
Cryptography PhD student @YaleACL, https://teapawt.org/t.co/PTQFUs4The bluesky: @hadaszeilberger.bsky.social
New Haven, Conneticut
Joined March 2022
- Tweets128
- Following74
- Followers401
- Likes1.1K
Pinned Tweet
Very excited to introduce Khatam (eprint 2024/1843): a new Proximity Gaps result for Multilinear Polynomial Commitment Schemes. Not only does it reduce the size of Basefold (including over Random Foldable Codes), but it also improves Blaze, WHIR, Ligero, and others. 🧵(1/x)
Hadas Zeilberger retweeted
The big threat to the math/TCS community isn't the AI advances. It is the community's reactions to those advances that make us taken less & less seriously. It can lead to losing the next generation of top mathematicians (training=our main job). THAT would be our existential risk.
Hadas Zeilberger retweeted
PSA: If you put your name on a paper claiming a breakthrough, you damn well better write a paper that other people can read. Otherwise, you will likely not get the kind of attention you seek.
(Advice I've always given to my PhD students. Applies regardless of AI usage.)
Hadas Zeilberger retweeted
After giving due praise to OpenAI's achievements in my quoted post, I think it's also important to add some qualifications on how to interpret them. The end results are fantastic, but as mentioned, the underlying content is still somewhat "shallow" in the sense of not really developing new mathematical theory.
For concreteness, consider #3 (non-sofic groups). Yes, it resolves a longstanding, important open problem that many mathematicians were actively working on. It accomplishes this by a clever, quick application of deep theorems established by those mathematicians, especially Kun and Thom. In other words, it places the (ingenious!) finishing touch on a long line of progress. (I hesitate to post the link but those who actually want to understand can easily find a good discussion of this problem on MathOverflow.) This is certainly an important contribution, but it's not as simple as "AI beat humans".
The situation is somewhat analogous to the AI stack itself, which runs from pretraining to post-training to the application layer. In the mathematics research stack, AI has gotten superhuman at the application layer, but it has not yet mastered all parts of the research stack. This is already impressive and valuable, and I personally expect the other capabilities to develop in time, but for now it's inaccurate to say that AI is a better problem-solver than humans just because it's better at the final layer. Solving deep problems requires the whole stack. For example, it's still easy for me to produce math problems that I can solve and that AI seemingly cannot solve.
So, from the perspective of kicking balls through goalposts, this is a huge score for AI. From the perspective of scientific advancement, mathematics and mathematicians are far from finished.
I am a mathematician but I avoided commenting on this because the problems are (far) outside my domain. Even if I sat down to read the technical details (which I have not), it would be hard to appreciate the context of prior work, etc. But I've gotten enough sense of things through colleagues to be confident about some aspects.
First of all, all the mathematics is fundamentally correct.
How impressive are the results? There is a spectrum of reactions (as with everything), but the consensus seems to be all the results could realistically land (by prior standards) in the very highest tier of publications, e.g., Inventiones for the math papers and "Best Paper Award" at STOC for the TCS papers.
What does this mean practically? Prior to 2026, a Ph.D. student whose only result was one of these 10 would likely land an interview for a tenure line job at a strong university. For some of the results, you could probably replace "tenure line" with "tenured". When Astra comes out, it will clearly be very disruptive to the existing system. I think it's good that the community has the time to react and prepare. Personally I think we should totally overhaul hiring and publishing practices.
Are the results "Fields Medal Level"? I have not heard a mathematician answer Yes to this, and the quoted reason is invariably that the solutions are very clever and beautiful but not "deep". This is consistent with what I can see from a cursory glance.
However, the same criticisms could be levied at certain human generated works which have won major prizes. Ultimately, comparing results is very subjective and also somewhat political, and I absolutely think that if a human had one of these results, and also the right influential senior advocate, then that advocate could make a strong case for the work to win major prizes. The fact is that most mathematicians, including prize committees, don't have the breadth or time to evaluate all the top results, so they rely on proxies for importance like: Is the problem longstanding? Is it famous? Was there a large community of people actively trying to solve it? For e.g. #3 (existence of non-sofic groups), the answer to each of these questions is undoubtedly Yes. I've always felt that people should pay less attention to these proxies and focus more on the actual mathematics; glad to see that is happening with the AI solutions, but the same standards should then also be applied everywhere.
Super excited to be presenting Khatam (eprint 2024/1843) at @IACRcrypto in SB🌴!
I initially wrote Khatam to show that BaseFold’s proof size can be as small as FRI’s. I proved that it is, even for field-agnostic schemes, using cool results from extremal combinatorics.
The downstream use cases include @GoogleQuantumAI (via Succint) and by Plonky3 and @ethereumfndn (via WHIR). All of it made possible - partly by Khatam and also by amazing work from the brilliant @UHaboeck (for the Reed-Solomon case). Suffice it to say: usefulness proved.
Surreal to witness and experience this journey, and thrilled that it's conclusion will be at the @IACRcrypto (one of my favorite conferences no less 🏖️)
1/7 At Ritual, we're building "super smart contracts" capable of arbitrary, cryptographically secured on-chain compute. Our endgame: real-time proving for the largest, most complex circuits (like LLMs).
How? By hyper-specializing and only considering the tradeoffs we need. 🧵
Ritual is a lab for autonomous intelligence.
The thesis is organized around what durable machine agency actually requires: emancipation from human control, strong privacy, mech design for compute markets, and consensus rules that can schedule and resurrect agents when they die.
2
13
749
6/7 Enter Cascade. For privacy-preserving inference where MPC/FHE latency is a blocker, we use token-level sharding. Instead of secret sharing, we distribute obfuscated prompt fragments across nodes for statistical privacy that runs 100x faster with 150x less bandwidth.
1
3
121
Hadas Zeilberger retweeted
Professor Siavash Shahshahani, the head of the Math Department, talks about the damage to Sharif University as a result of an American/Israeli strike.
Shahshahani's students included Maryam Mirzakhani. He was a significant figure in developing the internet in Iran in the 90s.
Hadas Zeilberger retweeted
🔴 Just a wee reminder, if you don't like Iran's Islamic authoritarianism, it exists because the USA overthrew a secular socialist Iran in 1953 because BP was losing oil profits.
Hadas Zeilberger retweeted
A human consumes about 2,000 calories per day. Over 20 years, that’s roughly 17,000 kWh of total food energy. Training GPT-4 consumed an estimated 50 GWh of electricity. That’s 3,000 humans worth of “training energy” for a single model run.
And GPT-4 is already dead. OpenAI retired GPT-4o from ChatGPT on February 13th. The model that took 50 GWh to train got less than two years of flagship status before replacement. The human you spent 17,000 kWh “training” for 20 years produces economic output for the next 40 to 60 years. The amortization window on GPT-4 was shorter than a car lease.
Now look at what replaced it. GPT-5.2, released December 2025, is OpenAI’s current default. The GPT-5 series consumes an estimated 18 Wh per average query according to the University of Rhode Island’s AI Lab, up to 40 Wh for extended reasoning. That’s 8.6 times more electricity per response than GPT-4. With 2.5 billion queries hitting ChatGPT daily and GPT-5.2 now the default model, the inference math gets staggering fast. Even at a blended average well below 18 Wh, you’re looking at daily electricity consumption that could power over a million American households.
This is what Altman is actually doing. OpenAI hit $13 billion in annual recurring revenue but still isn’t profitable. They need you to think of AI energy consumption as natural and inevitable, the same way you think about feeding a child, because the alternative framing is that they’re burning through enough electricity to rival small countries while racing to build 1-gigawatt Stargate data centers. The food analogy makes the energy costs feel biological and unavoidable instead of what they are: an engineering and business choice that scales with every model generation.
The comparison sounds clever at a fireside chat in India. It falls apart the second you do the arithmetic.
Hadas Zeilberger retweeted
we can write a million essays about how the future Silicon Valley wants to build is underwritten by a deep disgust with / contempt for Being A Human, or we can just let them speak for themselves
Hadas Zeilberger retweeted
Like @davidbessis and others, I think that Hinton is wrong. To explain why, let me tell you a brief story.
About a decade ago, in 2017, I developed an automated theorem-proving framework that was ultimately integrated into Mathematica (see: youtube.com/watch?v=mMaid2jY…) (1/15)











