Bacterial infections were solved and cancer was not, Richard Socher said, and the reason is not money or headcount — it is that knowledge has fragmented into thousands of sub-fields nobody can hold two of at once. His book counts 34,000 journals.
The usual explanation for slow discovery is underfunding. Socher's is the opposite: there is more research than any human can weave back together, and that job is the one he thinks AI is actually for.
"Anything you can simulate, AI will solve."
Socher, whom the show introduced as one of the most cited researchers in AI, was chief scientist at Salesforce when his team trained the first large language model for proteins. His new company, Recursive, has eight co-founders and raised a round he put at about $670 million.
I listened to the full interview so you can skip it. 74 minutes of audio, 27 minutes of reading.
Here are the 15 arguments that matter.
👤 Guest: Richard Socher, Chief Executive Officer and co-founder of Recursive, former Chief Scientist at Salesforce, founder of the search engine you.com, and author of "The Eureka Machine"
🎙️ Host: Matt Turck, a Partner at FirstMark Capital who invests in AI and data companies
📰 Published: 10 September 2026 on the MAD Podcast feed
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 1 hr 14 min | ✅ Time saved: 47 min
Key Takeaways
Science slowed down because knowledge split into sub-fields nobody can cross, not because of funding
Bacterial infections were solved outright; cancer and viruses were not, and no new foundational physics has arrived
AI is superhuman in any domain that can be simulated or verified, and nowhere else
That is why games, math and programming fall first and cells do not
Recursive's plan is to build AI that does AI research before it touches the natural sciences
The internal target is a system with the knowledge of 50,000 PhDs
Hallucination is a feature when the job is inventing proteins rather than answering questions
The same temperature setting that produces wrong facts produces molecules nobody has made
Biology is turning into an engineering discipline, and proteins are the first programmable part
His Salesforce team's protein model led to a company whose designed editors he says beat CRISPR-Cas9 on specificity
He does not believe in a hard takeoff, because trials and physics set floors software does not have
What changed is company shape: eight drugs in late-stage trials instead of one
Whether a job survives AI depends on the price elasticity of what it makes
Illustration demand did not rise a thousandfold when the price fell a thousandfold; code demand did
His AI economist recovered a proven optimal tax formula, then beat it once time was added
He says economists missed "148 of the last 150 recessions" and the field has no benchmarks
Compute, not data, is the binding constraint on the whole program
He said his own nine-figure compute commitment will look small in hindsight
We are nowhere near the ceiling on any kind of intelligence, which is his answer to the bubble question
He defines 10 separate spaces of intelligence and says each one runs to physical limits, not engineering ones
1. Why Science Slowed Down
Turck opened on the claim that begins Socher's book — that scientific progress has slowed — and asked why, given the number of researchers and the money going in.
His evidence is a comparison between two disease classes. Socher said bacterial infections went from a death sentence to a nuisance because antibiotics genuinely solved them, and that nothing comparable has happened with viruses or cancer
He made the same point about physics. Nothing as foundational as E=mc² or general relativity has arrived since, in his account, even though those theories led directly to nuclear fission and fusion research
The mechanism he blames is fragmentation. Fields understood their foundations, moved into engineering, and then split into thousands of sub-fields: "It is almost impossible nowadays to be the sort of general genius that can dabble in all of these different fields" — each one takes years to reach depth in
The consequence is too few people per niche, which he said others in the field have predicted would be a big part of the slowdown
Turck quoted the book's own image back at him — a shift from a body of knowledge to a labyrinth of knowledge, and "34,000 journals that might as well have no trespassing signs." Socher agreed
He offered himself as the example. Socher said he studied linguistics, then computer science, and has spent recent years studying biology on the side, where ten sentences into a conversation with a biologist the abbreviations and terms pile up until most people stop following
Academia's incentives do the rest, in his telling. Careers reward novelty up to a point and punish it past that: he said most of his first papers on neural networks for natural language processing were rejected from the main conferences around 2010 and only accepted in small sub-groups
"That certainly happened to me a lot in the early days", he said, recalling the first deep-learning workshop at NIPS as 30 or 40 people — "just like a couple of us renegades"
Even within biology the layers do not talk. Socher said a biology PhD now works at the cell, tissue, biochemistry or protein level, and asking one a deep question about another layer often gets no answer
2. AI As Calculus For Biology
Asked to state the book's premise, Socher gave an analogy rather than a forecast.
His framing is that AI is the tool for reassembly, not for discovery in the small. Science has become good at understanding smaller and smaller pieces; putting them back together is the unsolved part. "AI is kind of what calculus did for physics AI will do for biology"
His examples are systems whose parts are understood and whose behavior is not. The microbiome and the brain: individual neurons, their synapses, their firing and — he noted, often ignored in neural models — their chemistry are all reasonably well understood, and nobody can say why the assembled brain has a thought
The advantage he claims is tractability, not transparency. A neural network can exhibit patterns that are hard for people to interpret, he said, but studying a neural network is still easier than studying the biological system it stands in for
He said the evidence is breadth, not depth. Physics, chemistry, biology, neuroscience, medicine, economics and astrophysics all show small improvements at once, and it is the accumulation across all of them that he extrapolates from
3. The Language Of Proteins
Turck put the skeptic's case — that generative AI is a chatbot and a next-token predictor — and asked why the same technology predicts protein structures.
Socher said the field has made this mistake before, about itself. Ten or twenty years ago natural language processing researchers would have said one neural network could never answer every kind of question; he said his own paper describing what is now called prompt engineering, decaNLP, was rejected by every reviewer and the area chair as useless
His summary of that reception: "Not even humans have one system to answer all these different kinds of questions"
"It was like such a non-obvious thing to the field", he said of an idea now so obvious that it could not be invented today
The bridge to biology is that the model does not know what kind of sequence it is reading. Amino acid chains are a language no human evolved to speak, and "AI doesn't really care if it's English or a sequence of amino acids." "No human has been sort of evolutionarily trained and has learned to speak the language of proteins."
The consequence is generation, not just prediction. Socher said you can now produce entirely new proteins the way you produce sentences that never appeared in the training data, and that the same idea extends down to chemistry and molecules
Why next-token prediction picks up a world model, in his explanation, is that geography is latent in sentences. His example, adapted on the spot to where they were sitting: predicting the next word after "I was in New York and I was driving north to" pushes the model toward Boston rather than Yale, and toward the bigger city — so the model absorbs where cities are without being taught geography
The protein version of the same effect is measurable. Socher said you can inspect a network trained only on next-token prediction and find that residues which end up close together in folded 3D space are correlated in the network too
He defined the term for non-technical listeners. A token can be an English word or part of one, a protein, a few pixels of an image or video, or a piece of sound — anything discretized into a vocabulary
Turck asked where the analogy breaks. Socher's answer was that it is oversimplified and useful anyway: "All models are wrong. Some are useful." Proteins are not just sequences, and chemistry has loops within molecules, yet describing a molecule as a string works well for a lot of chemistry
He does not claim these systems produce new physics yet. Recombination of existing ideas is where he sees the gains, because "we stand on the shoulders of giants" in almost all of science — and he pointed to cases where Recursive co-founder Jeff Clune had an AI generate a novel evolutionary-algorithms idea that people later published independently
On the bigger prize — proving or disproving a string theory — he said research still has to be done
What he says is already in reach: "We can already cure a lot of diseases. We can already cure many of the aspects of aging." Plus better fusion reactors, new materials, and tax and subsidy questions answered against a stated objective
4. Anything You Can Simulate
Turck asked for the mechanism behind machine creativity, raising the Go move nobody expected that let the machine beat the human champion. Socher reframed the question as one about verification.
His prediction rule has one input. AI will be superhuman in any domain where there is a simulation or a verification tool, because it can then run an unbounded number of experiments — provided the simulation does not take years to run
That is why he was never surprised by games. Perfect-information games like chess and Go have no hidden variables, too many positions for brute force, and the option of self-play — which he tied directly to open-endedness and to recursive self-improvement, and to the rainbow-teaming work on language model security
Math is next and the field knows it. Socher said mathematics will change massively within a few years and the most prominent frontier mathematicians are already aware of it — the way manual feature engineering and manual architecture engineering stopped being useful skills in AI
His caveat is about motive, not capability: good if you want as many theorems proved as possible, not good if you love math for its own sake
He thinks competition against machines makes humans better, including physically. Chess is more popular now despite being dominated; he expects the same for athletes, citing a robot's odd running gait found in simulation that humans, having run their whole existence, had not thought of
The most valuable simulable domain is programming. Socher reworked Marc Andreessen's line — software ate the world, and now AI is eating software — and gave the simplest verifier he could: show a picture of a website, ask for code that matches it, repeat infinitely
He expects all of programming to change, and with it the digital and knowledge economy
The limit is the natural sciences, because they cannot be simulated yet. He said we cannot perfectly simulate a complex cell, let alone tissues, organs or whole humans, and that closing that gap needs far more data collected robotically
5. AI For AI, Then Biology
Turck asked how far generalization goes, and whether it is brute-force reinforcement learning domain by domain or something broader.
Recursive's sequencing is deliberate and narrow at the start. Socher said the company is fairly sure "we have to start with AI for AI" — making it very good at research that produces better AI, until it has "the equivalent of 50,000 PhDs" in knowledge and capability — and only then going after physics, chemistry and biology
He named the data suppliers he expects to matter. Tahoe Therapeutics and Parallel Bio, both of which he said appear in the book, as examples of companies generating much more training data over the next two or three years
The biology equivalent of next-token prediction is the perturbation study. Knock out one gene or add one molecule, observe what happens, repeat enough times, and he thinks the model eventually learns the underlying pattern the same way it learned that New York plus driving north implies Boston
His honest assessment of where that stands is blunt. We are "nowhere near having enough training data for biology" — which is why he wants organoids and pooled perturbation studies feeding a virtual cell the AI can then experiment inside
6. Hallucination As Feature
Turck raised the book's counterintuitive claim that hallucination is a feature rather than a bug in AI-driven science.
The distinction he draws is between recall and invention. Any computer can memorize, Socher said; what is interesting is "how well can you hallucinate" — how far outside the distribution the predictions go while staying reasonable
He said the same knob does both jobs. Raising the temperature makes a language model generate tokens less like anything it has seen, which is what you want when exploring novel proteins and not what you want on a factual query
"I do think hallucinations can be also very helpful for AI when you want it to explore novel kinds of proteins", he said, concluding that "hallucinations are in some cases a feature and not a bug"
For factual questions the fix turned out to be plumbing, not symbolic reasoning. Socher said people assumed symbolic methods would be needed; what actually worked was giving the model real search results to summarize, which is what you.com does with a search engine built for agents
Turck noted the book's historical aside that scientific discoveries have come from scientists in altered or unhealthy mental states. Socher named Heisenberg among the physicists and said that in some cases actual psychosis pushed a field forward
The everyday version of the argument closes it. A poem written for your wife should not sound like the poems already out there
7. Writing Biology, Not Reading
Asked about the shift from reading biology to writing it, Socher started with why he disliked the subject at school.
He said high-school biology was memorization and that the field has changed character. Processes learned for a grade and forgotten six months later, versus a discipline that is now programmable
His general claim is about how sciences mature. Once the basic pieces are understood, the low-hanging fruit is recombining them usefully, which is the transition into an engineering science
"biology is becoming a programmable science. It's becoming an engineering science." His example of what that buys: one molecular piece attaches to a cell while a different connector injects something into it, so molecules can be recombined like parts
The aha moment he dates to 2018 at Salesforce. Socher said his team trained the largest language models for proteins, in a paper called ProGen with Ali Madani as first author, while he was chief scientist there
That work turned into a company with revenue. Madani went on to found Profluent, which Socher said has closed multi-billion-dollar contracts with Eli Lilly on the strength of designed proteins that are better than CRISPR-Cas9 at gene editing — more specific and more targeted at changing particular genes in living people
He pointed to a trial that made a different drug for every patient, which he said is a first for the FDA, and expects more of it
His complaint is about where trials can be run. Socher said running clinical trials has become hard in the US and in Europe, and that people are moving to China or Australia
China is cheaper and faster, with the caveat that intellectual property may leak
Australia decentralized clinical trials so every hospital can run its own, which he said replaces one central gatekeeper with competition — though he noted the country does not have enough people
He would like a version of the Australian system in the US
8. No Hard Takeoff
Turck asked what portion of a 10- to 15-year drug development cycle AI actually removes.
Socher rejected the hard-takeoff framing while keeping the optimism. "And as bullish and excited as I am about AI, I'm not a believer in this crazy hard takeoff" — because some things take time for reasons of physics and the real world
His example of an irreducible delay is the long-term trial. You want to know whether patients have a problem three years after they stop taking a drug, and no model shortens that
The change he thinks matters is to the shape of a biotech company, and it is a contrarian call he says his firm makes. The old pattern: eight to ten years of spending, a forced IPO because there are not enough late-stage private investors, one or two drugs in phase three, and a dead company if phase three fails
What he says is happening instead: companies with multiple drugs in late-stage phase two within six to 18 months, going public with eight or more candidates, each more likely to succeed because the predictive models are better
His closing qualifier on all of it: "everything in biology just takes longer than it does in software"
9. Who Loves AI, Who Hates It
Turck put the industry's standard justification to him — that AI will cure cancer — and asked what is real about it.
Socher's opening answer was about position, not technology. "I think if you care about the outputs of an industry or a company, then you love AI. If you get paid hourly, you probably hate AI."
He framed the resentment as rational rather than ignorant. Someone whose work is being recorded hourly to train their replacement, with no equity in the intellectual property they are creating, is understandably unhappy
His theory of which jobs survive is an elasticity argument. Whether a profession benefits depends on how much demand rises when the price falls
"illustrators hate AI" because the world needs a certain number of illustrations: "If you got paid 200 bucks for one illustration and now it's worth 2 cents, maybe you hate it" — demand did not rise a thousandfold when price fell a thousandfold
Coding went the other way, which he called the Jevons paradox: cheaper code meant far more code, and he said there is "more demand for programmers now" because each one is more productive
His illustration of the induced demand: people ending up with dozens of personal apps on their phones, and ideas nobody explored because the market looked too small
On cancer specifically he said yes, with conditions. Socher said AI will play a big role in curing multiple cancers, pointing to trials where it designs a specific cocktail of drugs and specific RNA sequences
His reason is structural: a cancer is not one homogeneous thing, it contains sub-cancers, and treatment has to be specialized per person and per form
Turck put it to him that cancer is a systems problem, which Socher agreed is precisely why AI suits it
The constraint stays the same — even a perfect molecule still goes through years of clinical trials
10. Opting Out Of Progress
Asked what else he is optimistic about, Socher ran through fields and then turned to whether everyone wants any of it.
His most vivid example is engineered bacteria that eat microplastics and die once the plastic is gone, which he said would help the oceans — with an explicit warning that they must not mutate into eating something else
He treats ecological intervention as a track record of mixed results. His example is forest management: suppressing small fires leaves the underbrush uncleared and makes the big ones worse. "Everything is a system. Everything is a complex system."
He listed the physics and materials cases. Balancing plasma in a tokamak for fusion is already a machine-learning control problem; more efficient solar cells and panels are being designed with AI; better batteries using more abundant materials than lithium, which he said he has invested in
He expects the experts to be wrong about longevity the way they were about language. "I understand sort of the famous saying of like if you want to know why something doesn't work ask the experts" — true of natural language processing and neural nets then, and true of longevity and cancer now, in his view
Then the turn: he thinks whole groups will decline to participate. Socher described people already off-ramped from progress — someone on a Greek island who goes fishing and does not have to think about AI — and expects more of it
His institutional example is Bhutan: "Bhutan decided we will not measure our gross domestic product based on money but based on happiness", with the honest caveat that people who want to build startups are probably not as happy there
He blamed his own industry for the backlash. "I think the AI industry has done a terrible PR job in general", which is part of why he wrote the book
His counter to the moral panic is a list of earlier ones. An 18th-century German novel blamed for suicides that every German schoolchild now reads as high literature, then comic books, then computer games, and now chatbots
His asymmetry argument: hundreds of millions of people getting cheaper medical, legal and emotional advice do not post about it, while the harms are vocal
Turck's summary, which Socher accepted: the future as a whole needs better marketing
11. Building An AI Economist
Turck raised the book's economics chapter and the stat that economists failed to predict "148 of the last 150 recessions," then asked about the AI economist Socher's team built at Salesforce.
His diagnosis of the field is the absence of a benchmark. Computer science improves because a number settles arguments; economics has no such number, so "economics often becomes just a political field" where a department's political slant decides which papers get written
The papers were desk-rejected. Socher said Nature and Science rejected the two-level reinforcement learning work without review, in one case by an ethicist he said knew nothing about AI
The simulation itself was simple, and he was clear about that. From 2018: agents with a utility function and a number of hours a day they were willing to work, sampled from priors — "not everyone wants to work 14-hour days"
The agents gathered resources, built houses, and could block other agents from resources to build monopolies
A meta-agent watched all of them and chose how to tax and subsidize each group
The reward function is where the politics has to be made explicit. His team used equality multiplied by productivity — you want growth, but not one agent owning everything — and he said you want neither alone, which is why the two are multiplied
The result was a recovery and then an extension. Socher said a famous optimal-taxation formula in economics is provably optimal in a one-step economy where a single decision is made and never revisited; the reinforcement learning system recovered that same answer, and then went further once the economy became a sequence of decisions with agents adapting against the tax scheme
The behavior the agents learned was avoidance: dumping assets before a tax year and taking gains just after it
What he wants is scale, and he called the ambition by name. Socher said he hopes the paper eventually gets "a GPD3 moment" — someone building a realistic simulation big enough that a policy proposal can be run through billions of simulated years before it is adopted
He was explicit that humans keep the decision: no AI making policy without oversight, only suggestions
And that the hard part is human: societies have to formalize what they are actually optimizing for
Turck pushed on whether an economy full of fear and greed can be modeled at all. Socher's answer was "All models are wrong. Some are useful," plus evidence that prompting a model to act as a 43-year-old from a given region produces something close to what such a person says
He does not expect the US to use it. Identity politics, special interests and super PAC funding make it very unlikely, he said; Singapore and China are where he thinks an explicit objective function gets tried
12. The Four Pillars
Turck asked Socher to walk through the four pillars of the Eureka machine, and raised the data-quality problem of training AI on AI output.
On data quality he conceded the point and widened it. "AI often is only as good as the people, the data, the systems, the infrastructure, the rewards that we give it" — and he said it is still surprising how poorly engineered some frontier labs' environments and sandboxes are
Pillar one is human knowledge through language models, ingesting the world's recorded knowledge and reasoning over the combinatorial space of what already exists
His aside on how that knowledge moved: the large closed labs took what they could from the open internet, Chinese open-source companies distilled a lot of it back out of those closed models, and then open-sourced the result — putting the knowledge back where it started
Pillar two is a model of reality itself, built from instruments rather than text. Humans cannot perceive gravitational waves or gamma rays, he said, but machines can measure them, and much of that data has never been described in human language — and in some cases could not usefully be
His analogy for why interpretability is not the goal: you can list the five million parameters behind a predicted word and gain no intuition, just as nobody can say why they moved a particular muscle fiber in a pinky while turning a steering wheel
Asked whether pillar two exists, he said no. It would need many sciences, universities and labs collaborating on one foundational model of physics, chemistry and biology, and he noted that plenty of knowledge was never worth putting on the internet and is therefore still hidden
He is skeptical of the rebranding: companies building "foundational models," now often called world models, typically produce something narrow — a virtual cell good at estimating particular gene variants and nothing else
A virtual cell in full, he said, does not exist yet and is the goalpost that would require many teams to pool data
Pillar three is simulation, and he wants abstraction rather than completeness. In a perfect world it would be one simulation for everything; in practice most questions do not need quantum detail, and computer science is the field that is good at building the abstractions that let lower levels be ignored
He allowed that the abstraction can cost you something real — quantum biology effects that only show up in a larger simulation
What exists today, he said, is simulation "in many small bits and pieces": chess and Go at the easy end, a virtual cell as the hard target that is still far off
Pillar four is the real world. At some point a system has to be engineered and run to see what the simulation missed — a lab for physics, chemistry and biology, and eventually real machines and satellites taking measurements at every scale
13. Robot Labs, Agent Swarms
Turck asked about self-driving robotic labs and about what the agent swarm on top of the four pillars actually does.
He likes the first attempts and thinks the timing is early. Socher named Periodic Labs as a good example and said that from an investing perspective it feels a little early — but that in two to three years it will be right on time, once the software, the scientific data and higher-fidelity simulations are in place
What he wants then is expensive experiments, not cheap ones. Ask the AI to design experiments that take hours, days or weeks, run against better organoids and small cell systems derived from human stem cells
One regulatory result is already banked, and he made an animal-welfare argument out of it. Socher said Parallel Bio has FDA approval to skip certain animal trials, which saves money over the coming years and saves the lives of animals bred to be tested on and dissected — so "you can also love AI"
He also mentioned a chemistry machine that can assemble a small set of molecules, and Parallel Bio running trials with organoids
The agent swarm is modeled on how scientific communities actually work, exploring different ideas in parallel and recombining them — open-endedness, which he said is central to Recursive
Clune's microwave example is the argument for indirection. You do not get a microwave by demanding a faster pot, Socher said; you get it because someone working on radar noticed a chocolate bar melting in his pocket
His conclusion is that these recombination paths across research fields can now be modeled with agent swarms
14. Inside Recursive
Turck asked about the compute and data constraints, then about the company.
He named the binding constraint in three words. "Compute is the biggest constraint." Socher expects companies, and eventually humanity, to have to decide which problems are worth the compute, and expects new scaling laws describing how much compute a class of problem needs
On data he is less worried than the industry. Most of the public internet has been digested, he agreed, but new data keeps arriving — which is why you.com supplies newer labs with fresh search results about things that happened last week and were in no training set
Recursive's founding goal is stated as automation of a research process, not a product. Socher said it was built to create recursively self-improving superintelligence to automate knowledge and scientific discovery, and that the eight co-founders arrived at the same conclusion from different directions
Tim Rocktäschel and Jeff Clune came from open-endedness research and evolutionary algorithms
Socher came from a ladder of automation: feature engineering automated by word vectors, architecture engineering automated by one unified architecture, and ideation, implementation and validation of AI research ideas as the next rung
"And we believe that will be a great unlock" for applying that intelligence to other sciences
He revised the round size upward and dismissed the compute commitment as small. Turck put it to him that Recursive raised $650 million and committed $410 million of it to a single compute deal with Amazon; Socher said the raise ended up around $670 million, and that the deal "will probably be one of the smallest" compute deals in the company's future
He would not say what ships, but committed to a date. "I can guarantee you they will happen this year", he said of a couple of releases he could not detail
He dislikes the neolab label and drew a line under it. Socher said many neolabs will not succeed and that Recursive is "a real company not an academic lab" talking to real customers
Asked why most neolabs are not real companies, he said they are ideas for exploring one question rather than products, and that founding teams need shipping experience and not only research
The published evidence so far is narrow and concrete. Socher pointed to a blog post of milestones toward full recursive self-improvement: a first narrow instantiation of the Eureka machine that can already outperform "months and sometimes years of human endeavor" on particular problems
"We also showed that they can build new CUDA kernels which is very useful for faster inference" — useful, he said, to hyperscalers and anyone serving tokens, with positive feedback from users and from the Nvidia staff who built the benchmark
He expects the sequence to run general first, then progressively more specific
15. How Far Intelligence Goes
The book ends on how far intelligence can go, and so did the interview.
He was surprised nobody had defined intelligence properly. Socher said his own current candidates for the foundational building blocks are "prediction action and goals" and combinations of the three, and that there is no unit of intelligence the way there is a unit of energy — something he said he thinks about a lot
His model for the taxonomy is physics. Kinetic and potential energy, then chemical, mechanical and electrical energy, some still pure science and some pure engineering; he sees the same pattern in visual, language and physical intelligence
"I define these 10 different spaces of intelligence", he said, each with many dimensions
His worked example is vision, and it ends at physics rather than engineering. Humans see a narrow band of the electromagnetic spectrum; an intelligence could see from gamma rays to gravitational waves, and could have millions or billions of sensors rather than two eyes — until it runs into the speed of light and the light cone around each sensor
That is his answer to the bubble question. "Boy, are we far away from the true upper bounds of any of the spaces of intelligence" — the unit cost of intelligence may fluctuate with a number of factors, he said, the way energy prices do, but the field and the civilization can still go much further
Bonus Insights
His name for the early neural-network community was "renegades." Socher described the first deep-learning workshop at NIPS as 30 or 40 people who are now famous, arriving from different directions — feature learning instead of feature engineering, and neuroscience inspiration — and said the intersection of fields is both where the interest is and where publication is hardest
He said the novelty threshold for machine-generated ideas has already been cleared more than once, citing Clune's public accounts of AI-generated evolutionary ideas that people later published
On whether AI could model irrational human behavior, his position is that fidelity rises until the simulation crosses a threshold of usefulness, not that fear and greed get solved
Turck raised the claim that the industry has run out of new data. Socher's answer was that news keeps happening, which is the commercial case for the search layer he already runs
He recommended one book that is not his own. David Deutsch's "The Beginning of Infinity," for its account of how material problems were solved by better explanations
He pointed listeners to two places at the end: his own account on X and recursive.com
Socher's bottom line is that AI becomes superhuman wherever a result can be simulated or checked, which is why code and math are falling now and cells are not — and that the work worth doing is building the missing simulator for biology, because the pieces of science are already understood and nobody can put them back together by hand.
Products, Companies & Tools Mentioned
Recursive (His new company, eight co-founders, built to automate AI research and reach recursive self-improvement; he put the round at about $670M)
you.com (The search engine he founded, which supplies newer AI labs with fresh search results as grounding)
Salesforce (Where he was Chief Scientist when his team trained ProGen, the first large protein language model, in 2018)
Profluent (Founded by ProGen's first author Ali Madani; Socher says its designed proteins beat CRISPR-Cas9 on specificity and that it has closed multi-billion-dollar contracts with Eli Lilly)
Tahoe Therapeutics and Parallel Bio (His examples of companies generating the biological training data the field lacks; he says Parallel Bio already has FDA approval to skip certain animal trials)
Periodic Labs (Named as a first real attempt at a self-driving robotic lab — early for investors today, right on time in two to three years, on his estimate)
Anthropic and OpenAI (The closed labs that ingested the open internet, and whose knowledge he says Chinese open-source models distilled back out and re-released)
Nvidia (Its staff built the benchmark Recursive's machine-generated CUDA kernels were measured against)
Amazon (Counterparty on the single large compute deal Turck put to him, which Socher said will look small in hindsight)
AIX Ventures (The firm behind what he called a contrarian position on investing in biotech)
Books & Resources Mentioned
The Eureka Machine – Richard Socher (His own book, out 22 September, and the source of the four-pillar framework and the slowing-progress argument)
The Beginning of Infinity – David Deutsch (The one book he recommended, for its account of material problems solved by better explanations)
His decaNLP paper (The rejected paper describing one network promptable with any question — what is now called prompt engineering; he said the reviews are still online)
Recursive's milestones blog post (Where the narrow Eureka machine results are published, including the CUDA kernel work)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: