Recursive self-improvement: are the new billion-dollar AI labs building something beyond the capacity of businesses?
Investors have committed more than $1.7 billion to systems built to run autonomously for months or years. The organisations meant to receive them are not built to measure what they produce.
In The Hitchhiker's Guide to the Galaxy, Douglas Adams described a computer that designs a still greater computer, a form of recursive self-improvement. The machine was instructed to find the answer to life, the universe, and everything. Four decades after that fictional machine, investors have committed more than $1.7 billion in the last six months to two startups building systems designed to run autonomously for months or even years without human intervention.
Recursive Superintelligence, founded six months ago by former research leads from OpenAI, Meta and Google, is valued at more than $4.65 billion and has raised capital exceeding $650 million from investors including GV, Greycroft, Nvidia and AMD. Peter Norvig, who directed research at Google for twenty-five years, joined as an adviser. Ineffable Intelligence, founded in London in late 2025 by David Silver, formerly head of reinforcement learning at Google DeepMind, raised $1.1 billion in seed funding at a valuation of $5.1 billion, the largest seed round recorded in Europe. Its goal is a superlearner that acquires knowledge from its own experience without relying on human-generated data.
But these systems, if they arrive, will meet organisations that my research shows are not built to receive them.
A recent analysis of how 22 publicly listed organisations describe their AI programmes in prepared management remarks found 404 statements about adoption, deployment, usage, cost, efficiency and governance coverage. It found two statements about measured business outcomes directly attributable to AI reasoning or synthesis. It found 308 product and marketing claims. Both capability statements came from a single source, Alphabet. No other organisation reported a business outcome it could attribute directly to what an AI system reasoned or produced.
A parallel study of public policy documents across fifteen large organisations found that AI is governed universally through the technology and software regime. All fifteen classify AI alongside approved applications, cloud services, data handling, cybersecurity and acceptable use. The recurring vocabulary is technical and managerial: use, deploy, review, test, monitor, filter, secure, evaluate, approve, comply.
None of the fifteen governs AI as an external source of professional judgement, the way an organisation governs its lawyers, its auditors, or its outside advisers. No document in the sample requires a conflict declaration, an engagement letter, or a professional-judgement attestation when AI informs a strategic position. Two of the fifteen show this is deliberate. Oracle's AI terms require users to verify output independently and state that AI output may not be used as a substitute for professional judgement or advice. Salesforce prohibits the use of AI for individualised medical, legal or financial advice. Both drafters could see AI being treated as a professional source, and both wrote rules to prevent it.
Technology governance is treating AI as if it is software to be deployed, and monitored.
Richard Socher of Recursive Superintelligence has said his company needs years to build the underlying technology, which does not yet exist. The system these companies describe would generate hypotheses, design experiments, evaluate results and decide what to try next over periods of months or years, without returning to its human originator for permission. The open-endedness research field explicitly envisions software that pursues goals humans set at the start and does not return until the system stops.
Today's enterprises have built elaborate infrastructure for measuring distribution and almost none for measuring contribution. The recursive AI companies now raising capital are building machines to find answers. The businesses of the future that capitalise on those answers will need reporting that records what AI contributed and governance that treats AI as a source of judgement. Neither has been built.