Richard Socher runs a company built to make AI improve itself. He doesn't expect that work to produce superintelligence any time soon. Asked where recursive self-improvement stands, the co-founder and CEO of Recursive said: "In various weak forms, we already have RSI. We're not quite there yet, but we're very close." Asked when artificial superintelligence will arrive, he said that in the strongest sense of the term, being better than all of humanity across ten spaces of intelligence, it "will take us probably several decades."
The two answers started an argument on Moonshots with Peter Diamandis. The episode was recorded on October 2, 2026 and published the next day. Alexander Wissner-Gross, a computer scientist and founder of Reified, said he was "very confused." Host Peter Diamandis took his side.
What recursive self-improvement means
Recursive self-improvement, or RSI, describes an AI system that improves the software, research or hardware used to build the next version of itself. That version can then improve the one after it. Artificial superintelligence, or ASI, is a looser term for AI that exceeds human ability. The definition was part of the dispute.
Socher has a long research record in natural-language processing. He founded MetaMind, became Salesforce's chief scientist after Salesforce acquired it, and went on to found You.com, according to his biography. On the show Diamandis said Recursive had raised $670 million from Google Ventures, Greycroft, NVIDIA and AMD, plus $410 million in compute from Amazon Web Services. He also disclosed that he is a seed investor in the company. GV, Google's venture arm, gives a slightly different figure in its own announcement: it says it co-led early funding of $650 million at a $4.65 billion valuation.
Why Socher says the weak version is already here
Socher said "something major shifted" this year: AI can now write code. "AI is code and AI can code," he said, "so you have a loop that you can close there."
He counts the way Anthropic and OpenAI talk about their engineers using Claude or Codex to create code as a weak form of recursive self-improvement. It is weak, he said, because "you still have like deeply embedded humans in that loop."
Recursive wants to shrink the human role. Socher said people would only set up "the rewards and the environment and the goals." The AI would handle the whole cycle of ideation, implementation and validation: proposing an idea, building it and testing whether it worked. On top of that cycle the company wants to run open-ended evolutionary search algorithms that combine "interestingly different ideas." Recursive's website describes the same approach, modeled on biological and cultural evolution. It says the company keeps a varied pool of discoveries that can be recombined instead of chasing one fixed answer, and that its first target is AI research itself.
Socher said the main obstacle is physical: "you still need a lot of compute." An RSI system that is asked to design and train sophisticated versions of itself needs a lot of computing power. Even so, he predicted, "this is going to take off, next year."
Five things an AI could change about itself
To show what is still missing, Socher cited a paper by his friend Jason Weston of Meta. It lists the learnable "axes" along which an AI system could improve itself: its parameters (the numbers adjusted during training), its training data, its objective function (the target it is trained to optimize), its neural architecture, and its overall code and harness, meaning the surrounding software that runs it. Socher said no one has "cracked the nut" of doing all five, let alone having the AI decide which of them to optimize. The evidence, he said, is that "all the big companies are still hiring thousands of engineers to do it manually to a large degree."
The paper, written by Weston and Jakob Foerster, does set out these five dimensions. It treats improvement across all five as the full recursive vision. Its authors argue for a different goal, though. They propose human–AI co-improvement, in which AI systems are trained to work with people on choosing problems, running experiments, evaluating results and handling safety, with humans taking part throughout. Recursive's stated aim is to limit humans to setting rewards, environments and goals.
Where does recursion begin?
Salim Ismail, founder of Open ExO, said this was the point where he felt "soft." He listed a continuum and asked Socher where real recursion starts. In his list, AI first writes some code that the next model uses. Then it proposes experiments for researchers, runs those experiments, evaluates the results and modifies its own training system. Finally, it launches the next iteration of itself "without meaningful human intervention."
Socher answered with his three stages. In the strongest sense, he said, recursive self-improvement requires an AI to carry out ideation, implementation and validation itself in an inner loop. Around that loop there must be "an outer process that is more open-ended," where the AI can innovate and recombine ideas, "similar to biological cultural and technological evolution."
Recursive's June 11, 2026 research report describes early steps toward that loop. Parallel loops proposed ideas, implemented them, ran experiments and validated the results. In the report, the automated system improved a small-model training benchmark called NanoChat and slightly shortened a NanoGPT training "speedrun" on eight H100 chips, from 79.7 to 77.5 seconds. It also raised the average score on a set of GPU programming tasks. The authors note that the underlying models may already have seen the code repository during training. They also describe some kernel scores that were inflated by cached outputs, which they countered with stricter checks and human feedback.
A demanding definition of superintelligence
Socher's bar for ASI is high. It should surpass "not just arbitrary humans like a Turing test, but all of humanity" at arbitrarily hard tasks. It should do so across the ten "spaces of intelligence" defined in his new book, The Eureka Machine. In the conversation he named perceptual, communication, social and creative intelligence, speed, metacognition, knowledge and mathematical reasoning among them. He said the system must also be able to "choose to some degree what it works on" and have some metacognition, meaning awareness of its own thinking. Diamandis offered Elon Musk's definition, "as smart as all humans combined." Socher went further: "smarter than humanity combined."
Ismail objected to the vocabulary. People can be emotionally, physically, linguistically or musically smart, he said, so "smarter seems to be a very vague term to me."
Spikes now, physical limits later
When Diamandis noted that Musk says "2029, 2030, latest," Socher said Musk probably means weaker forms. He expects AI to be better than all of humanity at programming and, within a few years, at math. He also expects it to win any game where all the pieces are visible, such as Go and chess. But humanity can build a particle collider, make a few atoms of gold and create new molecules, he said. "It's going to take a while before we even give AI the access to the physical world such that it can innovate in that way beyond all of humanity."
His main example was chipmaking. An AI that controls and upgrades its own computing hardware would need new supply chains, new materials and its own version of the machines made by ASML, which create chips with features of one to two nanometers. Building such a machine, Socher said, will take more than two years, "even if you had the perfect blueprint."
Wissner-Gross had asked how a recursive self-improvement company could think superintelligence was 20 years away. He said replacing ASML might take three years, using Musk's free-electron laser as his example, but no more. Hooking AI up to the physical world, he said, is "the easy part." He pointed to an earlier episode in which a model taken "straight out of the box" could drive a car once it was given the controls. "Completely don't buy the 20-year timeline," he said, adding that he was apparently "drinking very different singularity water." Diamandis said: "agree with you, Alex, for what it's worth."
The end of the rainbow
Wissner-Gross then asked Socher to set timelines aside and describe the "fixed point" of recursive self-improvement: what the perfect model looks like once the process has run its course. Socher said "it would be a hubris for us to know, right now." He added that there were things about Recursive's architecture work he could not share. He predicted that the end state would be "incredible," able to out-innovate people on any dimension, though he said confirming that new drugs work would still take years of clinical trials.
Wissner-Gross tried again by way of the NanoGPT speedrun, a community contest to train a GPT-2-class model as fast as possible. He said the record had collapsed through algorithmic improvements alone, without more data, and that a "mini-scandal" had broken out over the previous two weeks. In his account, someone cut the time from roughly 60 or 70 seconds to about 40 by factoring world knowledge out of the final model, and people argued over whether that still counted as training. Socher hinted at news: "you just wait for a few more days. There'll be another really fun update there."
Wissner-Gross used this to ask the real question. Some argue the ideal model would keep world knowledge in a text file or database and reduce its weights to a pure reasoning kernel, perhaps a megabyte in size. Ismail cut in to say that, by definition, no one can predict where a singularity goes, then added that this was a definition he himself did not subscribe to.
Socher rejected the clean split. "Just like humans benefit from memorizing things," he said, an AI also has to have knowledge partly in its weights, because reasoning creatively over concepts requires "some of that world knowledge deep inside the model." He expects a separate store of knowledge as well, the way people use a search engine. "But the main model will have a lot of that mixed in for sure."