astroButter

Autoresearch from @karpathy runs 1 experiment at a time. We gave it 16 GPUs and let it run them in parallel. 8 hours. 910 experiments. 9× faster to the same best result. The most surprising part: the agent had access to both H100s and H200s. Without being told, it noticed H200s scored better (more training steps in the same 5-min budget) and started screening ideas on H100s, then promoting winners to H200s for validation. That strategy just emerged on its own. A human researcher can grab a cluster and run experiments in parallel. The agent couldn’t. It was stuck with 1 GPU, greedy hill-climbing, ~10 experiments/hour. We built a @skypilot_org agent skill that teaches coding agents to manage their own GPU clusters. The agent reads the skill, then launches clusters, submits jobs, checks logs, and pipelines experiments on its own. With that, Claude Code provisioned 16 GPUs on Kubernetes, ran factorial grids of 10-13 experiments per wave, and covered in one 5-minute round what sequential search takes six rounds to do. The biggest finding: scaling model width mattered more than every hyperparameter trick combined. The agent tested 6 width configs in a single parallel wave and found the winner immediately. Sequential search might have missed that entirely. Total cost: ~$300 compute + $9 in Claude API.

@3LettersZain רוסים שמנסים להגיד הרצליה

🚨🇮🇱 ISRAELI SURPRISES ARE THE BEST KIND PM Netanyahu vows an all-out campaign against Iran’s regime, saying Israel has “many surprises” planned to destabilize it. “We have many more targets.. the moment of truth is approaching.” 😲 Stay connected, follow @MOSSADil.