BIG IDEAS BY NEW ECONOMIES
00:37:00 · Richard Socher
[00:37:00] S0: Yeah, I wish I had a perfect answer for all industries and all software and stuff for that. But I don't think, I mean, these big companies are going to try to maximize their profits. That's their goal. And it is easier now to build something, but it's also easier to copy whatever someone's built. And now with decompiling, even other software that was compiled, like Photoshop and people used to buy, can just be decompiled, re-implemented by an AI and then boom. All the features are there and you have a much cheaper free version of it. In a weird way, I think China has shown the future there because in China, labor has been so cheap for so long that China never developed a deep SaaS ecosystem. If you are going to spend a couple million dollars on a software, you just hire a couple people for like 10,000 bucks a year and have them re-implement that software. Now the US kind of catches up to China where it's so cheap to build things, not because we have more people and they're getting paid less, but because they were so expensive, we have now AI that builds it. And there's a lot of things where that can be true for manufacturing, too. It was too expensive to do in the U .S., so a lot of manufacturing moved overseas. Now with robots it might become cheap enough to then come back to the US and you know I mean In some ways international trade is a force for peace but its also good I guess To be able to manufacture things everywhere You want them to So I don't know if I have Like Perfect answer for every startup and how they need to react to this. But I think generally, again, network effects, brand, marketing, and data modes are continuing to be powerful. And then there's also a funny mode that I would call the mode of boring. Like maybe I don't know, maybe. I need to come up with a better name, but like the boredom moat where People at Anthropic are not that excited about automating taxes. It doesn't feel like, oh, we got one step closer to AGI because we automated taxes, right? It feels cool when you can say, oh, I proved this theorem and it's unproven for a decade or longer. But it doesn't seem that cool to be like, oh, in Alabama, you're like, this income tax, blah, blah, blah. We found this reward hack niche where it's just like, solve this zero tax problem better. And so my hunch is that kind of...
00:39:31 · Richard Socher, Ollie Forsyth
The boringness of certain industries are a mo2. There's not enough people want to get into that. So that is another way to get away from these really large model companies. And then I think in some ways, if you can build in a niche that is kind of standard startup stuff, you find a niche that is too small for people to care about, but then somehow you're able to grow very quickly from that niche such that it's hard for other companies to catch up. And I think that's what is still possible in lots of different pockets of software. [00:40:06] S1: From a Geo standpoint, I feel the U .S. has always been far ahead in what AI can accomplish. Now we're obviously seeing a lot happening in China and a few other countries. When you compare China and the US together, we're going to probably get some hate in the comments, but this is okay. Who do you think is going to win in this AI race, China or the US?
00:40:29 · Richard Socher
[00:40:29] S0: I mean, it's very obvious that the US is better at coming up with truly novel ideas and really pushing the frontier. Arguably, AI is also in a state where we've developed enough things and enough innovations and ideas that Now it's just about scaling them, making them faster, better, cheaper. And that is something that China is very good at taking ideas from the West and making them faster, better, cheaper and manufacturing more quickly. And so that's one aspect of it. Of course, intelligence is a very large concept and we can go into how to define it, but like at a high level, It has many different spaces of intelligence. And it's also arguably like China is better at physical manifestations like robots than any other country right now, certainly in manufacturing them at scale. But the robots need the brain. So a lot of times right now, People feel like the brains of the robots are going to come out of Silicon Valley, but the physical robots and drones and whatnot will come out of China at scale. So there's hopefully a lot of good collaboration and practice, probably a lot of competition. And right now it seems like all the companies are moving up and down the stack. If you're a model company, you're building up the stack more and more real end-user software products on top. If you're a cloud, you want to build a model and inference. If you're an inference provider, you want to own some of your own GPUs and build your own compute substrates. The big ones like Microsoft and OpenAI and so on. Everyone is moving up and down the stack very quickly. And I think countries are going to try to do that too. And obviously China is ahead when it comes to open source. We're going to change that. I don't know when this podcast will come out, but hopefully it'll come out after We've already released a model that is optimal in its class, and it's the best model and better than any Chinese model in its class. And so we're very excited about that. But yeah, in general, China has pushed open source a lot, which is kind of similar to Hollywood. It's a lot of soft power over the world because you can basically... Nudge people in certain directions on how you want them to think. And that soft power is something that the US is currently retreating from in the world stage.
00:42:48 · Richard Socher, Ollie Forsyth
But China is like massively stepping into that role and trading is obviously continues to be a really important aspect to everything. And so I don't know if I have a definitive answer of who will win the I race. I also don't think as long as you do summing I, I don't think it's a winner take all. And certainly if there's open source, then that again equalizes a lot of the access and capabilities and potential in different regions to participate with AI. [00:43:17] S1: You just mentioned one of your latest models, what can users be expecting? [00:43:21] S0: About two weeks ago, we released Flint, which is the best coding model that people can run still on a laptop, like a beefy MacBook Pro. And we're really excited about it because it's much faster and more accurate than any other model in this cloud. Like mixture of experts that are fast and run locally on lots of different machines. And that basically enables everyone to be able to have an AI that can code all kinds of projects for them. So this just, you know, our first model, but we think it's a really important step in the direction of more people having access to this technology and have kind of unmetered and fully private AI. [00:44:03] S1: When you guys were thinking about launching Flint, there are so many coding agents already available. Do we really need another coding agent? I mean, this would have been a question you and your team would have been thinking about, right? Do we need another one? And how do you really define becoming that number one player, ideally?
00:44:22 · Richard Socher, Ollie Forsyth
[00:44:22] S0: Yeah, so you can define this based on a ton of different benchmarks and then you can kind of look at on the y-axis the average score on all the benchmarks and on the x-axis just how fast the models can actually produce something because obviously if you had a super fat You'd have to wait a year before it would program a simple app or website for you. So you want it to be fast. But the faster it is, usually the less smart it is. Local machines especially, it's such a trade-off where the top models are just not that useful yet. And we felt like we could push them just beyond that stage where they're actually useful enough to be a daily driver to be able to do a lot of things. And then, of course, because AI is code and now AI can code, this is also a really important milestone for us towards full recursive self-improving superintelligence. And this model is just the first artifact on our journey to this. And what was beautiful here is that we've done that entire model in just a few months, like three and a half months or so from scratch. We've never had a model, like never developed a model that the company just started earlier this year. And we worked on like other things before that. And so it was just in some ways a showcase also of what our Eureka machine can do and sort of our version of the full Eureka machine I described in the book. [00:45:49] S1: Wow. In three months, this is what you've created from [00:45:53] S0: scratch and so if you think about launching tools, just a curious
00:45:53 · Richard Socher, Ollie Forsyth
[00:45:53] S0: scratch and so if you think about launching tools, just a curious [00:46:09] S1: question. Does it matter? Does it have to be 100% perfect? To build this type of tool in three months is awesome, but also I'm sure there are things you still need to fix, right? Should founders just launch and build and ship fast and just put something out and see what users actually think? Can founders do that? Yeah, of course. I mean, we're doing it. [00:46:36] S0: Proof is, existence proof exists, is there. But yeah, I think... Obviously, we also have quite substantial resources. We raised $670 million. We had access to tiny amounts of compute compared to the big guys, but lots of compute compared to an individual hobbyist. And so that was important because it wasn't just us doing it. We had to have an incredible... It's a team of really, really smart researchers and engineers that we assembled at Recursive, but it's still a fairly small team. And so the majority of the inventions and so on came from them working with the Eureka machine that then automated the ideation, implementation, and validation of all these different ideas that led to this model being faster and better and still