LLMs are adapting their environments to themselves (ianbarber.blog)

  Post-training Claude on a particular shape of edit tool doesn’t just make Claude better on that edit tool, it encourages harness authors to support that shape. It propagates that shape of edit tool. That itself leads to more traces which use that edit tool, which propagate that technique into otherwise unrelated models. Claude isn’t just adapting to its environment, its creating it: constructing a niche where edit tools bend towards its preferred approach, despite that approach being, as Ronacher complains, undocumented.

What exactly is the difference between this and saying that the C programming language adapted its environment to itself by making its ABI a de facto standard? Or JavaScript adapting to its environment by making JSON ubiquitous? The ecological metaphor, like any biological metaphor about LLMs, is worse than useless.