“Reproducing the performance of the previously best open-source model costs ~$400 of compute.”
Allen Institute drops open-source coding agents that match proprietary models for $400 in compute. Built largely by one researcher. The 32B model fine-tuned on 8,000 samples beats 100B+ parameter teachers on specific codebases. This is how you commoditize AI. Open the weights, publish the recipes, let anyone train on their own code.