Anthropic shows AI that trains itself — and it actually works

An Anthropic researcher just published a paper on an AI system that writes its own upgrades. Given ten benchmarks for misaligned behaviors, the system improved on every single one without degrading overall performance. That last part is the kicker: most self-improvement attempts make things worse on the other tasks. This one didn't.

The key is a loop called AutoRFT. The model gets a new test, generates a fix, then trains on that fix — all without a human hand. It's a step toward the kind of autonomous agent Anthropic has been hinting at for a while, and it lands right in the middle of a race between Anthropic, OpenAI, and Google to build systems that can revise their own weights. The paper is thin on the exact training recipe; what matters is the pattern: a model that reads its own failures and patches them faster than a human could.

Why this matters for us: the faster these systems improve themselves, the faster they'll leak into the apps and services our families use — and the harder it becomes to audit what they've changed.