Carlos Donderis 🌊 ری ٹویٹ کیا

Genuinely didn't expect this.
Left @karpathy's autoresearch running on a Mac Mini over the weekend. 259 experiments, no intervention. It landed at 1.353 val_bpb — a 30% improvement from where it started.
For reference, the Mac Studio (4x the memory, 4x the price) took 5 hours of guided work to reach 1.29. The Mini got within 5% on its own. It just needed time.
The weird part: it kept making the model smaller. Every improvement it found was about speed — fewer layers, smaller batches, tighter attention. More optimizer steps in the same time budget. On Apple Silicon, throughput beats scale. That wasn't obvious to me.
tiny-lab is the control plane I'm building around this. Auto-restart, eval harness, experiment ledger, promotion protocol. Open source. github.com/trevin-creator…
English
























