The lessons from o1

Lets strap a rockets to a dumpster and launch it into space.

Sam Altman is dancing. OpenAI’s new model o1 is showing massive promise.

It is performing exceptionally well on various benchmarks, including math, physics, coding, and reasoning tasks.

It has been trained via a new method. Using RL and tree search to selectively improve logical paths.

BUTTTTTT…

There’s a bit of a hick-up. In many cases you can proxy o1’s results with good prompting.

Essentially its an automated system running the following prompt over multiple inference steps.

o1 system prompt

1. Carefully read and understand the problem or question presented. identify all the relevant details, requirements, and objectives.

2. List all the key elements, facts, and data provided. Ensure that no information is overlooked.

3. Examine the gathered information for patterns, relationships, or underlying principles. Consider how these elements interact or influence each other.

4. Develop a plan or approach to solve the problem based on your analysis. Think about possible methods or solutions and decide the most effective one.

5. Implement your chosen strategy step by step. Apply logical reasoning and problem-solving skills to work towards a solution.

6. Review your solution to ensure it fully addresses the problem. Provide a clear explanation of your reasoning and justify why your solution is valid and effective.

7. Provide a short and clear answer to the original question.

Lesson

I think this means that you can - if your good - get a lot more out of your AI. Specialized prompts - painstakenly crafted can get really good results.

I seriously wonder how much of the content people consume today is AI generated. You can make pretty damn good content through a sequence of specialized prompts.