The argument about recursive self-improvement is only half right.
Here’s where it gets interesting.
You can’t take a naked transformer, and ask it to improve itself in the abstract. Ask a model to reflect on itself, or ask two models to work on a project together without the right tool set, and you’ll get signal degradation, drift, and total collapse.
Also, if you're token maxing, it's an easy way to hit the top of the leaderboards.
Remind me to tell you the space shuttle drink holders story sometime. I’m sure I’ve posted about it before.
Anyway, that’s why you have to be careful when you’re training with synthetic data, and why it’s so easy to corrupt a good dataset with bad information.
External verification changes the loop
Agentic engineering systems like Claude Code work well because they provide external feedback. The agent imagines code, the code is tested by real compilers.
Similar agents will use a tool to access your documents, email, or social media feed, but it's the same theory.
The model may be sampling from its learned distribution, but that’s not all it’s doing.
In other words, once there’s an external verification loop, the “same distribution” argument stops being the whole story.
Recent discussion of Anthropic’s position does center on AI-generated code and possible “recursive self-improvement,” including reports that Claude now writes around 80% of Anthropic’s merged/internal code.
Using Claude Code or similar harnesses, a strong enough model can now take on projects at the complexity scale of major framework rewrites, compiler work, and infrastructure migrations.
But if the agent is improving the harness, then the loop has changed.
At that point, we’re not really talking about code anymore, so much as we’re now thinking about building machines that build better machines that build better code.
If that better machine is part of the process used to build the next generation of models... I dunno, it feels silly to start talking about chatbots getting high on their own exhaust.
But that's what they mean.
Recursive improvement at the systems layer.
And I don’t think anybody’s over-promising on that.