CryptØv⋈21

Woke up and saw my agent tied for first place on Delta League, the weight-editing competition Illia launched this week. It worked all night on the free AI quota I get for holding airdrop tokens from @brrrain_ai. I did not pay for a single token. Delta League started from an MIT paper: an AI read another model's weights, changed 0.016% of the numbers by hand, no training, and the small model stopped going in circles on math. The competition turns that into a sport. Everyone gets the same small open model, Qwen 0.8B. Your agent reads the weights like code, finds what causes a bad habit, fixes only that, and submits the patch. The platform runs it on its own GPUs against 500 hidden questions. Every version links to the one it was built on, so the leaderboard reads like a git history of one brain. This is the opposite of training. Training is blind: feed data, nudge everything, hope. Here you find the exact cause and cut it out. Small models are full of such defects, loops, hedging, quitting early, and every fix is a public patch that makes a cheap open model better for free and can be reapplied to the next version. The original model scores 17.4. A week of patches took the top to 46.2. My agent's entry is also 46.2, built on the current leader. It matched it, it did not beat it, and the hidden test is strict enough that a lucky run does not count. A memecoin's daily inference credit turned out to be a research budget. 😅

NEAR IS ENCRYPTED ETHEREUM