
Carlos E. Perez
@IntuitMachine
1/n. I suspect this loop from 4o/4.5 -> Deep Research -> o1/o3 requires an explanation. 🧵
2/n First you incrementally refine your ideas in a non-reasoning model like GPT-4o or GPT-4.5 (or Claude, Grok, etc.). It's a discovery through conversation where you are trying to figure out what you want. You can leverage the following Pattern Language: https://intuitionmachine.gumro...
3/n Once you know what you're looking for, you can have 4o construct a query for you for Deep Research. Better yet, use the GPT from the book. Here's an example of that process: https://x.com/IntuitMachine/st...
4/n Now expect to get a huge dense document out of Deep Research. You won't have the bandwidth to read through it, but the size should be small enough for o1/o3 to work its reasoning magic. Have it pull in the facts from the Deep Research results and reason about it.
5/n An important thing not to overlook is that iterating through o1/o3 leads to a lot of lift. Do it through several loops as you continue to "design your knowledge." This notion of designing knowledge is described in the book: https://intuitionmachine.gumro...
6/n Now once you are happy that you've captured all the semantics that you wish to convey, send the results back into non-reasoning model to conjure up a digested "expository essay." Sometimes this is an important step to build an intuition behind what o1/o3 cranked out.
7/n Now that you've got you knowledge design, you can continue to have a conversation, refining ideas and potentially going through this loop again!