On AI

by Kei Kebreau — Thu 12 November 2026

Myths are only as true as the worlds their bearers create, and the bearers of the myth of "Artifical Super Intelligence" have created quite a hell of a world. I remember trying to use GPT-4 as a first-year PhD student to explore what should have been basic questions about algorithms and logic. I immediately ran into "hallucinations". Just two years later, overreliance is much more my concern. I used tools like ChatGPT, Claude, Gemini, and co. in my work to bypass the tedium of reviving and running old academic machine learning code. My productivity seemed like it went up, and I would read papers or think about experimental design while AI monitored my experiments and analyzed my data. But some things about using these tools made me feel icky and uneasy.

Aside from the very real and unnecessary labor, political, environmental, and socioeconomic externalities, I couldn't inspect or control the systems I depended on. It's difficult to be properly responsible for such a system. If my AI agent injects someone else's intellectual property into my codebases or writing, am I responsible for copyright violation? Can I trust my AI agent to not hack me or whoever the agent is writing code for? If I'm using a big closed model like Fable, is this AI agent really even mine to begin with?

Perhaps the answer, insofar as we're all going to be forced to use LLMs, lies in using more transparently developed models like Olmo or Apertus. For now, I know that the producitivity increase is not generally worth it for me in my personal work. It remains to be seen how well this all will fare for software engineering, let alone knowledge work as a whole.