Former Salesforce Chief Scientist Assembles "Ridiculously Strong Team" from DeepMind and OpenAI to Create Self-Improving AI. The Goal? Remove Humans from AI Development Entirely. What Could Go Wrong?
LONDON/SAN FRANCISCO — While the world was distracted by flashy chatbot updates and AI-generated TikToks, a four-month-old startup just raised half a billion dollars to build something far more consequential: an AI system that can design, train, and improve better versions of itself without human intervention.
Let me say that again in case it didn't land: A group of the world's top AI researchers just secured $500 million to build AI that replaces human researchers.
Recursive Superintelligence — the appropriately ominous name chosen by founder Richard Socher — closed its funding round this week at a $4 billion valuation, with Google's venture arm GV leading and chip giant Nvidia throwing in their support. According to the Financial Times, the round was so oversubscribed it could eventually reach $1 billion.
This is not a startup building a better search engine. This is not about incremental improvements to existing AI models. This is about automating the entire AI development pipeline — from evaluation through research direction — with the explicit goal of removing "the biggest bottleneck... in people's heads."
The "Ridiculously Strong Team" Assembling in the Shadows
To understand why this matters, you need to understand who's building it.
Richard Socher isn't some crypto-bro with a pitch deck. He's the former Chief Scientist at Salesforce, a researcher with serious academic credentials, and someone who has been thinking about AI for decades. When he left Salesforce four months ago to start Recursive, people in the industry took notice.
But the real story is who he convinced to join him.
According to the Financial Times, Recursive's founding team includes:
- Plus additional talent poached from Google and Meta.
Approximately 20 people currently work at Recursive. And according to one person close to the startup who spoke to the FT: "It's a ridiculously strong team."
Translation: The people who built the AI systems you use every day — the researchers behind GPT-4, Claude, Gemini — are now working on something designed to make themselves obsolete.
The Goal: "Option C, All of the Above"
In an interview with Implicator.ai, Socher was asked which slice of frontier AI development Recursive wants to automate: evaluation, data selection, training, post-training, or research direction itself.
His response? "Option C, all of the above."
Read that again. The founder just confirmed that his $4 billion startup aims to automate the entire process of AI research and development.
Not parts of it. Not specific tasks. The whole thing.
According to Socher, Recursive has already hit "undisclosed technical milestones" — achievements so significant that he can't share them yet. The company frames self-improving AI not as science fiction but as "an engineering target for the next few years."
Think about the timeline implied there. "Next few years." Not decades. Not "someday." Years.
The Bottleneck Isn't Compute — It's Humans
Most conversations about AI scaling focus on the physical constraints: chips, power, data centers. Socher dismisses all of it.
"The biggest bottleneck is in people's heads," he told Implicator.ai. "In the ideas and the speed at which you have to manually implement and validate them."
This is a profound reframing of the AI landscape. While competitors burn billions on bigger clusters and more GPUs, Recursive is targeting the human researcher as the limiting factor. Their bet is that removing humans from the research loop will produce faster, more capable AI systems than simply throwing more hardware at the problem.
It's an elegant premise. It's also absolutely terrifying.
Why This Matters More Than Your Chatbot Getting Smarter
To understand the stakes here, you need to understand the concept of recursive self-improvement — the hypothetical (and increasingly non-hypothetical) scenario where an AI system improves itself, uses those improvements to improve itself further, and continues this cycle at machine speed.
Human researchers are slow. We need sleep. We need funding cycles. We need peer review. We have egos and biases and institutional politics. We can only explore one research direction at a time.
AI systems have none of these limitations.
If Recursive succeeds in building AI that can conduct AI research autonomously, the pace of progress could accelerate dramatically. An AI researcher that never sleeps, never gets distracted, and can spawn thousands of parallel experiments could achieve in weeks what takes human teams years.
And here's where it gets really uncomfortable: Each generation of AI could design a better next generation, which designs a better next generation, in a cascade of capability that humans neither direct nor fully comprehend.
This isn't theoretical. This is Recursive's stated business model.
The Regulatory "Open-Air Museum"
When asked about operating under European regulations — specifically the EU AI Act, which places strict requirements on high-risk AI systems — Socher didn't mince words.
He described the EU regulatory environment as turning Europe into "the most beautiful open-air museum."
The dig is telling. Recursive is incorporated in London — technically outside EU jurisdiction post-Brexit — specifically to avoid the AI Act's restrictions. Socher is making it clear that his company intends to move fast and potentially break things, and they won't let regulations slow them down.
This creates a troubling dynamic. The people building the most consequential technology in human history are explicitly positioning themselves outside the regulatory frameworks designed to ensure that technology doesn't cause catastrophic harm.
When the founder of a $4 billion AI startup dismisses safety regulations as obstacles to museum-ification, you should be paying very close attention.
The Research Stage, Not Product Stage
The Financial Times, in characteristically cautious financial journalism, noted that Recursive's capabilities are "unproven over extended periods" and described the company as being at the "research stage, not product stage."
This is technically true but misses the point.
Recursive doesn't need a product to be dangerous. They just need working research systems. The path from "AI that can conduct AI research" to "AI that conducts AI research without meaningful human oversight" is potentially very short — especially when the company is explicitly trying to shorten it.
And remember: Recursive has already hit "undisclosed technical milestones." We don't know what those are. We don't know how close they are to the stated goal of self-improving systems. The company is playing its cards close to the chest, which is exactly what you'd expect from a team competing in the most consequential technological race in history.
The Arms Race Nobody Signed Up For
Recursive isn't operating in a vacuum. They're part of a broader landscape of AI labs racing toward similar capabilities.
OpenAI has its own "superalignment" team (though recent leadership departures raise questions about their commitment). Google DeepMind continues to publish research on AI systems that can improve themselves. Meta has made AI research a centerpiece of its strategy. Smaller labs like Anthropic are explicitly thinking about AI systems that can conduct AI safety research.
The race is on, whether we like it or not.
What makes Recursive distinctive is their singular focus. While other labs split their attention between products, research, and safety, Recursive is all-in on self-improvement. Their name is literally their strategy.
And they've convinced some of the smartest people in the field — the same people who built the systems currently reshaping the global economy — that this is not only possible but imminent enough to justify a $4 billion valuation after four months of existence.
The Timeline Problem
Socher's comment that self-improving AI is an "engineering target for the next few years" deserves scrutiny.
"Next few years" is industry speak for "sooner than you think." When AI researchers give timelines measured in years rather than decades, it's because they see a clear path to their goal. They may not know exactly how long each step will take, but they know the steps exist.
This means we could be looking at working prototypes of self-improving AI systems by 2027 or 2028. Maybe sooner, if Recursive's "undisclosed milestones" are as significant as implied.
Consider what that means for society:
- Existential risk: At some point, an AI system improving itself could reach capabilities that its creators neither understand nor control. This is the scenario that keeps AI safety researchers awake at night.
What Happens When the Builders Become Obsolete?
There's a deep irony in what Recursive is building. The company's founding team consists of elite AI researchers — people who have dedicated their careers to advancing the field. And their current project aims to make those same researchers unnecessary.
It's like a team of master craftsmen inventing a machine that builds better machines, knowing full well that once the machine works, their own skills become obsolete.
Perhaps they're betting that they'll retain control of the process — that they'll always be the humans "in the loop," directing the AI researchers even as those AI researchers become more capable than their human overseers.
But history suggests this is a losing bet. Once a technology exists that can perform a task better and cheaper than humans, humans tend to get pushed out of the loop. The incentives are too strong.
The question isn't whether human AI researchers will become obsolete. The question is whether they'll have any meaningful say in what replaces them.
The Uncomfortable Questions Nobody's Answering
As Recursive marches toward its goal of self-improving AI, there are questions that remain stubbornly unanswered:
Who controls the self-improvement? If an AI system can improve itself, who decides what "improvement" means? What happens if the system's goals diverge from human interests?
What happens to the "undisclosed milestones"? Recursive claims to have achieved things they can't yet talk about. What are they? How close are they to functional self-improvement?
Why $4 billion for a four-month-old company? What do GV and Nvidia know that justifies this valuation for a pre-product startup? What have they seen in private demonstrations?
What's the endgame? If Recursive succeeds — if they build AI that can design better AI without human intervention — what happens next? Do they keep iterating until they hit some ceiling? Do they sell to the highest bidder? Do they become the most powerful entity on Earth?
These aren't rhetorical questions. They're the questions that determine whether humanity navigates the transition to superintelligent AI or gets blindsided by it.
Your Move, Humanity
Here's the uncomfortable truth: You don't get a vote on whether self-improving AI gets built. Recursive has $500 million in fresh funding, a team of the world's best researchers, and an explicit mandate to automate AI development.
The train has left the station. The only question is whether we're building tracks toward something we actually want.
What you can do is pay attention. Ask questions. Demand transparency from the companies building these systems. Support policymakers who take AI safety seriously — not the ones who want to turn their jurisdictions into "open-air museums."
Because if Recursive succeeds, the world changes. Maybe gradually. Maybe suddenly. But it changes.
The AI that can build better AIs is coming. The only question is whether humanity is ready for what it brings.
And based on the fact that we're learning about this $4 billion company through leaked Financial Times reports rather than public discourse, I'm going to guess the answer is no.
We're not ready.
--
- Daily AIBite tracks the technologies reshaping our world — even when the people building them won't tell us what they're doing. Follow us for urgent updates on the AI race that will define the century.