Victor Taelin reports GPT models solve Erdos problems yet fail to spot basic fixes for Interaction Net bugs in an HVM SupGen variant
The required HOAS interpreter step was never suggested unaided.
@VictorTaelin That’s the problem statement part (+grading) of their pipeline. Clearly doing search.
Taelin@VictorTaelinGPT is able to solve Erdos problems but still not come up with simple solutions on Interaction Net programming... I left 5.5 fixing a bug on a SupGen variant overnight and it failed. Obviously it did: the solution requires writing a HOAS interpreter on HVM, and doing so is physically impossible unless you have a key insight that allows it to go through. It is a very beautiful idea that completely reshapes how you think about the domain and the fix itself is 2-3 lines long. Of course, I just taught it the solution so it can fix the bug it had. But I wonder if it could find that on its own. I really with I could set up this experiment with whatever model solved Erdos. If it could rediscover my solution independently that would be one of the most shocking moments of my life
#1047Taelin@VICTORTAELINGPT is able to solve Erdos problems but still not come up with simple solutions on Interaction Net programming... I left 5.5 fixing a bug on a SupGen variant overnight and it failed. Obviously it did: the solution requires writing a HOAS interpreter on HVM, and doing so is physically impossible unless you have a key insight that allows the proof to go through. The reason this idea is so beautiful is that it is as "out of the box" as it gets; literally, you cannot even *express* this proof in Lean because it requires questioning the nature of type theory itself. Yet, the solution itself is 2-3 lines long. I think it was one of the most creative solutions I found in my life. Of course, I just taught it the solution so it can fix the bug it had. But I wonder if it could find that on its own. Would be really cool to set up this experiment with whatever model solved this Erdos problem. If this or a future model manages rediscover my solution independently, that would be the definite AGI moment to me. I'll make an extra effort to not publish it
Taelin@VictorTaelinthis is super cool but I still do not understand how they get a model to coherently and usefully reason for that amount tokens and at this point I'm to afraid to ask