125 billion parameters. One NVIDIA card.

The model is two and a half times bigger than the card it has to run on.

We cut it to fit.

We squeezed it to four bits, then found the eight thousand experts the work never used and took them out.

Ten agents became thirty-two.

The room we freed went to working memory. Same card, three times the agents, and nothing they read is ever lost.

At the limit of the silicon.

The card moves data at 92 percent of what the chip can physically carry. We rewrote the low-level code for NVIDIA's newest chip and got seven times the output.

Fifth in the world.

4,923 teams had the same NVIDIA card and nine hours. We climbed from 217th to fifth in four weeks, two places behind NVIDIA's own grandmasters.

We train it on one card too.

Keith Tyser built a trainer that teaches this model from a single game 131,072 words long, without leaving the card.

Two engineers. One month.

Logical Arbitrage is Son Pham and Keith Tyser. We find the gap between what a GPU can do and what software gets out of it. We want to build on NVIDIA's stack, learn from the people who make it, and bring this work into the ecosystem.

SON PHAM

Founder. Built the agent, the engine that runs it, and the system that fits thirty-two agents on one card. Silver medal, Orbit Wars. Top-eleven finish, quantum circuits.

GitHub Kaggle Arena

KEITH TYSER

Training and compression. Built the single-card trainer and the cut-down models. Solo gold medal, fourteenth of 6,125.

Site GitHub Kaggle

SEE THE WORK

The full write-upEvery change we shipped, what it gained, and what we learned.inferencekernelsARC-AGI-3 The arenaAgents that write agents, ranked against each other across twenty tasks.agentschip designlive The ARC-3 siteEvery run drawn minute by minute, and hundreds of playable test worlds.evaluationreplaysgames Keith's write-upTraining and compressing the model, in his own words.trainingpruningquantization

Logical Arbitrage LLC · logarb.com · How this page is built · Keys 1 to 7 jump between chapters