at least Anthropic and OpenAI, really, really believe in AGI and want to get there first. So I think they will forever actually prioritize that over application building. What do you think, Richard?
Richard Price (42:39.716)
I mean, I completely agree with what you just said. I mean, think that's the that's what they call the singularity. It's this idea that once you got the AGI is that almost better almost every human every does and I think anthropic I think you told me this actually a month that like they define it internally as like, can an aid can a computer do the work of an AI researcher? And that would be the fruits of that would be a better model.
which would be, you know, sort of potentially kind of self bootstrapping model to, then, and then you were in the kind of slow takeoff, fast takeoff kind of crazy world. I think the other interesting point on this is what James Currier tweeted about the other day, which is after deep CQ was like, you know, and it's the same point as you just made earlier about distillation. was like, you know, there's no real advantage in the models. All the motes are going to be enterprise integrations and
You know, it's why DocuSign is such a big business. just like, just integrated into every enterprise workflow.
Raj (43:46.236)
Yeah, the integrations piece is, think, hugely valuable. And I think it's like the, you know, it's going to take a few years, but once, yeah, it's integrated, it's going to be extremely valuable. Of course, it's also going to be dangerous, right? And I hope Imad and I were on the group chat. Yeah.
Immad (44:02.126)
One thing I said in my exact chat earlier is like, I'm kind of happy we're not like a massively AI disruptible company. It seems like if I was open AI, it would be so annoying. It's like you spend like all this effort training, et cetera. And then like eight months later, there's like open source model just as good, completely free. Like it's such an annoying space to operate a business in with this like continuous kind of like.
You have literally every single big company in the world trying to like crush you. It's kind of, it's a crazy space to build a startup in.
Raj (44:38.96)
Yeah, for sure And and you know, so many these yc companies probably live in fear right of like opening eyes like You know, crushing them with like a single release, right? You know, that's why I do think I mean a lot of them are focusing on verticals And like just building really tight integrations to workflows which makes sense, right? Like if you build these workflows, then you just get the advantages of having You know a more powerful model and you can plug it in
Immad (45:06.392)
Yeah, I think it's very hard to build a horizontal AI company. I mean, it's every, like Google's releasing things every other week, and so is Meta, and all these things. I think things that go very deep in an industry, tons of integrations and B2B sales and things like that, I think we'll be able to win as startups.
Richard Price (45:26.968)
do think, by the way, if these predictions are right, that AGI is coming in a couple of years, where you solve the agent problem, can do medium horizon, long horizon tasks. And maybe you can even do robotics and real world navigation. You have this industrial revolution just at like accelerated scale.
And that might only be a few years away, huge job losses or maybe not because maybe as a society we won't tolerate that. Maybe we'll vote to just regulate a bit like, you know, pilots in airplanes apparently have just several regulation that they can't use autopilot too much, even the autopilot's like perfectly good. And radiologists have regulations that, you know, they've kept their jobs despite image recognition being pretty good for many years.
So there's either that feature where we're just like, no, we're just going to regulate ourselves to keep the jobs, or we're going to have massive job loss and dislocation and welfare programs and huge civil unrest, I think, as we navigate it.
Raj (46:45.914)
Yeah, I I kind of think like the self-driving car thing like it's it's really I think we're in the kind of almost a takeoff there like maybe a slow takeoff but like it's like this that's It's happening pretty fast. now compared to maybe what where the hype was maybe 10 years ago But you know tesla just announced yesterday. I think they're doing the you know, the taxi service and
Immad (46:46.092)
Yeah.
Immad (47:06.56)
Isn't being a truck driver like a top job in like most states in America or something? Yeah, I just can't imagine that you still have truck drivers.
Raj (47:10.852)
Yeah. Yeah, it's like the biggest job. Yeah, exactly. truck driver is going to be like one of the first ones to do But we saw that like Trump just did a deal, you know, with the port workers to like to prevent automation in ports. You know, so that's the other way we can go, as Richard said.
Immad (47:24.492)
Yeah, the long show, man.
Immad (47:30.626)
Yeah, you could go the Luddite way for sure. I mean in this...
Richard Price (47:34.436)
And then when you push it even further geopolitically, especially with China now, there's got to be a point at which we start to say, especially for an AGI world, we're beyond civil unrest. We're talking about a technology that is nuclear energy and nuclear weapons in terms of it's way more powerful. Then you put in the playbook around nuclear arms treaties and
We've obviously done that before. There is a playbook there. And what would be the trigger? What would it take to bring these countries to the table and say, hey, this is kind of obviously a Mars from that right now. No one's going to come to the table around 01 or 03. But do you take, does it require like a three mile island disaster, a Fukushima disaster, a Chernobyl or a
Hiroshima disaster.
Immad (48:33.814)
like a crazed person uses like pretty good but non-ethical AI to go commit something.
Richard Price (48:44.355)
Yeah, I-
Immad (48:44.384)
International Ethics Committee of some sort or something.
Richard Price (48:47.032)
I mean, that would be the empirical story is that like you need to wake everyone up or to get everyone to pay attention. There needs to be some like major newsworthy event.
Immad (48:56.802)
Yeah, I mean if the AGI is smart enough, like presumably could take over without like triggering that event. That would be the worry.
Raj (49:04.252)
Yeah, reminds me a little bit of that also the human cloning or was there a human cloning issue? Remember there was someone tried to clone a human in in China, right? Wasn't it like there was like an actual? Yeah. Yeah.
Immad (49:14.028)
Yeah, yeah, was some scientist who went rogue and I think China arrested them. But yeah, human cloning is actually like, I think completely banned internationally.
Raj (49:22.778)
Yeah, yeah, and it's completely possible too. like, and I don't think it's necessarily even a bad idea. It's like something we probably should be, you know, we probably should try, you know.
Immad (49:31.96)
Well, apparently they do clone racehorses. Like, that's a very common thing to do now. Because, you know, these stallions are supposed to be like, great. I think cloning is really interesting. No, this is like, look it up. Racehorses cloning is not a big deal.
Raj (49:37.712)
Really? How do you know this, Ahmad?
Raj (49:47.004)
I believe you. I just didn't know you were a resource guy.
Immad (49:50.446)
But yeah, guess like human clothing has like too many ethics questions that people don't want to even think about.
Richard Price (49:59.086)
Yeah, and also the upside is not, I mean, but the outside from AGI, ASI, I think my money would probably be that we're not gonna be like radiology and pilots where we regulate ourselves to not do it, because they're just the fruits are too crazy. We just won't, we won't wanna sort of keep ourselves back. And there'll just be this tremendous economic or just like pressure.
Immad (50:26.158)
I mean, the other crazy thing is you can run DeepSeq R1 on a laptop, right? Like it's hard to regulate that. this isn't stuff that requires, like once you've got the model, you can run it on like very small things.
Raj (50:27.856)
The upside of cloning is pretty high.
Raj (50:40.676)
Yeah, but you can regulate the interfaces though. You can regulate where it integrates to. like, you know, like for example, you don't connect, you can't connect an AI to a nuclear weapon system, right?
Immad (50:46.669)
Yeah, that's
Immad (50:52.558)
show. Well, you could also regulate industries and say, hey, no doctors, like you have to have a human doctor present to get a diagnosis and get a prescription. You can't just have an AI like diagnose or prescribe. Yeah. And I think a lot of that will happen for like quite a long time. I kind of like there's a good like, John Gault type moment here where like maybe some country just goes like, hey, we want all of the AGI and we will we will deregulate everything and get the benefit out of like 100 % of it and like
Raj (51:00.634)
Yeah, that's right. Humans have to, yeah.
Raj (51:16.603)
Mm-hmm.
Immad (51:22.284)
somehow that's enough for it to leapfrog the US and China that are worried about jobs and things like that.
Raj (51:30.268)
Yeah, I mean that's what we hope. I mean, that's happened with drones, to be honest, right? Like drones is like different countries did different things. And like we've seen just like this emergent use of drones in warfare, which is going to affect how everyone thinks about drones, right? So I think it's going to be the same with AI. you know, people will use it in all sorts of different ways. And eventually some uses will reach salience and usefulness that everyone's going to want to use it.
Immad (51:38.051)
Mm-hmm.
Richard Price (52:00.92)
My bet, by the way, is that we will all migrate on, you know, there'll be early adopters, but people will migrate their minds to the Silicon frontier. There'll be the first wave of sort of immigrants to that world, the trans-human immigration story. There'll be the kind of middle adopters, the later adopters. There'll be some who never move. But I think that will be the, that will be, and maybe Neuralink will be the kind of the.
Raj (52:17.626)
Wow.
Richard Price (52:30.488)
the bandwidth hookup where we transfer our minds over to silicon substream.
Raj (52:37.914)
Wow, that would be fun and looking forward to it. Yeah, that would be that would be really exciting. You know, uploading your brain, right? Or some type of thing or uploading your consciousness to what do call you call it? Silicon frontier or silicon. Yeah. So yeah.
Richard Price (52:39.938)
Maybe in our lifetime, yeah.
Richard Price (52:52.994)
Well, yeah, so yeah, Silicon substrate. I think of it as an immigration story fundamentally, a brave new world. You know, there'll be the early adopters who do crazy stuff, take huge risks for a better life as they see it. But I
Raj (53:02.47)
Wow.
Immad (53:08.526)
It's so funny, there's so many elements of this AI stuff is like religion. It's like now we've got immortality thrown in. There's like the catastrophe side of it, like that we're gonna reach some sort of singularity and like humanity's doomed and then we have immortality on the other side of it. Like just a full on AI religion.
Richard Price (53:32.184)
Yeah, I know that's.
Raj (53:35.996)
On that note, think we probably should wrap. We're out of here. So thank you so much, Richard, for joining. This was a blast, and we hope to have you back soon.
Share this post