4dimcube (@4dimcube) header
4dimcube (@4dimcube) profile picture

Professional circle packer. Does math sometimes.

Joined July 2018
Photo by 4dimcube (@4dimcube)Photo by 4dimcube (@4dimcube)
Replying to @QiaochuYuan
I think a lot of the usefulness comes down to the structure of the proof. If P vs NP was solved with a polynomial time algorithm for a NP-complete problem that might turn the world upside-down, but a proof of P!=NP just tells us we can trust all of the work that assumes it.
1
60
I've known Justin for the better part of 4 years through some pretty awesome chatrooms for discussing frontier ML/AI research, and I think his philosophy and approach are going to make waves. Super excited to see him come out of stealth here!
This quoted post is unavailable.
3
263
Replying to @jxmnop
Reminds me of Kingma's dissertation. Develops deep insight, drops VAEs and ADAM in a blitz of fantastic papers, goes on his merry way.
1
151
To be clear I agree wholeheartedly with your initial point, I just see scaling challenges that strongly decrease the odds of an "intelligence explosion" relative to something more like "optimal but still incremental/polynomial progress on heuristics"
19
Replying to @VictorTaelin
I think this requires some deeper consideration that I haven't yet done, but the way that you describe utility of functions here sounds either uncomputable (via Kolmogorov Complexity) or at least NP-Hard (via reduction to something like generalized circuit-SAT).
1
29
Replying to @staysaasy
This, but we hit stage 4 by Q3 2026
97
Replying to @VictorTaelin
I lost vision in my right eye from the chiari complications and then again after the meningitis, but thankfully it came back (to my severely myopic status quo) after the intracranial pressure subsided both times. Sorry to hear that, can definitely relate to human body being shit.
94
The meds required to keep you alive during an acute meningitis presentation (e.g. steroids to drop intracranial pressure, narcotics to get you to stop screaming) also tend to make you feel *awful* afterwards, at least until you can get back to an even keel.
31
Replying to @VictorTaelin
N of 1 here, but I can tell you that it tends to get dramatically better as the primary symptoms subside. Had meningitis as a complication of brain surgery; almost died but didn't. Swelling and inflammation in the brain or spine feel fucking awful but are thankfully temporary.
1
1
117
Replying to @cloneofsimo @FAL
I don't think this is correct; the original is about comparing an expectation to a single known value (the weight not the sample). Not seeing how that follows from here but maybe that's on me 😅
394
Or in other words, the answer to your question is "yes"
1
5
Replying to @getjonwithit
The emperor has been standing naked in our sprinklers at 3:32am every third Tuesday of the month for at least a few years now. (LLMs really suck at extrapolation because it's underspecified. A decent rule of thumb is "can the output I want be interpolated from training data?")
2
1
63
Replying to @MilesCranmer
It's interesting, for sure. Some counterexamples too, in Go and Starcraft iirc. I'm skeptical of the "foom" sort of superintelligence for more basic computational reasons, but it's interesting to consider consequences.
1
1
87
In this case that seems unlikely; fake data to support weak (or more likely entirely fabricated) papers is far more probable.
26
Replying to @jxmnop
Other people have already said various smart things so this is just another reminder: 🥒=🙃
52
Yeah, it's all about quantum effects related to the available electron orbitals. More charge in the nucleus means it pulls harder on the electrons, but more electrons means they repel each other (due to charge and degeneracy pressure). Noble gases have full shells so they're big.
2
81
Atomic radius isn't about mass, it's about configuration! Helium has a smaller radius than hydrogen, but both of them are larger than average. Most transition metals have similar or smaller radii.
1
1
77
To be clear: I think you can build effective models that do not explicitly express a posterior over the input space, but I contend that internally the model must capture this information if it is complete. It *is* a generative model, just without a decoder in the input space.
1
1
188
I agree with @ylecun on a lot of things but this seems off: by the data processing inequality, if a system "understands" (captures) an input then it *must* contain the information required to generate that input. Perhaps the point is that you don't need a generative architecture?
Replying to @ylecun
Video generation systems will get better with time, no doubt. But learning systems that actually understand physics will not be generative. All birds and mammals understand physics better than any video generation system. Yet none of them can generate detailed videos.
1
2
374
Replying to @ylecun @balazskegl
This strikes me as an interesting hypothesis. I'm led to wonder if the experience of dreaming is even uniform across people - for example, might those with aphantasia have a radically different experience?
51