I built a tool to analyze ChatGPT's "Thinking" and JSON code to uncover the topical map of its brain.
I built a tool to analyze ChatGPT's "Thinking" and JSON code to uncover the topical map of its brain.
See more here: https://x.com/arcknight_tech/status/1988821849212416409
Interesting angle. I’ve been testing multi layer reasoning too, but more on the execution side. Have you noticed patterns in when the model deviates from its initial plan?
Thanks! I do notice themes and patterns but it depends on the keyword/phrase and how much the search intent aligns. If the keyword doesn't have much demand, it tends to be wonky and all over the place with less themes present
Interesting. So it’s basically tied to how strong the underlying structure is in the data.
When intent is low, the model has to “guess” the themes and the map becomes noisy.
Have you found any prompt or preprocessing tricks that reduce that drift?