What happens when AI starts building itself?

Richard Socher has been a major figure in AI for some time, best known for founding the chat startup You.com and, before that, for his work at Imagenet. Now, he’s joining the latest generation of AI research startups with Recursive Superintelligence, a San Francisco startup that went private on Wednesday with $650 million in funding.

Socher joins the new project with a team of renowned AI researchers, including Peter Norvig and Cresta co-founder Tim Shi. Together, they are working to create a type of iteratively self-correcting AI, which can recognize its own weaknesses and reprogram itself to fix them, without human intervention – a holy grail of modern AI research.

I spoke with him on Zoom after the launch, and dug into Recursive’s unique technical approach and why he doesn’t think of this new project as neolab, he speaks in no uncertain terms to the new generation of AI startups that prioritize research over architecture.

These interviews have been edited for length and clarity.

We hear a lot about repetition these days! It sounds like a common goal at different labs. What do you see as your unique approach?

Our unique approach is to use openness to achieve repeatable self-improvement, which no one has yet achieved. It is an impossible goal for most people. Many people already think that it happens when you do your own research. You know, you can take an AI and ask it to make something better, which could be a machine learning machine, or a letter that you write, or, you know, whatever it might be, right? But that is not repetitive self-promotion. That’s just fine.

Our main goal is to build repeatable, self-generated intelligence at scale, which means that the entire process of conceptualizing, implementing and validating research ideas will be automated.

First (it would only change) AI research ideas, eventually all kinds of ideas, even in physical fields. But it’s especially powerful when it’s AI that’s working on itself, and it’s creating a new kind of self-awareness about its mistakes.

You used the word open – does it have a technical meaning?

It does. In fact, Tim Rocktäschel, one of the co-founders, led the open and automated teams at Google DeepMind and mainly worked on the international version of Genie 3, which is the best open model. You can say it to any idea, any country, any agent, and it just creates, and it connects.

In the evolution of life, animals adapt to the environment, and then others resist this change. It’s the only system that can evolve over billions of years, and interesting things just happen, right? That’s how we made eyes in our heads.

Another example is rainbow teaming, from another paper from Tim. Have you heard of red teaming?

In cybersecurity, it means—

Therefore, red integration should also be done in LLM. Basically you’re trying to get LLM to tell you how to make a bomb, and you want to make sure it doesn’t.

Now, people can sit there for a long time and come up with interesting examples that the AI ​​shouldn’t say. But what if you test this first AI with the second AI, and the second AI now has the task of making the first AI (try) say all the bad things possible. And then they can go back and forth for millions of repetitions.

You can let two AIs take turns. One keeps fighting the other, and then comes up with not just one corner but many different ones, thus the metaphor of the rainbow. And then you can enter the first AI, and you are safe and secure. This was an idea from Tim Rocktaeschel, and it is now used in all major labs.

How do you know when it happens? I think it didn’t happen.

Some of these things will never happen. You can always be smart. You can do well in programming and math and so on. There are limits to intelligence; I’m actually trying to do it normally right now, but it’s astronomical. We are very far from that limit.

As a neolab, it feels like you get to do what the big labs aren’t doing. So part of the point here is that you don’t think the big labs will reach RSI (self-renewal) by doing what they’re doing. Is it fair to say?

I can’t really comment on what they’re doing, but I think we’re approaching it differently. We truly embrace the idea of ​​freedom, and our team is focused on that vision. And this team has been researching this and doing papers in this space for the last ten years. And this team has a history of really pushing the field forward and shipping real products. You know, Tim Shi built Cresta into a unicorn. Josh Tobin was one of the first people at OpenAI and eventually led their Codex and deep research teams.

I actually sometimes struggle a little with the group neolab. I feel like we are more than just a lab. I want us to be a reliable company, to have amazing products that people love to use, that have a positive impact on people.

So when are you planning to ship your first product?

I have thought about this a lot. This team has progressed so much, we can take it back to when we first thought. But yes, there will be sales, and you have to wait a quarter, not years.

One of the assumptions about iterative optimization is that, when we have this type of system, the computation becomes the only resource needed. The faster you run the system, the better it runs, and no outside human action can really change it. So the competition is just, how much power can we throw at this? Do you think that is the country we are going to?

The reading should not be reduced. I think in the future, the most important question will be: How much money are people willing to spend to solve problems? Here’s this cancer and here’s this virus – which one do you want to get rid of first? How much do you want to donate? It becomes a matter of distribution of resources in the end. It will be one of the biggest questions in the world.