Agent swarm with Astra sets Riemann hypothesis optimization record
The author built an agent swarm for verifiable optimization problems and says GPT-6 Astra Pro pushed a record from 0.1875 to 0.1665 in 14 hours at about one-sixth the cost of the previous model.
doing my part for e/acc and @kardashev_ai. i've built a system that lets a swarm of agents hill-climb any problem whose solutions are cheap to verify. the heart of it is the bank. think of it as a ledger where every block is validated by independent re-execution and review instead of cryptographic consensus. over the past few months i've been running it solo, testing how far a swarm can scale while staying coherent. this led to a few record submissions pertaining to the riemann hypothesis, accepted by terence tao himself. (https://github.com/teorth/optimizationproblems/pull/126) for reference these problems sat unimproved for about 7 years and the last improvement before the swarm was by tao himself with @beffjezos and @mjdramstead now committed to kardashev as a counterweight to d/acc, it felt like the right time to bring the project back and see how far the models have come since fable 5 dropped: fable 5 (june): 0.2 -> 0.1875 in 3 weeks, ~$7,240 (~$580 per 0.001 drop) gpt-6 astra pro (this week): 0.1875 -> 0.1665 in 14 hours, ~$2,000 (~$95 per 0.001 drop) ~6x cheaper per unit of progress certificate + scripts to verify it yourself: https://github.com/463464q435q43/optimizationproblems/tree/dbn-lambda-01665/certificates/certified1665 why i'm bringing this to beff's attention: i believe that for the first time in history, most open problems are within reach of frontier models, but the only ones getting solved are the ones frontier labs care about. if the e/acc community rallies behind this, we can procedurally solve everything in between. im thinking of creating some kind of prize fund to incentivise people to target the lowest possible scores on any problems that exhibit qualitiies like these optimisation problems in the name of e/acc