#23: Embedded reasoning for ChatGPT

Knock-knock. Good morning. Here comes another AI breakthrough.

So, OpenAI has released O1 preview and mini, and in short, I’m thrilled. The model "thinks" before responding and solves complex tasks significantly better than previous models.

Last night, I decided to test it by writing a service in Go with its help, which collects statistical metrics from GitLab. The model succeeded. More precisely, it first built the project structure, filled in the files with content it suggested, and left placeholders where it couldn't write. Then, I iteratively asked it to handle each metric individually. No issues at all. Some work had to be done manually (like copying content between files), but it was very straightforward.

I specifically chose Go, a language I know little about, so I wouldn’t interfere with the model’s process using my own experience on how things should be written, focusing purely on "I can read the code in any language," while the model should write it itself.

How does it work? According to the documentation, OpenAI first trained an internal reasoning model using reinforcement learning (where the model succeeds or makes mistakes, receives feedback, and adjusts its weights). The internal model is designed to output reasoning tokens based on input tokens (essentially a preliminary response in the form of reasoning about the query, an embedded chain of thoughts). Then, the input plus reasoning generate output tokens. Tokens from previous reasoning aren’t involved in further communication to avoid overcomplicating things.

It’s already available for those with a subscription, and I highly recommend giving it a try. There are limits, but mini should be enough for writing code (limit of 50 messages), while preview has a limit of 30.

For general users: https://openai.com/index/introducing-openai-o1-preview/

For tech enthusiasts: https://platform.openai.com/docs/guides/reasoning