Is O1 a New Model?

OpenAI’s new o1 model, released yesterday into preview, is impressive. It seems to have the ability to do some new things:

• Explain it’s planning.
• Read and adjust its own responses.

We’ve seen this kind of thing before, in tools like langchain, but they’ve never worked this well.

I’m curious whether o1 is really a single “new model”, or whether it’s a new set of tooling that actually runs a series of requests back through the model until its happy with response.

You can do this yourself now. In ChatGPT, ask the model to make a plan, ask it to execute parts of the plan, feed the response back into it, and you can get it to hone in on a very good response, but it’s the human that’s guiding it.

The real innvoation here may be the tooling to orchestrate this process.