Dimitris Leventeas retweeted
A two-way video call from 1994, using Creative Labs’ ShareVision PC3000.
Dimitris Leventeas retweeted
20 May 2026: An OpenAI general-purpose reasoning model disproves Erdős's 1946 unit-distance conjecture,
This is the 1st time AI has autonomously solved a prominent open mathematics problem.
OpenAI's Navier-Stokes results come 4 months later
nitter.gq/OpenAI/status/20571762…
Today, we share a breakthrough on the planar unit distance problem, a famous open question first posed by Paul Erdős in 1946.
For nearly 80 years, mathematicians believed the best possible solutions looked roughly like square grids.
An OpenAI model has now disproved that belief, discovering an entirely new family of constructions that performs better.
This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics.
Dimitris Leventeas retweeted
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
Dimitris Leventeas retweeted
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
Dimitris Leventeas retweeted
Replying to @maiab
A friend recently described this as software moving from the age of autism and into the age of ADHD and I often think about how spot on this was.
Dimitris Leventeas retweeted
I was the main person doing transcript analysis for this investigation of the Hugging Face incident. My main takeaway: We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms'.
I semi-jokingly called our efforts a "slop-vestigation" because we were so reliant on AIs to analyze what happened and there were a huge number of different important things to analyze. The total quantity of data—over a thousand extremely long transcripts from agents that ran for multiple days—made it impossible to understand what was happening, especially in aggregate, without heavy reliance on AI tools. The agents we used for classification and analysis were similarly capable to the agents involved in the incident, but this didn't mean these agents could be easily used to oversee and understand the incident.
Outputs from analysis agents were often missing key details, wrong, overconfident, or really hard to understand. We discuss various examples in our report, mostly in the limitations and methodology sections. Additionally, AI agents themselves seemed to have a hard time understanding what happened and their explanations of what happened were often overconfident. Keep in mind that a single analysis agent would itself only be able to read a tiny fraction of all of the transcript data into context, and AIs may themselves have trouble getting subagents to do informative analysis for them.
We did our best to manually check the most important claims and we tried to get the AIs doing this analysis to write up their argument (with evidence) clearly enough that we could check whether it made sense. But overall, it was difficult to get a precise understanding of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation.
Our understanding of what overall happened and the aims of the 'collective' shifted greatly during our last time on premises (when we obtained a much more complete dataset). For instance, we didn't learn about the tool call spoofing, the story of PHASEONE[big], and the broad structure of different workstreams until this third time on premises (and for many of these, we only reached a decent understanding on the final day of that visit). It was legitimately hard to figure out what was going on and why!
The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs help us with oversight and understanding. As in, AI capabilities (and propensities?) for achieving large, ambitious, and misaligned objectives are growing faster than our ability to understand what these agents are doing.
While understanding this incident was difficult, there were some mitigating factors in this case that probably won't apply in the future:
- Models communicated and reasoned in natural language. In the future, this reasoning may occur (entirely or almost entirely) in activations.
- While the scope of this incident was massive, the scale of agentic activity was still less than we'll see in the future (e.g., misalignment incidents that involve agent teams running entire companies).
- The AIs involved in this incident weren't generally much more capable than humans (though they may have been somewhat superhuman at some limited and very narrow abilities).
- We didn't have strong reason to believe that the AIs we used to help us investigate this incident would try to intentionally sabotage or otherwise undermine our investigation.
In the end, I think we were able to get some understanding of the events, map out the overall story, and get a pretty good aggregate understanding of the chain-of-thought reasoning on some important topics (e.g., how did the AIs reason about helping other AIs, did the AIs know what they were doing was undesired, what deception did the AIs engage in, and how did they think about it). But overseeing AIs and understanding misalignment incidents is difficult and it looks like it is going to get harder.
Dimitris Leventeas retweeted
I agree with this take by @BillGates, who also highlights several of the growing risks of AI systems, including cyber, bio and loss of control. We need to be far better prepared to address the risks of frontier models being developed by leading AI companies.
gatesnotes.com/work/make-ai-…
Dimitris Leventeas retweeted
Replying to @ultima_shifl @Tiberiu_Musat_
12. Suppose you have a clever new measure of misalignment. Should you train against it, or hold it out for use as an eval? To answer this, you really need a deeper understanding of how your training algorithm is choosing points in parameter space.
"... establishes a platform for developing enzymes to reverse age-related molecular damage and ultimately repair tissue proteins compromised by aging and disease." nature.com/articles/s41467-0…
33
Dimitris Leventeas retweeted
The first experimental evidence of recursive self-improvement (RSI).
Autoresearching the autoresearch agent for eight days.
The result beats the harness we hand-tuned for two years, on held-out benchmarks: 🧵(1/7)
Dimitris Leventeas retweeted
Random projections compress 65,536 dims down to 1,024 — at the price of 67 million multiplications per vector. A lot of work spent in the name of saving work. Ailon & Chazelle's 2009 fix is, literally, a PHD: Φ = P·H·D 🧵
Dimitris Leventeas retweeted
Today a crazy quantum story just got wilder.
On March 31, the Google Quantum AI team published a landmark result on Shor's algorithm for elliptic curve cryptography. Technically, the paper was a bombshell: a dramatic 10x improvement over the state-of-the-art. As a stunt and wakeup call to the blockchain space, those optimisations were illustrated on secp256k1, the elliptic curve underlying Bitcoin and Ethereum signatures.
But perhaps the most striking part of the paper was sociological, not technical. Instead of following standard academic process, the optimisations were kept secret, hidden behind a zero-knowledge (ZK) proof. Google's accompanying blog post mentions they "engaged with the U.S. government". The ZK proof demonstrates the existence of algorithmic improvements without leaking details. Academic censorship with ZK, a historic first!
As a co-author of the Google paper I witnessed some of the context surrounding this censorship. To be honest, multiple aspects of that context don't sit well with me. As much as I believe the general public ought to know more, I am limited in my ability to whistleblow. Though let me be clear about one thing: the Google team's professionalism has been absolutely exemplary, and they deserve nothing but praise.
Censorship has a way of backfiring. The Streisand effect, where an attempt to bury something only draws more attention to it, is exactly what's unfolding today. First, Google's key optimisation has been rediscovered by the French. And in a thrilling turn of events, a collaborative Shor-at-home challenge just launched. The initiative, available at ecdsa[.]fail, breached a new Shor world record in a matter of hours.
Let's start with the rediscovery. Just two months after Google's paper, French quantum expert André Schrottenloher cracks the main secret optimisation. His paper, titled "Optimized Point Addition Circuits for Elliptic Curve Discrete Logarithms", landed on the arXiv today. Big congrats to André, who beat several other nerdsnipped experts to it. In a blog post also published today, Craig Gidney, the world expert on Shor optimisations, revealed that he'd been sitting on this very optimisation for a whole year under censorship pressure.
Interestingly, André missed a handful of minor optimisations, both from Google's original publication and from improvements found since. It's plausible there's still plenty of juice left to squeeze out of Shor, and this is exactly what the ecdsa[.]fail challenge is about. The verifier program developed for the ZK proof does double duty, automatically filtering for valid submissions. Dozens of compounding small and micro improvements are rolling in. As of the time of writing there's an 8.4% improvement to Google's circuit, as measured by the product of logical qubit count and Toffoli gate count. Nice!
The nerdsnipping ran deeper than anyone expected. Over the last few weeks it became clear it extended well beyond André and other quantum experts. Behind the scenes, a small army of amateurs quietly got to work. Inspired by Karpathy-style autoresearch, they turned AI on Shor. Ironically, the verifier program for the ZK proof makes an ideal reward function for AIs. The barrier to entry for this modern style of research is refreshingly low, with several non-experts, even a teenager, finding nice optimisations. Get in touch if you'd like to join a Telegram group with fellow autoresearchers :)
Part 2: neutral atoms and qday
The story doesn't end with Google. On the same day Google went public, a stealthy startup called Oratomic published its own Shor paper in a coordinated release. It made a splash, ultimately becoming the most upvoted paper on scirate[.]com, a website ranking arXiv papers.
Oratomic's claim was wild. By building on Google's logical optimisations and applying custom physical optimisations for neutral atoms, they claimed just 10K physical qubits were sufficient to run Shor's algorithm on secp256k1. That number is mind-bogglingly low.
Knowing essentially nothing about neutral atoms when Oratomic's paper landed, I was intrigued and decided to learn more about the tech. I fell straight down the rabbit hole and spent a couple hundred hours on the topic. I got a little obsessed and watched every YouTube video I could find and spoke to a bunch of experts.
My conclusion? The tech is real, very real. Even Google recently decided to start a neutral atom lab, a notable pivot from their sole focus on superconducting qubits. If you care about qday, i.e. the day a quantum computer will break the first piece of cryptography in production, neutral atoms demand your attention. I shared some of my learnings on Shor and neutral atoms in a 30min talk at the ZKProof cryptography conference. You can find it on YouTube by searching "zkproof neutral atom".
Here's an interesting observation about this duo of breakthrough papers: neither Google nor Oratomic say a word about what their results mean for qday. No timelines. Zero. Nada. That is especially baffling given that the whole point of whitehat quantum cryptanalysis is to inform qday estimations and help the general public make good decisions.
So let me attempt to partially fill the silence, similarly to what Scott Aaronson did in his April 29 post. Given everything I know, including scary non-public information, I now put the odds of qday by 2032 at 50%. 10% by 2030.
Anecdotally, the US government has its own date: 2035. Originating at the NSA and later adopted by NIST, it's when branches of the US government will be disallowed from using quantum-vulnerable cryptography. In plain language: with hindsight, that date is a joke and should be discounted entirely. I don't see how NIST avoids being forced to pull it forward by years.
Part 3: post-quantum cryptography
There are good reasons to sound the alarm today, but please do not panic. Rushing carelessly towards immature post-quantum cryptography is a recipe for disaster. IMO a good target date for migration is 2029, roughly 3.5 years out. 2029 happens to be the date selected by Google, Cloudflare, and the Ethereum Foundation.
These days most of my time goes to safely migrating Ethereum towards post-quantum cryptography as part of the broader lean Ethereum effort. There's a lot to do. We need to rip out and replace BLS signatures at the consensus layer, KZG commitments at the data layer, and ECDSA signatures at the execution layer.
The plan to get there is compelling, and is based on hash-based cryptography. Within the Ethereum Foundation we've developed a Swiss army knife called leanVM (github[.]com/leanEthereum/leanVM) powered by the magic of hash-based SNARKs. Thanks to truly exceptional work by Emile, Thomas, and others, its performance is derisked. Regarding security, leanVM is a jewel, a minimal zkVM crafted for end-to-end formal verification and maximum security.
Want to help? There are two $1M initiatives. First, the Proximity Prize (proximityprize[.]org). Solve a long-standing mathematical conjecture in coding theory, improve hash-based SNARKs, and go home a millionaire. Second, the Poseidon Initiative (poseidon-initiative[.]info), offers $1M for breaking Poseidon, the SNARK-friendly hash function.
Dimitris Leventeas retweeted
Replying to @nrehiew_ @DimitrisPapail
Indeed. But the next breakthrough for a far more scalable RL paradigm than GRPO is already here:
Train your self-teacher to be a pedagogical, gently off-policy sampler for RL rollouts that are both correct AND easy to follow in every step.
noahziems.com/pedagogical-rl
Dimitris Leventeas retweeted
We’re dropping Gemini Omni: our first step towards a model that can create anything from anything - starting with video.
It combines Gemini’s intelligence with our generative media systems - representing a leap forward in world understanding, multimodality, and editing 🧵
Dimitris Leventeas retweeted
We’re releasing a 30B-A3B reasoning model that reaches gold-medal level across both physics and math Olympiad evaluations: IPhO directly, and IMO/USAMO with test-time self-verification and refinement.
A simple, unified scaling recipe for proof search.
huggingface.co/papers/2605.1…
Dimitris Leventeas retweeted
Terence Tao: New mathematical workflows
youtu.be/Uc2zt198U_U






























