Or Zamir (@OREAXEAX) header
Or Zamir (@OREAXEAX) profile picture

Assistant Professor, department of Computer Science, Tel Aviv University @TelAvivUni Formerly: @the_IAS, @Princeton

Joined January 2024
Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)Photo by Or Zamir (@OREAXEAX)
What do I think makes more plausible fine-grained hardness conjectures? Problems for which we know absolutely no useful structural mechanism, and can hope there isn't any. When we had non-trivial factoring algorithms, a poly quantum (or classical) one shouldn't be shocking. (1/2)
1
8
1,690
When there's a near-linear depth decision tree for 3SUM, a non-trivial algorithm is again not surprising. That also made SETH questionable, but other SAT variants seemed good. Even now then, it is still reasonable to cautiously conjecture hardness of Circuit-SAT, OV, etc. (2/2)
1
6
480
So what changes? Most works on exp-time algos are now obsolete (incl. some of my own), and much elegance in fine-grained complexity is lost. Many intermediate conjs still hold (eg, OV), but the slickness of "we have a central assumption that sounds like a strong NP!=P" is buried
2
2
50
4,959
To slightly calm down the external discourse: Breaking ETH does not directly imply anything about P vs NP, nor that every NP problem has a better algorithm; in fact, even for SAT (rather than k-SAT for a constant k) we still don't know if (2-eps)^n time is possible!
4
5
2
73
6,293
Oh wow!!
An internal model at OpenAI has produced a claimed refutation of ETH! Russell Impagliazzo, Mohan Paturi, and I have been asked to help organize an open and independent examination of this claim. 1/4
2
19
2,353
just click on "im not a robot" bro
2
482
Terry's analogy is indeed silly for math, and yet: If human researchers found a vulnerability in the drug trial process, e.g. how to modify proxy biomarkers but not the effect they are meant to predict, they wouldn't knowingly exploit it. Models based on modern optimization would
17
3
6
52
17,807
Without being whiny, we should rationally decide how to treat the OpenAI dump: Most results are likely true and important, but some are incorrect, and all are unreadable. So we can't yet view them as proofs, and we should make sure not to normalize this as a "publishing" method.
19
2
1
95
5,137
Or Zamir retweeted
The question of genocide in Gaza is not just a matter of “semantics,” as many are now claiming, quietly backpedaling from their earlier confident assertions. It is not a borderline case, or a matter of interpretation, or an issue on which reasonable people can differ. It is an obscene lie, contradicted by everything we know about Israel’s conduct during the war (which is not to absolve the IDF of war crimes). Here are 76 retired senior military leaders and legal experts flatly rejecting the genocide allegation: “Israel has not committed genocide but rather set a high bar for how to conduct urban warfare against an enemy who will stop at nothing, including the death of its own civilians.” The fact that the rector of my own university wanted to prohibit questioning the “Gaza genocide,” as though it were the equivalent of flat-earthism or climate denialism, still fills me with disbelief and rage. One day it will become clear who are the real flat-earthers here. jinsa.org/open-letter-from-m…
210
544
33
1,986
76,042
Concerning.
Replying to @ElliotGlazer
There also seems to be a bunch of formalizations that I found that have been weaker than the paper that i found using opus 5.5, i posted it on the leanprover zulip! Im not sure who to contact so the team could fix them :(!
10
63
10
593
37,486
I see dismay, rather than excitement, about yesterday's slop-drop, especially among students, sadly fanned by the more senior. For what it's worth, nearly all results on this list are surprisingly "within distribution" of what we'd expect seeing, just faster and far less readable
4
3
56
7,563
With, reportedly, 372 solved problems out of around 4000 tried, I think my earlier heuristic remains roughly accurate. Maths is very much not solved! (Though having a magical proof producing machine that works 10% of the time does change it a lot...)
FWIW, my current heuristic is that frontier AI has 5-10% chance of being able to solve or make substantial progress on any well-known conjecture. That's quite high! But also not. 1/
72
25
6
410
29,808
Oop. To me, the surprising ones in TCS are the DFT/Integer Multiplication improvements.
We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math
2
1
29
2,658
This week we mark 3 years since the #October7th massacre. TAU lost 124 students, staff, alumni, and family members on that day and in the ensuing wars. Together with the people of Israel everywhere, we mourn their loss among all those who perished. May their memory be a blessing.
6
104
10
356
17,913
(1/4) Awesome breakthrough from Anthropic+@pinkfloydie+@firebat03! Many other conjectures (e.g., SETH) are very plausibly also false. Nonetheless, here is a short thread about what seems to have made this problem especially "ripe for AI" (in hindsight). arxiv.org/abs/2610.06783v1
3
3
3
72
8,476
(3/4) The matrix multiplication improvement comes from a technical variant of the existing FMM algorithms, which are in turn also very technically complicated (and for many years, the SOTA ones relied on computations).
2
1
9
1,279
(4/4) This made this problem especially ripe for AI tools: what was needed was not a conceptual breakthrough, but a (big) technical one, on a non-clean problem formulation that was identified in past research. This is all said in hindsight, so can be taken with a grain of salt.
1
1
19
1,149
UK Greens, who polled at 20%, just cemented antisemitism and the objective of destroying Israel into their charter. Labour repeatedly copies their insane policies for votes. Open and proud Antisemitism across Europe, 3 generations after nearly eradicating its Jewry, is abhorrent.
1
557
Uploads to core TCS categories jumped, but not yet absurdly so: seems like +50%, consistent with the increase in conference submissions even pre-gpt5.6. Also visible in this graph: the labs may have solved a millennium problem, but not the inclination for absurd prompt-references
4
884