What Is Going On Inside OpenAIs Strawberry (o1)?

This week OpenAI released a bombshell of a model or a bunch of models from the series Orion (possibly that’s the full name). This project was code-named Strawberry and the models are called o1-preview and o1-mini. More are soon to follow. Now that we have more details, it’s time to go a bit deeper into its possible architecture. Please note that OpenAI has not released a paper, all we have is a model system card. With a bunch of early testing, results, and the information present in the model system card, we are going to try and understand what’s new in this model and why this is a big deal for all the upcoming models. This marks a momentous shift, and the models coming from other companies will follow soon.

So, without further ado, let’s begin.

Table Of Content

• What Exactly Changed In o1?
• Is It A Glorified Chain of Thought (CoT)?
• How Do We Know o1 Is Using CoT?
• What Is Happening In Inference Time Compute?
• Is It AGI Finally?
• Opinions On o1 Seems Quite Divided
• Conclusion

What Exactly Changed In o1?

I’m sure all of you had the experience where if your initial prompt was not good, the entire trajectory the model takes to answer kinda sucks. But then you change the starting prompt and it reaches the correct answer. Do you know why?