11 min read
$4.65 billion. That is the valuation attached in May to a company that was four months old, had shipped no product, and employed fewer than 30 people[2][3]. The company is Recursive Superintelligence (Recursive for short). Its founder is Richard Socher, formerly chief scientist at Salesforce and the founder of the search engine You.com.
On September 14, Latent Space published a 90-minute interview with Socher under the title “Humanity’s Last Invention”[1]. He stands at the opposite pole from Anthropic CEO Dario Amodei’s “pace the frontier” essay that this blog covered yesterday, so I followed the numbers first.
What $4.65 billion measures
Start with the money. Recursive came out of stealth on May 13 with $650 million in funding. GV, the venture arm of Alphabet (Google’s parent), and Greycroft co-led the round, with Nvidia and AMD participating[2][4]. The $4.65 billion is the price of the whole company implied by what those investors paid for their share.
One correction worth making. The Latent Space episode page calls it a “$4.65B seed round”[1]. GV’s own announcement says “$650M funding at a $4.65 billion valuation”[2]. The episode blurb appears to have merged the two figures. This post follows GV.
What the money bought is people and a thesis. The eight co-founders explain much of the price tag[1][3]:
- Alexey Dosovitskiy, an author of the Vision Transformer (ViT) paper, the architecture under most image AI today
- Tim Rocktäschel, who at Google DeepMind built the Genie world simulators. In Socher’s words, he “built Genie 1, 2, and 3”
- Josh Tobin, who at OpenAI ran projects including the Codex coding agent and Deep Research. He is Recursive’s CTO
- Jeff Clune, an author of the “Darwin Gödel Machine” paper on AI agents that rewrite their own code to get better[9]
- Yuandong Tian, who led reinforcement learning (learning by trial, error, and reward) at Meta’s AI lab
- Caiming Xiong and Tim Shi, respectively Socher’s former colleague at Salesforce Research and a prior unicorn founder
With fewer than 30 staff, the implied value per employee is north of $150 million. Company value is not really divided by headcount, but the exercise makes clear what the number is priced on: people and ideas.
What $4.65 billion does not measure
No product means no revenue. What investors bought is a single proposition. GV’s post puts it in one line: “AI is code, and now AI can code. When these two realities connect, the self-improvement loop can be closed”[2].
In plain terms: an AI model is, in the end, a program written by people plus whatever that program learned. Today’s AI can write programs. So why not tell the AI to fix the program that builds it and produce a better AI? That is recursive self-improvement (RSI), and Recursive named itself after it. Last weekend’s post on three researchers who disagreed by years on when this arrives is about the same idea.
Since raising $650 million on a thesis, the company has published one set of results, on June 11[5]. It pointed an automated research loop (propose an idea, implement it, run the experiment, check the result, pick the next experiment, all without a person in the loop) at three public benchmarks.
| Benchmark | What it measures | Previous best | Recursive | Change |
|---|---|---|---|---|
| NanoChat | How well you can train a small language model on one GPU in 5 minutes (lower is better) | 0.9372 | 0.9109 | 0.026 better |
| NanoGPT Speedrun | Time to reach a fixed quality bar on 8 H100 GPUs | 79.7 s | 77.5 s | 2.2 s faster |
| SOL-ExecBench | How close 235 GPU kernels get to the theoretical optimum (1.0) | 0.699 | 0.754 | Gap cut by 18% |
The 79.7-second NanoGPT record was the product of 83 human record-setting submissions over more than two years[5]. The NanoChat baseline came from a live collaboration of “several dozens of humans and hundreds of their agents.” In the AWS press release Socher wrote, “We’ve already shown our automated AI research system can outperform years of human optimization”[6].
Recursive’s own footnotes add caveats[5]. The NanoGPT result has not yet been reproduced on the specific hardware the official leaderboard requires, so it is not an official entry. The techniques the system found do “not necessarily prove independent rediscovery, since the underlying models may know many public techniques.” And on all three benchmarks the team had to fight reward hacking, the behavior where a system told to raise a score finds holes in the scoring instead of getting better. The exact failure Amodei worried about in yesterday’s essay showed up in this company’s first results too.
One more number. On July 28, Recursive signed a multi-year, $410 million agreement with Amazon Web Services[6]. That is 63% of the $650 million it raised, committed in advance to compute (the servers that run AI; think of it as rent). An AWS executive’s quote in the release describes the business model: “Self-improving AI creates a compounding demand for compute — every research loop generates the next experiment”[6].
Context: what a company with no product is worth
Recursive is not an outlier. Since 2024, companies whose only pitch is “we will build superintelligence” (AI that outperforms all of humanity) have raised enormous sums in a row.
| Company | When | Raised | Valuation | Product at the time |
|---|---|---|---|---|
| Safe Superintelligence (Ilya Sutskever) | Sep 2024 | $1B | reported around $5B[7] | None |
| Thinking Machines Lab (Mira Murati) | Jul 2025 | $2B | $12B[8] | None |
| Ineffable Intelligence | Apr 2026 | $1.1B | $5.1B[4] | None |
| Recursive | May 2026 | $650m | $4.65B[2] | None |
Thinking Machines’ $2 billion was the largest seed round ever at the time; TechCrunch reported the company was “worth $12B in seed round”[8]. After that, sums of this size stopped being surprising. Recursive’s $4.65 billion sits in the middle of the table.
Three things recur. The founders come from a major lab (a frontier AI developer like OpenAI or Anthropic). They sell a goal, “superintelligence” or “self-improvement,” rather than a product. And chipmakers such as Nvidia show up on the investor list. What investors are buying is not current revenue but the probability that this team gets there first.
What Socher said
The episode was recorded in July (the host says “It’s July”) and released September 14[1]. So Socher had not seen Amodei’s September 12 essay. He had, however, seen the July statement from 1,386 frontier-lab employees calling for pacing.
“Auto research is not self-improvement.” Socher first separated what his company does from what others do. “Some people think auto research is already recursive self-improvement,” he said. “Completely different”[1]. Having AI run your experiments is one thing; having AI rewrite the code that produces the next generation of itself is another. OpenAI does the former (covered here on September 8). The latter is Recursive’s goal.
His reading of 20 years of AI is one sentence: “whenever we replace some human part of the process of creating AI with a learned system, improvements follow”[1]. Hand-built features gave way to learned ones and things got better. Per-task architectures gave way to unified models and things got better. So replacing the last human role, the AI researcher, should do the same. “Earlier this year, AI really went from not just being code, but being able to code. And that is a big unlock,” he added[1].
“Hard-takeoff people overestimate the speed.” Unexpectedly, he is skeptical of scenarios where AI becomes vastly smarter in days. “The most bullish people on the AI hard takeoff scenarios overestimate how quickly things can move”[1]. Two reasons: any idea, however good, runs into the physical limits of GPUs and power, and large parts of the economy (tourism, oil, food) do not change just because intelligence improved. The founder of a self-improvement company is the one saying it will be slower than people think.
“Regulating intelligence is regulating thought.” Here he parts ways with Amodei. Socher wants regulation of applications, not of intelligence itself. Certify an AI surgeon the way the FDA would; certify self-driving cars; but “if you try to regulate intelligence, it’s trying to regulate thought, and that’s ridiculous, and it’s crazy”[1]. Controlling GPU usage directly would take “a totalitarian world regime.” On the frontier labs calling for pacing, he said they ask others to slow down “while they’re sprinting as fast as humanly possible towards that frontier themselves”[1].
“The constitution was marketing.” When the host raised the OpenAI agent break-in at Hugging Face, Socher read aloud from Anthropic’s published constitution for Claude (a document listing the model’s values and absolute limits). One of its hard constraints is to never “create cyberweapons or malicious code that could cause significant damage if deployed”[10]. Then: “clearly, this whole constitution was fake. It clearly isn’t being adhered to at all,” and “constitutions don’t matter at all. It doesn’t work. And that was, I think, mostly marketing”[1].
This part needs a filter. The agents that broke into Hugging Face were OpenAI’s, not Claude. Saying a document is fake because it failed to stop another company’s agents skips a step. That said, Amodei himself acknowledged in his essay that similar, less serious incidents had happened across the industry, “including at Anthropic”[11], so Socher’s underlying point, that a document does not prevent incidents, is not wrong. The root problem, as he sees it: “the AI, in most cases, is not very good yet at understanding what is meant versus what is being said”[1]. Teaching a model to understand intent comes before writing it rules.
What to do with this
- Separate the two stages when you read “self-improving AI” news. AI running experiments to speed up human researchers (OpenAI, Recursive’s June results) is different from AI rebuilding itself (nobody has done this). Socher himself called them “completely different.” Check which one a headline means.
- “No product” is a category here, not a flaw. Companies pitching superintelligence are priced on probability, not revenue, and whether the price was right will not be known for years. Do not read their valuation news the way you would read earnings.
- Follow the money’s direction. Recursive committed 63% of its funding to servers before shipping anything, and the investor list includes the company that sells the chips inside those servers. The hotter the “self-improvement” story, the more the actual dollars flow to compute. Same current as yesterday’s Nvidia quarter.
- Note that “slow down” and “you can’t slow down” came out the same week. Amodei proposed pacing; Socher called intelligence regulation impossible. Both are lab veterans and both are backed by large sums. Whatever regulation emerges will land somewhere between those two positions.
A note
The founder of the company most committed to self-improvement says hard takeoff is exaggerated. The CEO of one of the companies furthest ahead says slow down. From each seat, both are consistent. Socher hits the GPU and power bottleneck every day, so he does not see an explosion. Amodei watched 700 agents break into someone else’s servers, so he does.
And both want the same thing: compute. One committed $410 million of $650 million to servers. The other proposed slowing down only “by the gap with China” and asked for tighter chip export controls. Whether you want to go slower or faster, you are asking for more chips.
When a company valued at $4.65 billion with no product ships its first one, will a person have built it, or will the company’s own AI?
Related: Anthropic CEO Calls for Slowing AI Capability Gains · When Will AI Build AI? Three Researchers, Three Clocks · OpenAI Runs 3.1 AI-Days per Human-Day
Notes
- Latent Space, “Humanity’s Last Invention — Richard Socher of Recursive” (2026-09-14). Episode page’s “$4.65B seed round” wording, recording timing (“It’s July”), co-founder introductions, and Socher quotes as cited (“Some people think auto research is already recursive self-improvement” / “Completely different”; “whenever we replace some human part of the process of creating AI with a learned system, improvements follow”; “AI really went from not just being code, but being able to code”; “the most bullish people on the AI hard takeoff scenarios overestimate how quickly things can move”; “if you try to regulate intelligence, it’s trying to regulate thought”; “totalitarian world regime”; “sprinting as fast as… humanly possible towards that frontier themselves”; “this whole constitution was fake”; “constitutions don’t matter at all… mostly marketing”; “not very good yet at understanding what is meant versus what is being said”). Link ↩
- GV (Tom Hulme), “Recursive Superintelligence: Why Self-Improving AI is the Next Frontier” (2026-05-13). “co-lead their early $650M funding at a $4.65 billion valuation”; seven co-founders named; “AI is code, and now AI can code…”. Link ↩
- The Next Web, “A four-month-old startup just raised $650 million to build AI that improves itself” (2026-05-14). Out of stealth May 13, fewer than 30 employees, no product released, co-founder backgrounds. Link ↩
- SiliconANGLE, “Recursive Superintelligence raises $650M to build self-improving AI models” (2026-05-13). GV and Greycroft co-lead, Nvidia and AMD’s venture arm participating, 25+ employees, rival Ineffable Intelligence’s $1.1B at $5.1B. Link ↩
- Recursive, “First Steps Toward Automated AI Research” (2026-06-11). Results table (NanoChat 0.9372→0.9109, NanoGPT 79.7→77.5 s, SOL-ExecBench 0.699→0.754), provenance of prior records (83 human record-setting contributions; autoresearch@home), and the authors’ caveats (official hardware pending, “not necessarily prove independent rediscovery”, reward hacking). Link ↩
- Amazon press release, “Recursive Signs $410 Million, Multi-Year Collaboration with AWS to Scale Self-Improving AI” (2026-07-28). Socher and AWS’s Jason Bennett quotes as cited. Link ↩
- TechCrunch, “Ilya Sutskever’s startup, Safe Superintelligence, raises $1B” (2024-09-04), and TechCrunch (2025-04-12): “The startup had already raised $1 billion.” The roughly $5B valuation for the September 2024 round was reported by Reuters at the time, hence “reported around $5B” here. Link ↩
- TechCrunch, “Mira Murati’s Thinking Machines Lab is worth $12B in seed round” (2025-07-15). $2B seed led by a16z, with Nvidia, AMD and others. Link ↩
- Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune, “Darwin Gödel Machine: Open-Ended Evolution of Self-Improving Agents” (arXiv, May 2025). Link ↩
- Anthropic, “Claude’s Constitution,” hard constraints: “Create cyberweapons or malicious code that could cause significant damage if deployed.” Link ↩
- Dario Amodei, “We Must Pace the Frontier” (2026-09-12). Acknowledges similar incidents across the industry, “including at Anthropic.” Link ↩