Richard Socher: The Biggest AI Breakthroughs Haven’t Happened Yet

Richard Socher (0:00)

If you're a multi-billionaire or a fairly standard sort of middle-class teenager in the United States, you're using the same phone. That is the magic of technology at scale.

And if you now think about what are other goods and services that currently only billionaires have access to, you realize, wow, a lot of those are bottlenecked on intelligence. And hence, as intelligence gets cheaper and cheaper and more accessible to everyone, we can all have a personal assistant, a personal tutor for our kids, a personal health care team.

LG Doucet (0:30)

What's up, everybody? It's LG Doucet here, and welcome to Milk Road Stocks, the daily market show that is clearly not ready for the next phase of AI advancement because we're still out here arguing about Neo Clouds every single week. Today is September 21st, 2026 We've spent the first phase of the AI boom, investing in the companies building intelligence. But what happens when we start investing in the discoveries that that intelligence is going to produce?

I don't just mean like software companies using AI to make their product better, faster, smarter. I mean scientific process that could redefine drug discovery, biology, digital cells, economics, neuroscience, physics, astronomy, materials and energy, all that stuff is coming. Today, we're talking to Richard Socher, CEO at Recursive, u.com, AIX Ventures, and the author of The Eureka Machine that comes out this week on just how quickly and possible this revolution to our lives may occur, and most importantly, how the hell could anybody ever try to invest in something that just sounds so big? Before we start, I want to remind you to like and subscribe to our channels, whether on iTunes, Spotify, or YouTube. It helps us out a lot, and leave a comment as well as to what you thought about the show.

Richard, welcome to Milk Road.

Richard Socher (1:39)

Hi there, nice to be here.

LG Doucet (1:40)

We're excited to have you. I want to talk to you about this book that you said you've been writing for over a year. It's called The Eureka Machine, and we have a lot to unpack today, but what is The Eureka Machine? What does that mean?

What exactly does that stand for?

Richard Socher (1:53)

The Eureka Machine is the ultimate invention that might be the last organic invention that it needs because afterwards, it can invent most of the future dimensions for us.

And so it's a system, a community of AI agents that essentially does the three steps of innovation, which is ideation, implementation and validation of novel ideas. And we'll start first applying this as Recursive to AI itself, so AI research on AI by AI. But eventually, our goal is to apply it to all kinds of other sciences, and that's kind of what I lay out in the book.

LG Doucet (2:39)

What does that mean to apply AI research to AI? I know that that's what Recursive means, but what does that actually look like? What does that progress look like? How do you measure it?

Richard Socher (2:51)

The simplest one, one really cool paper from one of our co-authors, and almost all the authors of that paper, except one are at the company too, is called the Darwin-Girdle Machine, where you have AI agents that make suggestions on how to change some part of their code. Could be in pre-training, it could be the model architecture, it could be in post-training, all kinds of different changes, and then they can just evaluate. So it's not a girdle-like provably better, but you can evaluate it on either benchmarks or rewards or inside environments, and then say, well, if it's better, or if it's not, but it was interestingly different, then I just created a child node of this parent node in these phylogenetic trees, that basically you keep building a little bit like evolution, where you combine an open-ended fashion, interestingly different ideas, and they co-evolve in a relationship of others and their environment to prove further and further.

In the simple case of AI, you can just evaluate it on some benchmarks, Sweet Bench, like for software engineering tasks, for example.

LG Doucet (4:00)

Is this also one of the things that people are scared about too? Is that you're giving AI the reins to keep training itself, and as a result, that's where something like ASI maybe has already been achieved.

No, it's hard to measure. Is that where this concern comes from? Is that we're just letting it go? Yeah.

Richard Socher (4:19)

Superintelligence, we think, is really exciting stepping stone to then automate the scientific method and create all kinds of new inventions. For us, it's less about can we replace this job, and we're not trying to work our AI on hacking other computers and so on. We're trying to get it to be good scientists, and that's not a zero-sum game compared to a lot of other things.