Alexander Wissner-Gross says AI can now edit other AI directly

Alexander Wissner-Gross, a computer scientist who founded Reified, said on the podcast Moonshots that AI can now edit other AI directly. He spoke about new MIT research in an episode posted Wednesday.

Why you'd care

For most people this is about how fast AI improves and who controls it. If one AI can fix or add skills to another without starting over, new abilities could reach the apps people use sooner. It could also make safety limits easier to strip out.

He described a paper from MIT's computer science and AI lab. An AI edited 0.016% of the weights of a small open model, the numbers that set how it behaves. The model had been stuck in endless thinking loops. Its math score rose from 7.9% to 48.1% without retraining. Wissner-Gross said he thinks the work matters a great deal, though some people pushed back after it came out. He expects AI to help explain other AI. He said at 1:05:11, "The machines will understand the machines."

The subject is in the news today. Yahoo ran a story headlined "Together Brings Recursive Self-Improvement To Open Models". It reports that Together AI is building what it calls an agentic compiler. The aim is to bring recursive self improvement to open source models. That means AI helping to build better AI.

Wissner-Gross named a downside. Someone could remove the guardrails from an open model that anyone can download. He said that wouldn't work on closed systems like ChatGPT. He called the result very early and small scale. He expects the benefits to outweigh the risks, which is his own view.

Watch the 8 minutes where he explains it.1:01:57