The Recursive Self-Improvement Claims Are Multiplying — and So Are the Reasons to Be Skeptical
On a single day, three separate accounts described AI systems that supervise their own training, rewrite their own research agents, and slash compute by orders of magnitude. The claims deserve scrutiny before they deserve alarm.
The most important story of the day is not any single breakthrough — it is the sudden clustering of recursive self-improvement claims arriving in parallel from multiple sources, none of them fully documented. Taken together they paint a picture of AI systems that are increasingly involved in building the next generation of AI. Taken apart, each one raises more questions than it answers.
The headline claim came from @RocaCap, who reported that OpenAI's GPT-6 Astra was the first model to use other AI systems to supervise its own training, yielding what the post described as a 70% improvement in token efficiency. That framing — AI supervising AI training — is exactly the kind of loop researchers have theorized about for years. It is also exactly the kind of claim that requires a technical report to mean anything. A 70% efficiency gain is significant if measured against a real baseline and marketing if measured against a favorable one.
Get our free daily newsletter
Get this article free — plus the lead story every day — delivered to your inbox.
Want every article and the full archive? Upgrade anytime.
No spam. Unsubscribe anytime.