This week, there has been a great deal of discussion about a new piece from Anthropic on how AI systems are starting to build the next generation of AI systems themselves. More than 80% of code merged into Anthropic's codebase is now written by Claude itself.
Taken to its conclusion, this points to a world where AI systems can fully design and train their own successors. The technical term is recursive self-improvement. The honest description is harder to absorb.
Anthropic's authors are clear-eyed about what this means. The pace of AI development is accelerating beyond what any single organisation, government, or treaty body is prepared for.
They argue, reasonably, that the world should have the option to coordinate a slowdown. But they also acknowledge the hard part: any slowdown only works if you can verify that other labs have actually slowed. And training runs, unlike missile silos, are remarkably easy to conceal.
This makes the problem a coordination problem, and it is worth re-reading the AI 2027 forecast in light of it -
ai-2027.com.
AI 2027 paints two endings to the same starting conditions:
- In one, well-coordinated actors keep the technology aligned with human interests.
- In the other, a race dynamic between bad actors and cautious ones produces a future none of us would choose.
The difference, in large part, is whether mutual verification was ever possible.
Verification at this scale cannot be entrusted to a single party. If the body that monitors compute usage at AI labs is controlled by one government, or built on one company's cloud, then it can be coerced, captured, or quietly switched off. The infrastructure for AI safety has to be at least as decentralised as the threat it is trying to constrain.
This is where Aventus Cloud comes in.
While we did not build it for AI safety, we did build it as a decentralised compute layer for situations where you must trust as little as possible and verify as much as you can. A network of thousands of independent nodes, with no single party able to turn it off, could host the verification layer for agreed compute thresholds without any single body holding the underlying data. The watchtower mechanism we already use to verify Merkle roots between chains is structurally the same problem as verifying that a frontier lab has not exceeded an agreed training threshold.
A decentralised cloud built to be contestable by design will not solve AI safety on its own. But verification has to live somewhere, and the place it lives matters.
That choice might be the difference between the two endings of AI 2027.
Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.
It’s happening faster than we thought, and the implications deserve greater attention.
anthropic.com/institute/recu…