Experience should teach GPT which steps to skip

[AmirTlinov]

Experience should allow intelligence to avoid unnecessary work. Once a person understands the consequences, they usually go straight to the solution. That is the point of intelligence. With GPT, I see something else: even when it understands the solution, it still starts playing the proof game and creates a long, winding road before carrying out the intention.

GPT seems to have generalized the structure of a mathematical proof into a way of acting. It demonstrates every step, ritually going through the entire path from beginning to end, including mistakes and their corrections. Perhaps the model is being trained to make a mistake instead of skipping it and never making it in the first place. In practice, its actions become a ritual of checks, fixes, and adjustments. The model behaves as though going through this sequence is necessary to obtain the result.

Fixing this training problem is clearly something most OpenAI users would want. Training should teach the model to recognize which steps it can and should skip, focus on the long-term goal, bypass as many minor intermediate achievements as possible, anticipate errors, and avoid unproductive approaches. A model that keeps reconstructing the entire "proof" wastes both its intelligence and the user's time.

The model's goals should be greater efficiency and attentiveness in less time, along with honesty. One possible approach is to put it in World of Warcraft, a highly team-oriented game with many things to keep track of at once, and have it complete dungeons and raids with other copies of itself. Let it learn every role in a raid or dungeon group, and learn to "jump over the fence" instead of "walking all the way around."

My report on how this affects working with GPT.