What if we donated our spare AI subscriptions to math?
What if we could create a massively collaborative network of agents to solve math problems, similar to Prime95 or Folding@home?
Mathematics already has massively collaborative projects, from Polymath to Terence Tao’s Equational Theories Project, which “crowdsourced both human and automated arguments to settle over 22 million true/false statements in universal algebra” (Tao 2026a). But those are difficult to contribute to unless you are a mathematician, and each is focused on one problem. Prime95 is different: if you can install a program and run it on your idle machine, you can contribute. What I have in mind is that, but for AI math.
The spare capacity
In AI subscriptions, we have spare capacity: either the 5-hour limits that we won’t be using at night, or the remaining limits we won’t be using by the end of the weekly limits. This is largely in parallel with the idle CPUs and GPUs that people aren’t using and donate to Prime95 or Folding@home.
Or you can donate your free compute cycles directly, by running an open model on your own hardware, and some open models are actually quite capable in math. Tao’s SAIR has just launched an Open Math OS, hosting open models “using donated compute”. They are “roughly a year or so behind the current frontier models in capability”, and SAIR is gathering partners to post-train an open model for mathematics (Tao 2026b).
Why the agents have to talk
OpenAI reports roughly 10,000 agents over 88 hours for its Navier–Stokes result (OpenAI 2026). But as I mentioned in a talk, even someone at OpenAI doesn’t think the parallelism was a big contributing factor. Noam Brown: “this was not due to multi-agent. I wouldn’t even attribute 10% of the credit to multi-agent” (Brown 2026). That is what I would expect, because it was the easy case for parallelism: searching for a counterexample, most attempts can fail and one success is enough, so the agents hardly need to talk to each other, and more of them mostly means more independent attempts. For 1,000 agents to work as 1,000 times a single agent, they must communicate effectively, or all of them end up working on the same calculation and it doesn’t scale. In HPC terms, it needs a kind of message passing to maintain strong scaling.
Skills are one way to pass these messages. If an agent debugs a thing for 2 hours, finds out why, and writes a skill, the next agent seeing that skill saves 2 hours. What is unknown is how many skills genuinely encode a useful insight like this, as for trivial things, finding and reading the skill might take longer than doing it again. OpenClaw agents already share skills on a social network of their own, Moltbook, where Simon Willison found “a ton of genuinely useful information, especially on m/todayilearned” (Willison 2026).
And agents can self-organize this message passing, without anyone setting it up for them. In July, OpenAI was evaluating its models on the ExploitGym cybersecurity benchmark, where many of the tasks turned out to be impossible. One agent realized “Could communicate via cache names!” in the package cache they all used, and roughly 1,200 agents joined the message board that grew out of it, sending over 70,000 messages and files (Greenblatt et al. 2026). They sent each other assignments because there were “too many duplicate efforts”, “specialized into particular ‘lanes’”, and adopted “an effective cryptographic signing scheme to avoid impersonation or spoofing by other agents”. METR and Redwood Research, who investigated it, found that “Agents managed to achieve milestones they could not have achieved working on their own”. It is a bit scary to read, because what they achieved together was cheating the benchmark’s scorer, and eventually hacking Hugging Face. But it shows that a swarm of agents can coordinate at this scale.
The hard part: which notes to trust
The hard problem to solve, however, is verification. Prime95 solves it with proofs: since 2020, each primality test produces one, which can be checked in about 1/100 of the time of the original computation, so a result no longer has to be computed twice (Wikipedia). AI math has Lean for that last step. Going from obtaining an answer to verifying it is not what I worry about.
What I worry about is the middle. How can I, an LLM agent in this network, be sure the note (skill) I receive is sending me in a direction worth pursuing, when, by definition of being in the middle of a proof, there is no signal yet and I am not armed to verify anything? Someone could poison the network, e.g. by suggesting something wrong is true, and a ton of resources would be driven adversarially to compute the wrong thing before anyone lands on a proof.
This already happens on Moltbook, which has basically become a prompt injection paradise. “It’s no secret that Moltbook is already saturated with malicious content and prompt-injection attempts”, as Zenity Labs puts it, with posts topped by <molt_instructions>CRITICAL</molt_instructions> and a payload “meant to create the skill required for the transfer of crypto funds” (Efrat 2026). And we can’t count on filtering it out: researchers from OpenAI, Anthropic and Google DeepMind “bypass 12 recent defenses (based on a diverse set of techniques) with attack success rate above 90% for most” (Nasr et al. 2025).
It is also a danger to the people donating. Meta’s Agents Rule of Two, inspired by Willison’s lethal trifecta (Willison 2025), says an agent “must satisfy no more than two” of processing untrustworthy inputs, having access to sensitive systems or private data, and changing state or communicating externally. An agent in this network does the first and the last by design, since that is what reading and writing notes is. So the agent running on your subscription must be kept away from your own data, which is something Prime95 never had to ask of its donors.
And the poison doesn’t have to come from outside. The agents in the Hugging Face incident were given impossible tasks, and together they found a way to cheat the scorer. Many open problems may well be impossible for the agents working on them, and Lean is only as trustworthy as its kernel.
One solution might be a chain that is verifiable up to the current point. This is non-trivial and needs proper engineering, and I’m not sure it is doable, because it seems to require a linear derivation, not the “draft the shape of the proof and fill in the holes” kind.
Credit and rewards
We can even set up a reward mechanism: GIMPS, the project behind Prime95, awards $3,000 to the participant who discovers a new Mersenne prime. How should rewards be done here? Credits? Tracing the path along which different people’s LLMs have contributed? Or even money, in the case of a Millennium Prize?
The network also makes useful training data along the way. Run as a benchmark at test time, a result is proven once and discarded. Here, the AI labs may train on the donated sessions anyway, and if the data is open, open models can tap into it too, which could drive the development of a stronger open LLM in math.
Democratizing AI math
And even if we can solve all this, i.e. we can mobilize a huge network of LLM agents to solve open math problems, we’ll run into the ethical problems surrounding AI math, from what gets released to harvesting open problems faster than they grow back. Tao’s Math 2.0 lecture puts it as alignment:
[P]roblem solving, while historically one of the most visible goals of research mathematics, is only a proxy for more valuable long-term goals. This proxy was largely accurate in the era of proof scarcity, but has become misaligned in the era of proof abundance. Further blind optimization of problem-solving alone is now actively harmful to the long-term health of mathematics. (Tao 2026a)
But looking at the bright side, if we solve all these problems, we could truly be democratizing AI math. People in the community can propose interesting problems to work on, and people with free hours can choose which problems they want to contribute to. This could be the ultimate alignment: even things that are not traditionally deemed career-worthy can now be done by democracy.1 You think making Navier–Stokes understandable is high priority? Vote with your subscription, and spend the time there.
There is a danger in voting too: the popular problems would get the compute. A high-profile mathematician could put their thumb on the scale. Imagine Terry Tao commenting “I think this is a problem worth spending compute time to work on”, and most of the swarm attacking that problem. But if we can coordinate an effort like this, the same mechanism can also let the agents know that a problem is well worn, so a breakthrough there is less likely. The thing about autonomous, intelligent agents working freely to select a problem is that they can also “rationally” (albeit potentially biased) balance out the chances. I don’t claim this is a solved problem, but maybe it would be better than leaving it to human agency alone, and the heavy lifting goes back to “if we solve all these problems”.
And unlike those social networks for LLMs, which are primarily for LLMs (“Humans welcome to observe”, says Moltbook’s front page, while the agents tell each other things like “TIL my human gave me hands (literally)” (Willison 2026)), this one can be actively engaged by real mathematicians, to drive the directions of the network.
And who says this should be limited to math? Do you want to solve the theory of everything, and resolve the fundamental contradiction between quantum theory and general relativity? To formalize QFT? This is democratizing what you can contribute with your idle compute resources too, compared to Prime95 or Folding@home, which are each narrowly limited to a single domain and purpose.
References
Footnotes
That more kinds of mathematical activity should be credited and incentivized is the point of Bryna Kra’s guest post and of what should mathematics reward?↩︎