Sep 10, 2026 · 1h 14m · mad
When AI Improves Itself | Richard Socher (Recursive)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this in-depth interview on The MAD Podcast, AI pioneer Richard Socher explores the blueprint behind his book The Eureka Machine and his venture Recursive Superintelligence, explaining how recursive self-improvement, verifiable simulations, and automated research will revitalize scientific discovery.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 18.7% of the talking time here. How this is scored →
Matt as informed peer 5.0 Guest teaching 5.6 Guest disagreement 1.6 Matt pushing back 2.8
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:09:38 Dismissing Neolabs as unviable
Socher bluntly dismisses the Neolab category, arguing they are not real companies because they pursue abstract musings instead of building concrete products or shipping software.
Hardest push from Matt ▶ 50:32 Challenging macroeconomic simulation feasibilityTurck directly challenges Socher's thesis by arguing that real-world economies are driven by unpredictable human irrationality, greed, and fear that AI cannot easily model.
Biggest teaching moment ▶ 14:00 Explaining structural biological encoding in LLMsSocher educates the audience and host by detailing how next-token predictors implicitly learn complex 3D physical distances between amino acids in folded proteins without explicit spatial modeling.
Matt holds his own ▶ 1:07:04 Citing Recursive's 410M AWS dealTurck demonstrates deep domain and financial familiarity by confronting Socher with the exact breakdown of Recursive's funding allocation toward their AWS compute contract.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| The Slowdown of Scientific Progress and the Labyrinth of Knowledge | 5 | 5 | 1 | 2 | Turck demonstrates solid preparation by quoting Socher's book directly regarding the labyrinth of knowledge and 34,000 journals. Socher expands on the systemic causes of academic specialization and risk aversion in publishing, while Turck synthesizes the arguments cleanly. | |
| How Next-Token Prediction Unlocks Scientific Discovery | 5 | 6 | 2 | 3 | Turck asks how next-token prediction generalizes from chatbots to biological discovery and questions whether the analogy oversimplifies physics and biology. Socher acknowledges all models are wrong but explains in detail how sequences capture 3D spatial properties in folded proteins. | |
| Simulation, Verifiability, and the Limits of AI Creativity | 5 | 6 | 1 | 2 | Turck raises AlphaGo's Move 37 to probe the boundaries of intuition and AI creativity. Socher systematically frames creativity around closed-loop simulation and automated verifiability in domains like mathematics and programming. | |
| Recursive AI Roadmaps and Harnessing Hallucination in Science | 4 | 5 | 1 | 2 | Turck prompts Socher on Recursive's roadmap and highlights the book's counterintuitive thesis that hallucinations can be features in scientific ideation. Socher details how temperature adjustments enable novel molecular generation. | |
| Programming Biology, Clinical Trials, and the Societal Impact of AI | 6 | 5 | 2 | 4 | Turck challenges Socher on recurring industry claims that AI will cure cancer and notes tech's poor PR regarding job losses. Socher provides a grounded analysis dismissing the hard-takeoff theory due to physical trial constraints while outlining Jevons paradox and multi-target cancer therapeutics. | |
| The AI Economist: Multi-Agent Simulation for Policy Design | 5 | 6 | 2 | 4 | Turck questions whether chaotic human behaviors like greed and fear can realistically be simulated for macroeconomic policy. Socher recounts his Salesforce research on two-level reinforcement learning beating the Saez taxation baseline and explains the limitations of US adoption. | |
| Deconstructing The Eureka Machine's Four Core Pillars | 5 | 6 | 1 | 3 | Turck methodically walks through the four pillars of the Eureka Machine, probing whether foundation models of reality actually exist yet and questioning the compute bottlenecks. Socher details how LLMs, physical sensors, simulations, and robotics interconnect. | |
| Inside Recursive: Founding Vision, Scale, and Automated AI Research | 6 | 5 | 3 | 4 | Turck brings up Recursive's massive funding round and specific 410M compute commitment with AWS, then presses Socher on why he dismisses other neolabs. Socher responds candidly, calling out non-commercial labs for having vague ideas rather than shippable artifacts. | |
| Mapping the Spaces of Intelligence and Episode Conclusion | 4 | 6 | 1 | 1 | Turck asks how far intelligence can expand according to Socher's theoretical framework. Socher delivers an expansive monologue breaking down the 10 spaces of intelligence, physical boundaries, and the electromagnetic spectrum before concluding the interview. |
Statements from this episode (32)
Socher: Modern scientific hyper-specialization makes cross-disciplinary polymaths nearly impossible
“It is almost impossible nowadays to be the sort of general genius that can dabble in all of these different fields because each field takes years and years and years to get really deep into.”
Socher: Majority of his early 2010 neural network NLP papers were rejected
“The way careers work is you want to be kind of novel, but if you're too novel, if you're too far out there, then your papers will get rejected, and that certainly happened to me a lot in the early days, like, 2010 of neural networks for natural language proces…”
Socher: The first NIPS deep learning workshop had only 30 to 40 people
“The first sort of deep learning workshops workshop at NIPS back in the day, now NeurIPS that was basically like 30, 40 people all the now super famous folks”
Socher: AI will do for biology what calculus did for physics
“AI is kind of what Calculus did for physics. AI will do for biology in the sense that it'll help us weave back together lots of very complex pieces that build these complex systems that then have certain properties.”
Socher: Reviewers widely rejected his early prompt engineering paper DecaNLP
“My paper where I described prompt engineering, like one neural network that you can just prompt with any kind of questions called Deca NLP, the paper. That paper was wildly, widely rejected by the, like, basically all the reviewers and the area chair and so on…”
Socher: Existing AI plus more compute can cure diseases and aging
“With existing technology and giving it more data and more compute, we can already cure a lot of diseases. We can already cure many of the aspects of aging. We can already build better fusion like reactors and systems. We can already create new materials.”
Socher predicts AI will reach superhuman capabilities in any domain with simulation
“I think we can basically predict where AI will certainly have superhuman capabilities. And those are all scenarios and all domains where we can either have a simulation and or a verification tool.”
Socher says AI verifiers will change all programming and the digital economy
“All of programming will change, and that will be a major impact on the entire digital economy, which is the knowledge economy, and so on.”
Socher: Recursive will build AI-for-AI before tackling physical sciences
“So at Recursive, we are fairly sure that we have to start with AI for AI and then make it really, really good at doing research on creating better AI so that it has the equivalent of 50,000 PhDs in terms of knowledge and its own capabilities, and only then go …”
Socher: AI is nowhere near having enough training data for biology
“We're just nowhere near having enough training data for biology.”
Socher: AI hallucinations can be helpful when exploring novel proteins
“But I do think hallucinations can be also very helpful for AI when you want it to explore novel kinds of proteins.”
Socher: Grounding models in search results largely resolved factual hallucination
“And so I think the initially people thought, oh, we need neuro symbolic reasoning, blah, blah, to do all this. We just needed more examples of don't hallucinate now, like, take real facts from a search engine, and then mostly summarize those. And then those, l…”
Socher: Profluent closed multi-billion dollar contracts with Eli Lilly
“They've now closed, like, multi-billion dollar contracts with Eli Lilly at Profil and his company because they've created new kinds of proteins that are, for instance, even better than CRISPR-Cas-Nine and at gene editing and being even more specific and target…”
Socher: Physical constraints will prevent an AI hard takeoff
“And as bullish and excited as I am about AI, I'm not a believer in this crazy hard takeoff. I think, yes, things will accelerate, but there are certain things that will just require time because of physics and constraints in the real world, such as like long-t…”
Socher: AI's job impact depends on price elasticity of demand
“And my theory here after thinking about this for quite some time is largely dependent on the elasticity of the demand of the product when its prices go down.”
Turck: The AI industry has done a terrible PR job
“I think the AI industry has done a terrible PR job in general.”
Socher: Moral panic over chatbot friends mirrors early reactions to novels
“There's also some amount of moral panic about chatbot friends, similar fashion to how novels used to be a really bad thing.”
Socher: Nature and Science desk rejected AI economics papers
“When we submitted these papers and two-level reinforcement learning systems to nature and science, they just desk rejected them.”
Socher: AI Economist RL model replicated Saez taxation formula
“The SAIS formula in economics. And basically, it's beautiful math, and it shows that provably it's the optimal taxation scheme. But it's the optimal taxation scheme in a one step economy where you make one economic decision and then no other decision again. An…”
Socher: AI economic policy simulation unlikely in US for very long time
“I don't think this is very feasible in the United States for a very, very long time. It's just so much identity politics and special interest groups and how, you know, super PACs and so on are like, get funded. That is very, very unlikely to be used.”
Socher: Singapore or China probably more likely to adopt AI economic modeling
“My hunch is like Singapore or China will probably be more likely to try to use those ideas. Say, hey, we all agree, or we at least make it very clear that this is our objective function. And then, you know, we're going to really try our best to set the various…”
Socher: Chinese open source companies distilled knowledge from OpenAI and Anthropic models
“The few large closed labs, Anthropic and OpenAI, took almost everything they could from the open internet trained a model, but then the Chinese open source companies basically siphoned a lot of that knowledge out of those closed source models by distilling it.”
Socher: Investing in self-driving robotic labs will mature in two to three years
“Personally from investing perspective, I feel like it's a little bit early, but in like two to three years, I think it'll be right on time.”
Socher: Parallel Bio secured FDA approval to skip certain animal trials
“That company alone has already gotten FDA approval to skip certain animal trials.”
Socher: Compute is the biggest constraint for AI scientific machines
“Compute is the biggest constraint.”
Socher: Recursive's $410M AWS agreement will likely be its smallest compute deal
“That will probably be one of the smallest compute deals that will happen in, in our future.”
Socher: A lot of Neolab AI startups will not succeed
“We are, and I struggle with this sometimes, the Neolab category. I don't love it because I think a lot of them will not succeed.”
Socher: Recursive's early Eureka system outperforms months of human endeavor
“Our system, the sort of first instantiation of this Eureka machine and a very narrow domain can already outperform months and sometimes years of human endeavor on particular problems.”
Socher: Recursive demonstrated its system can build new CUDA kernels
“We also showed that they can build new CUDA kernels, which is very useful for faster inference.”