Simon Mackenzie retweeted
The fall of the Riemann hypothesis has revealed an even BIGGER mathematical challenge, easily worth trillions:
Characterizing what it means for a complex program (e.g., Lean proof) to be human-cognitively irreducible, and developing techniques & tools for automatically simplifying it.
For the first time in God-knows-how-long, the biggest challenge facing mathematicians is aligned with the biggest challenge facing technologists. I'm so excited.
There's a subset of mathematicians who really need to read up on what lean is before they post their opinion about the RH result on X.
1
1
46
Simon Mackenzie retweeted
Also, I wish people would stop calling it the "Claude proof", as I am seeing on this website, and instead credit the person involved: Levent Alpoge.
Levent is not just an Anthropic employee, he is a top-level mathematician. Others with similar access to frontier AI models/tokens (and trust me, there are many such people) were not able to come up with a proof of RH.
Levent has been on a truly generational run since summer, proving time and again his ability to solve conjectures thought out of reach. Jacobian conjecture, Hadamard matrices, complex structure on S6, etc. And now GRH.
I think the mathematical community, a large part of which is anti-AI, has treated him poorly, at least from what I've seen on social media.
The honour of proving RH, if the proof holds up, belongs to him.
Simon Mackenzie retweeted
In lesser news, Tristan and Levent claim to have proved blow-up of UNFORCED Navier-Stokes
Simon Mackenzie retweeted
Astra 6.2 found an expected O(n^25) algorithm for uniform Latin Square sampling.
No polynomial-time algorithm was known before.
Research notes and paper: github.com/Mira-Cult/math/tr…
I keep reading academics defending AI disclosures. All I'm seeing in practice is a majority of people downplaying the role of AI in their work, sometimes blatantly so. How is this adding any value? Is the hope to still reward work that was human generated as if it had been done 2 years ago? Is it to better assess people? Effectively it is basically asking people to self-report how much credit they think they should get which is guaranteed to produce dishonest outcomes.
8
610
Simon Mackenzie retweeted
For maths, my first Astra 6.2 impressions are positive. For a bunch of bounds where all the recent models were making no or epsilon progress, Astra got quite a bit further, seemingly effortlessly. Comprehensibility hasn’t improved but also hasn’t worsened (on first sight).
Simon Mackenzie retweeted
I've spent a good deal of the last two weeks working as an early tester of the new Astra model. This will be a longer post about my experience as a working scientist and mathematician, so stay put because this will be a wild ride.
First thing - I always use Astra in agentic mode (my preferred version is CLI mode in my terminal). I like seeing what is going on with the model (all the gory details), keeping control and seeing all the instructions which are executed. New Astra is a very perceptive model and keeps doing things incredibly fast and efficiently. Keep that in mind. No YOLO mode. Always stay a bit alert.
Capabilities: for mathematics, especially mathematics driven by hardcore computations, this is a killer model. I usually start with a very specific auto-research task, driven by the labyrinth skill which I developed, trying to figure out the constraints for what makes the particular theorem possible and why. This model is a heavy weightlifter. It requires patience. It is very common that once the setup, preps and first few runs are done, the model goes on a research hunt for a FEW HOURS with a simple "Carry on". Yes, most of my runs in the cycle were about 1–4 hours each. This is a lot of reasoning, a lot of grinding and a lot of new ideas coming through.
Partner: this is a model which I use both in conversation and in agentic loops. It can discuss very fine details of research with you and never get derailed or distracted. Doing things is very natural once you master the work mode. You should focus with this model (I prefer Max over Ultra) on a particular task. It's capable of going very deep. In my work on hypergeometric motives (all minimal polynomials of finite ones), the model proposed a very mature strategy. That was the work of a very capable mathematician.
Speed: when you turn on fast mode or lower the effort, you get a smart agent which is amazing for day jobs, but the most powerful version of Astra 6.2 comes from working in the Max tier and then it gets a bit slow. It processes a ton of data and its own chain of thought. But the output mostly works in one-shots.
Writing: the writing skills got much better with this model. I finally like reading the production drafts which the model produces. It works best with a specific prompt, but it's much, much better than Astra 6.
Agentic: this is what I got the most excited about. I was classifying semigroups of order 7 (almost a million objects), rewrote Mizar into a fully modern kernel with modules and tactics, and got a breakthrough in my research on elliptic divisibility sequences. All of this worked because the agentic orchestration and division of work were relatively natural. You need to work hard on this part yourself - the vision of what you want to achieve. The model won't eagerly set up a ton of agents for you, but you can trigger it to do so and even make communication channels between them. Powerful thing.
Lean: this is where I find this Astra really cool for maths. Once you have a ton of AI slop with great ideas, doing a Lean formalization attempt is a piece of cake. If it's harder than expected, Astra will point you to the hard parts and let you focus on checking the theorem formalizations. For my semigroups project, it produced a few million lines of reusable and very structured Lean code in two weeks. With AI-generated maths, I think Lean is a necessity.
Overall, this premiere comes with no bells and whistles, but it's a powerful tool for power users. Sol is a tiny bit less capable but is a much faster and cheaper agent. It's my go-to tool for everyday work. Both combined make a totally awesome tandem.
Replying to @OpenAI
GPT-6.2 Astra—a frontier model built for your hardest projects. It sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT‑6 Astra.
You’ll see the difference in things like slideshows, websites, and 3D designs.
IMO, the current challenge for mathematicians/TCS researchers is to map out the interesting frontier, identify where AI models still fail, make these challenges salient and public, and also design incentive systems for people to engage meaningfully in them.
More importantly, I think we need to rethink how research itself is conducted. Make problem-solving more dynamic, collaborative, and public. We should encourage sharing partial results, failed approaches, and emerging ideas as they arise. This will allow others to build on them in real time rather than waiting for the traditional cycle of polished papers and conference vetting. That cycle has become too out of sync with the reality of research.
The current incentive systems overwhelmingly reward polished, private contributions, made often after months or years of private development. Insisting that discovery be bundled with all the subsequent work of polishing, digesting, and presenting it creates an enormous burden, diverting time and attention away from further discovery. Not everything needs to be digested. The important parts will be, and that is valuable. A researcher can selectively target what they want to understand, what they need to understand for further discovery.
Discovery, verification, and exposition are all important. The issue to me is requiring them to be bundled together. Papers as the primary deliverable of research no longer works in my opinion.
3
461
Simon Mackenzie retweeted
"Ignore the Writing: A Proposal for Peer Review in the Age of AI."
arxiv.org/abs/2610.38062v1
The exponential barrier for cake cutting has been broken! @mashbat6 and I recently released a preprint with the first polynomial bound on the number of expected queries.
Following subsequent improvements from Alireza Farhadi and @MTHajiaghayi, we've now managed to get the bound down to Õ(n⁵). Let's see if anyone can get it lower!
Õ(n⁵) draft: github.com/SimonWMackenzie/e…
It is becoming clearer and clearer to me that many people are borderline obsessed with gauging papers on the authors' perceived merits rather than the scientific value of the paper itself. I think we need to move away from this way of thinking.
1
124
I think there is a real schism forming in tcs/mathematics between the conception of mathematics as a fundamentally human-centric endeavour vs mathematics as a collective epistemic enterprise expanding the boundary of what is known.
1
3
182
I think the mentality of litigating how much AI was used to generate a paper is ultimately futile. I believe soon it'll all be heavily AI assisted with humans steering more or less well. The romantic era of mathematics/TCS is over but so many new possibilities are emerging.
2
118
Simon Mackenzie retweeted
New paper with @SimonWMackenzie & @mashbat6:
Sampford’s apportionment is the 1st known rule to satisfy threshold monotonicity, quota, and exact expected proportionality.
papers.ssrn.com/sol3/papers.…
#apportionment #socialchoice















