00:00Fran, thank you very much indeed. Yes, very pleased to say I am joined by the co-founder of Recursive,
00:04Tim Rogteshul.
00:06And this is a company that is exploring a new paradigm when it comes to AI.
00:10You've raised about 650 million US dollars. You're valued at north of 4.5 billion, Tim.
00:16And you're trying to build AI systems that can self-improve without human intervention. Why?
00:24We believe that there's a future in which AI can actually be part of the discovery process.
00:31So getting to a point where AI itself can make innovations and where as you use more compute, this AI
00:38is making discoveries, starts to build on these discoveries and take part in, for example, the way we're developing AI
00:44systems.
00:44I think for us that's a very exciting future because if you look at the past, I don't know, hundreds
00:49of thousands of years, the reason we got so far as a species is because we were able to attain
00:55knowledge that then solves a lot of problems.
00:57We've been able to show some first signs of life.
01:00This morning we released a blog post where we show that we can reach a state-of-the-art in
01:05some of the benchmarks for training very small language models.
01:10We were also able to get state-of-the-art in NVIDIA's automated kernel engineering benchmarks.
01:15So these are little software programs that make basically running these AI models faster on hardware.
01:21Is it fair to describe this as a paradigm shift if you succeed in this, in self-improving AI systems?
01:28How significant would that be for the ecosystem?
01:30What would that do to the ecosystem if you succeed?
01:33Yeah, so basically we've been through different paradigm shifts in AI.
01:37We had initially, 15 years ago, used a lot of our energy, like brain power basically, to build very specific
01:47AI models where we have linguists, for example, defining specifically the features that are useful to get to certain capabilities
01:56with AI.
01:57Then deep learning happened.
01:58We were able, as a community, to get rid of a lot of this feature engineering and instead have AI
02:04models learn autonomously what are the right features to basically capture in order to make good predictions.
02:11Then a lot of people spent time, like 10 years ago, in terms of building specific neural network architectures for
02:17specific problems.
02:18Then the transformer came and all of that got generalized.
02:21So basically every time we were able to, you know, get AI to learn things by itself instead of having
02:27human strong biases built into the models, the capabilities got better.
02:32And we think the logical conclusion of that is to actually try now to automate that kind of scientific loop
02:37of ideation, implementation and validation and getting AI to autonomously build, you know, very strong AI models.
02:44What time frames are you working on to get to these points?
02:47Two years.
02:48Two years?
02:49You think you can get self-improving AI within two years?
02:51So self-improving AI, we always had versions of that in the sense that, for example, if you do reinforcement
02:56learning within an AI model, it learns from its own experience, right?
02:59It generates its own data as it's experiencing things in the environment.
03:03But I think the paradigm shift is to get AI really to, you know, in language even, come up with
03:09ideas and explanations and build upon discoveries that it's made before.
03:12So if you succeed, what happens to the valuations of OpenAI and Anthropic?
03:16I don't know what will happen to their valuations, but I think it's going to be generally, I think, for
03:21the community and I think for the scientific community, broader scientific community, a really exciting moment because now you have
03:28AI really contributing, you know, a lot to the discovery process of, you know, of inventions in scientific domains.
03:35How do you build guardrails, safety guardrails around self-improving AI?
03:40So there's multiple things one can do.
03:44First of all, we take AI safety very seriously.
03:46If you design such a system, this is like a, you know, it has to be baked into the system.
03:50I think to some extent you could argue that already when AI is not just pushing its capabilities into the
03:57model weights,
03:58but you're actually making it much more transparent in the sense that it has to actually come up with the
04:02ideas and explanations and show the empirical evidence,
04:05you already get an audit trail, right, that you can look at and we actually open source a lot of
04:09the inventions that our system made this morning so people can actually check them and build on top of them.
04:13But beyond that, we've been in the past already been able to show that you can actually use AI itself
04:22to, you know, check some of the safety constraints.
04:26So we use AI, for example, to be able in self-play come up with jailbreaks for AI, but then
04:32also train the AI to be more robust against these jailbreaks.
04:34And we could show that that actually translates also to robustness against human, you know, jailbreak attacks of these systems.
04:40So whenever you basically get a capability gain in AI, you can also try to then use that on the
04:45defense side.
04:46You're based, you have an office in San Francisco.
04:49You're a professor as well at UCL, as well as being the co-founder of Recursive.
04:54How well placed is the UK right now when it comes to AI?
04:56What are you concerned about?
04:57What are we not doing well in when it comes to supporting the AI ecosystem?
05:02I think the UK is placed really well in the sense that we have fantastic AI talent.
05:06So the reason why we started Recursive from day one in the San Francisco and in the London office is
05:11because we're able to draw a lot of talent from, you know, the ecosystem in UK.
05:17You know, DeepMind was funded here.
05:19We had, you know, Jeff Hinton starting the Gatsby unit like a few decades ago in London in the UK.
05:25So the talent density is enormous.
05:27Somebody was joking that all the AI talent, for example, in UK is basically five-minute walk away from the
05:33Shum in King's Cross.
05:34It's now called the Knowledge Quarter.
05:36So that's going really well.
05:37What I'm concerned about is a build-out of data centers, energy prices.
05:41If we really succeed at Recursive, right, at some point, it's mostly a question of, like, how much energy compute
05:46do you want to spend on solving a certain problem, a scientific problem domain, right?
05:50And then, you know, if you have very high energy prices, that becomes, you know, tricky for, I think, countries
05:56to then really, you know, take advantage of.
05:59Okay.
05:59Tim, thank you very much indeed.
06:00Yes, the energy in compute demands will...
Comments