I think you have a point. The paper has load bearing reliance on other preprints. I think soon we see a workflow where AI (ChatGPT) can identify precise transitions in the argument that do not require full formalization to falsify. Link - https://chatgpt.com/share/693cc655-ca94-800c-870a-a5c78fb10d...
The mathematics here are far beyond me, but it's interesting that Gemini more or less concurs with chatgpt with respect to "load bearing reliance on other preprints".
Gemini's summary (there's much more in the link above that builds up to this):
The mathematics is largely robust but relies on heavy "black box" theorems (Iritani's blowup formula, Givental's equivariant mirror symmetry) to bypass difficult geometric identifications in the non-archimedean setting. The primary instability is the lack of a non-archimedean Riemann-Hilbert correspondence, which limits the "enhanced" atoms theory. The core results on cubic fourfolds hold because the specific spectral decomposition (eigenvalues $0, 9, 9\zeta, 9\zeta^2$) is robust and distinct22, but the framework's extension to finer invariants (integral structures) is currently obstructed.
Okay, wait though like I want to know the full transcript because that actually is a better / softer benchmark if you measure in terms of the necessary human input.
This comment seems misguided. I am a grad student working with LLMs in a CS department. I promise you CS has (and will have) much to say about theoretical underpinnings of transformers.
I can't see what you might be referring to. Even topics around quantization are fundamentally about information theory and not really computer science per se. It could just as easily be an explanation that your CS department is better-funded and more interested in tackling those problems -- statistics departments tend to be a bit behind the curve in this regard, but that doesn't magically move an entire branch of science somewhere else in the taxonomy.
Maybe if you can explain what you mean instead of merely providing some vague assertion, we could have a real conversation about it. In my experience, though, the hype train around software engineering has incited CS departments to retcon a lot of things as CS-related so that their grant writers have an easier time securing funding.
Is the pytorch convention that one or zero should be used for mask values you do not want to attend to?
ChatGPT will give a correct answer and sample implementation, if the implementation is broken or uses a non resistant api then I just tell it what's wrong and ask it if it knows better. Getting to the correct answer has proven vastly more efficient than wading through google results.
Alternatively if I'm asking a conceptual question, chatgpt will give a few directions, I can ask to dive deeper on one, and then I ask it for a citation or google search terms to confirm the result. If I fail to confirm, I'll tell it that and see if it corrects.
If I have some code to write - and I'm feeling lazy. I can just tell chatgpt my requirements and ask it if it has any questions. I keep clarifying the questions until it says it can write the code ( which it usually does with 95% correctness for non trivial asks) - if you tell it where it made mistakes it will usually correct, and if not porting the code to a working state just means changing a few function calls which chat gpt hallucinated into existence.
