Victor Taelin reports GPT models solve Erdos problems yet fail to spot basic fixes for Interaction Net bugs in an HVM SupGen variant

@VictorTaelin > you cannot even *express* this proof in Lean

is it one of those weird termination checks again?

@tribbloid no, it is categorically worse than that. it is more like real numbers not being able to express all roots of polynomial equations

@tribbloid termination aside, it is impossible to write a fully linear "toHOAS : LTerm → HTerm" function in Lean, due to this clause:

toHOAS d (LLam f) ctx = let (ctx,f) = toHOAS (d+1) (f (Var d)) (x : ctx) in (HLam \x -> f)

do you see the problem?

@VictorTaelin it’s not the same GPT.

Can you elaborate on what's going on here beyond termination concerns? Why do we need linearity?

It is true that it cannot prove certain things about a sufficiently powerful type theory in Lean (for the sorts of reasons you're describing) but, on the other hand, it is still very possible to prove things about some type theories.

@VictorTaelin well Lean does have complex numbers :D

how about posting the code or/and the error message? Lean positivity & termination constraints can be buggy and old sometimes

@VictorTaelin They take the good stuff for themselves and serve us the bare minimum.

If you were around for the launch of Opus and 4.5, 4.6, 4.7 there is always a very brief period of OMG THIS IS AGI.

It is fast, effective, no problem can get in its way.

Think about it...

@VictorTaelin That’s the problem statement part (+grading) of their pipeline. Clearly doing search.

@VictorTaelin @tribbloid Usually, you can implement HOAS in a proof assistant using PHOAS. Does that sort of approach not work in this case?

@curiousgangsta @VictorTaelin Yes, it is OpenAI internal model, and most likely also a custom harness.

@VictorTaelin I don't believe this. IMHO you need to get out of your own codebase for a bit. Have it make some specs in @leanprover. Have the benchmark suite look at buffer use and POSIX IO use.

Have you tried generating any sort of "gospel of interaction nets"?

I guess that is sort of your code base, but I did something for geometry like this that really helped for earlier models that struggled with S3, SO3, sort of stuff and higher dimensional toplogy.

Basically.I. copied in like 20 wikipedia pages that were key, had it run deep research and chase all the followups, and then had Pro write out a huge 1000 line markdown file full of all key equations and the reasoning for their connection. Than I would paste in this geometry bible into any geometry converation I was having.

@VictorTaelin >If this or a future model manages rediscover my solution independently, that would be the definite AGI moment to me.

I think this task based framing of AGI is wrong

@VictorTaelin I have to imagine that labs are RLing on your chats in particular as a user of interest, I wouldn't assume that your problems suddenly becoming in-distribution implies AGI

@VictorTaelin tbf tho, solving Erdos problems is pattern matching, fixing SupGen is actual engineering intelligence

@VictorTaelin el problema erdos fue resuleto por fuerza bruta no razonamiento no razono nada cogio una forma existente y solo la aplico pero no razono la forma ,simplemente saco el resultado a pura fuerza bruta ,de voy a aplicar 50 formas y la que caiga cayo

@VictorTaelin you are getting the giga nerfed system prompted no goblin mode, they are using unleashed full goblin

@VictorTaelin Models are nerfed right now many are pointing back to GPT 5.4. Also the model is likely not trying to solve it and only pretending is trying to solve it.

@VictorTaelin Add it to the training set