Seeing Anthropic’s self-improving AI demo, and it gives you 10 misaligned behavior benchmarks—each one gets fixed by the system itself. My first reaction wasn’t excitement, it was more like that feeling of finding at midnight that a coworker has refactored your code and hasn’t opened a PR yet.
After coding for ten years, I’ve gotten used to finding bugs myself, fixing bugs myself, and then writing tests to prove it’s correct. Now the model just changes the weights directly, without even a git diff. In the future, debugging might shift from “why is this line wrong” to “what exactly did it change.”
Feels like I’ve gained an unspeaking teammate
After coding for ten years, I’ve gotten used to finding bugs myself, fixing bugs myself, and then writing tests to prove it’s correct. Now the model just changes the weights directly, without even a git diff. In the future, debugging might shift from “why is this line wrong” to “what exactly did it change.”
Feels like I’ve gained an unspeaking teammate