E8

ML researcher/mathematician

[E8 @_196883]

@deanwball If open weight is so bad for AI capex why is Nvidia pushing perhaps the most open source of all models (eg they publish basically everything about how Nemotron was trained, much more than Chinese labs), it’s almost like open weights/open source is good for Nvidia

[E8 @_196883]

@swombat Retrovirus-Like Intercellular Signaling association with the ARC gene also complicates the simplistic model of neurons

[E8 @_196883]

@TheStalwart The answer is that it’s possible but no one can be bothered for production systems. More specifically it’s possible for an LLM to have superhuman chess intuition, but you’d still need to RL it into planning properly

https://t.co/UZSao3RK0x

[E8 @_196883]

@NoahChrein I suppose more the point is that the “algebro-geometric” perspective of higher category theory is in my opinion a better direction than the model category/simplical perspective philosophically speaking

[E8 @_196883]

@NoahChrein I personally believe Lurie went in entirely the wrong direction, I’d say the correct “higher topos theory” perspective was from moduli theory in the direction of Diligne, and Grothendieck’s “Teichmüller stacks”

[E8 @_196883]

@giffmana @vanishingideal @ChShersh Ngl anyone who lists Mathematica is not a serious SWE

[E8 @_196883]

@littmath I suspect any real attempt would be making claims about the standard conjectures rather than going for Hodge directly

[E8 @_196883]

@veryfinesalt @miniapeur I suppose what I meant to say is that channels like standupmaths cover “unserious” mathematics, in the sense that it’s given in the tone of a high school class, whereas 3b1b feels like it’s aimed at a first year audience, that is to say it has the feel of university mathematics

[E8 @_196883]

@lisatomic5 My guess is that this is total bs, and the actual reason is that he wants to force his engineers to solve vision so it scales to cheap robotics

[E8 @_196883]

@karpathy Economists already realised this as limitations of utility theory, hence the creation of behavioural economics

[E8 @_196883]

I think the right way to think about it is that humans have fields of “aspirations and concerns”, one dimensional utility functions are simply not expressive enough to capture this, perhaps the ML community shouldn’t have sidelined Schmidhuber as hard as it has (although his erratic behaviour has certainly been unwarranted)

[E8 @_196883]

LLMs are actually not great at Hodge theory currently, they basically need hints to connect to different areas and don’t really have a good intuition for this. But what I’d say is that in the future this is somewhere where bandwidth ends up being a huge factor, for instance being able to connect super deep results in harmonic analysis and GMT with say Shimura varieties is something where the initial limiting factor is basically just understanding all the areas deeply enough to connect them

[E8 @_196883]

The best comparison I would say it to institutions, it’s like say “the Harvard math department as a black box is more intelligent than Grothendieck”, I mean sure, but they’re not really comparable, I think LLM intelligence is much better compared to institutions than to individuals

[E8 @_196883]

I think the math and ML community has in general over indexed on problems humans find difficult, for instance LLMs have a huge advantage over humans in terms of bandwidth, for deeply connected areas like Hodge theory, no human has the time to learn absolutely every area it connects to at an expert level, the bandwidth advantage will almost certainly lead to “superhuman” math ability at some point in the future, but it’s a sort of trivial brute force advantage that I suspect most will not feel constitutes AGI even if it outperforms humans in this domain

[E8 @_196883]

Or maybe you’re just a schizo who doesn’t understand finance, blackrock basically just manages retirement funds, blackstone is a PE firm that does invest in residential real estate, SWF’s are generally more influential in global finance than PE firms or asset managers like blackrock due to sovereign immunity. Most of blackrock’s money is tied up in ETF’s so they don’t actually have much liquidity, that is to say they don’t actually have much choice in what they do with their money unlike PE and SWF’s, “big number” does not necessarily mean “most influential”

[E8 @_196883]

@loganvasquez126 @alz_zyd_ @httr9k Also quantum consciousness theories have very little backing in academia for good reason, perhaps if you took the time to actually learn more about physics rather than deferring to others you’d understand why

[E8 @_196883]

@loganvasquez126 @alz_zyd_ @httr9k Would you say that AIXI is true reasoning?

[E8 @_196883]

@loganvasquez126 @alz_zyd_ @httr9k Only people with either a basic knowledge of math, or a basic knowledge of ML would deny it at this point. Most math PhDs are equally incapable of true originality just like LLMs, so I think it’s fair to call gpt-5-pro “PhD level”

[E8 @_196883]

I’d say gpt-5-pro is almost immediately useful for harmonic analysis, getting it to be useful for algebraic geometry (particularly for deeply connected areas like Hodge theory) requires extensive interdisciplinary intuition which LLMs currently do not possess, but this can be prompted easily