How to get o1 pro to write thousands of words

o1 pro likes context, and context can be generated from long text documents and dialogue.

I recently posted a 3,000-word essay, written primarily by o1 pro. When I say “primarily” here, I mean that o1 pro wrote around 95% of the words. Its original draft had some infelicitous phrasing which I cleaned up, and some formatting issues which I fixed, but aside from that: it was all derived from silicon. A few people asked me how I got o1 pro to write thousnads of words. This short post explains the process.

The process was the followng:

• I copied the transcript of Dwarkesh Patel’s interview with physicist and AI researcher Adam Brown and pasted it into ChatGPT, and selected the o1 pro model.

• I asked ChatGPT to parse the transcript and summarize all of Adam’s comments about AI. (Across the 2.75 hour interview, he spoke on a wide range of topics, aside from AI. I was only interested in his comments about AI.) Prompt: Here is a transcript of a podcast interview with Adam Brown, a physicist at Stanford. Can you summarize for me all of his comments about AI?

• Once I had that information, I further prompted o1 pro: I’d like you to speculate a bit about the implications of his observations. He is obviously a smart and well-informed person when it comes to the intersection of the state of the art in AI and the study of physics. What implications follow from his observations, again assuming that, given his position, his observations are directionally correct?

• After o1 pro answered that question, I asked a follow up question: Do these implications extend to other areas of science aside from physics? In some very rough, possibly facile, sense, physics drives chemistry, which drives biology, which drives medicine, etc. Does the promise of advancements in physics due to AI extend to advancements in other sciences as well?

• The next question I asked of o1 pro was: So is it reasonable to assume, then, that given the decade timeline Adam Brown sees for physics to be revolutionized by AI, we should also see broad advancements across all other sciences over the next decade?

• Finally, I asked o1 pro: Can you synthesize this dialogue into a 2,000-word essay? o1 pro ignored my length request and returned a 3,000-word essay.

The last question I asked contains part of the key here: in order to elicit a coherent 3,000-word essay from o1 pro, I engaged it in a dialogue, through which we created a lot of context. All told, over the course of the dialogue before I asked o1 pro to write its essay, we generated 34,000 words of context, including the podcast transcript. That’s 51 pages of text, give or take. Most of that context is the podcast transcript and the rest is my dialogue with o1 pro.

All of this is to say: o1 pro was able to generate a coherent 3,000-word essay because I provided it with a lot of context before I had it write the essay. And that’s the key, really: dump a lot of context into o1 pro before you have it do the thing you’re interested in. The podcast transcript was part of the context, but so was all the dialogue which preceded the essay. And that’s how I was able to generate a coherent 3,000-word essay from AI.