All the stories I’ve read, even Gwern’s recent one feel surprisingly abstract. To me the obvious, very concrete story for an intelligence explosion looks like this:
Run a program that does the following: while true:
Run Codex on its own source with the prompt: “Improve the performance and efficiency of this coding model”
Train a new version of Codex using the modified source code.
Run tests and benchmarks to check it is actually better. If so, update your local version of Codex
Wait until it is amazing / you are dead
Obviously Codex isn’t nearly good enough to do this and you would need the benchmarks to include very difficult tasks, so that as it starts to take off it still has room for improvement. But I don’t see why it would require a different kind of model.
All the stories I’ve read, even Gwern’s recent one feel surprisingly abstract. To me the obvious, very concrete story for an intelligence explosion looks like this:
Run a program that does the following:
while true:
Run Codex on its own source with the prompt: “Improve the performance and efficiency of this coding model”
Train a new version of Codex using the modified source code.
Run tests and benchmarks to check it is actually better. If so, update your local version of Codex
Wait until it is amazing / you are dead
Obviously Codex isn’t nearly good enough to do this and you would need the benchmarks to include very difficult tasks, so that as it starts to take off it still has room for improvement. But I don’t see why it would require a different kind of model.