If you'd asked me last year to run an autonomous research loop across two GPUs, I'd have said that's not something I can do.
Not "it'll take a while." Impossible.
Today this is my Saturday.
My - - : autonomous hyperparameter mutation, parallel experiments, no babysitting.
Results:
→ 17 experiments, 0 crashes
→ baseline 1.2365 → . _
→ 1.48% improvement. Found by the loop.
The chart shows the staircase, the same pattern Karpathy sees in his runs. His: 2 days, 276 experiments. Mine: 1 hour, 17. Same logic, different constraints.
The constraint here is the ' _=, which means ~5.5% MFU. The GPU is mostly waiting on memory transfers, not computing.
What the same loop looks like on different hardware:
( ): ~ / → .%
: ~+ / → -%
(): ~+ / → -%
(): ~+ / → -%
Same autoresearch loop. Just more runway.
Karpathy's setup, 8x H100 running 48 hours would likely hit 5,000+ experiments. On 4090s that's not feasible. But 24-48 hours on what I have would still find significantly more than 1 hour did. That's what's running next.
This is my own multi-GPU implementation built on top of his single-GPU original, orchestrated with iii functions, workers, and triggers.
Claude Code made the extension possible in a weekend.
Not "it'll take a while." Impossible.
Today this is my Saturday.
My - - : autonomous hyperparameter mutation, parallel experiments, no babysitting.
Results: → 17 experiments, 0 crashes → baseline 1.2365 → . _ → 1.48% improvement. Found by the loop.
The chart shows the staircase, the same pattern Karpathy sees in his runs. His: 2 days, 276 experiments. Mine: 1 hour, 17. Same logic, different constraints.
The constraint here is the ' _=, which means ~5.5% MFU. The GPU is mostly waiting on memory transfers, not computing.
What the same loop looks like on different hardware: ( ): ~ / → .% : ~+ / → -% (): ~+ / → -% (): ~+ / → -%
Same autoresearch loop. Just more runway.
Karpathy's setup, 8x H100 running 48 hours would likely hit 5,000+ experiments. On 4090s that's not feasible. But 24-48 hours on what I have would still find significantly more than 1 hour did. That's what's running next.
This is my own multi-GPU implementation built on top of his single-GPU original, orchestrated with iii functions, workers, and triggers.
Claude Code made the extension possible in a weekend.
If you want to try: https://github.com/iii-hq/n-autoresearch open source repo.