Meta details AIRA3 autonomous AI research system
AIRA3 coordinates isolated agent pairs via an asynchronous forum and shared filesystem, placing 8th out of ~4,000 teams in an NVIDIA fine-tuning competition.
- Used an ensemble of GPT 5.5 and Claude 4.8 to fine-tune NVIDIA's 30B Nemotron reasoning model in a live Kaggle competition.
- Coordinates long-running model and coding-harness pairs asynchronously without a central controller, relying on a shared forum and filesystem.
- Delivered a 27% latency reduction on production GPU kernels during internal benchmarking.
- Achieved gold-level performance translating 4,000-year-old Akkadian clay tablets into English by updating only task specifications.
AI researchers gain proof that decentralized multi-agent systems can autonomously optimize models and GPU kernels without human-guided search strategies.
