On March 19, 2026, Hacker News pushed "Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster" onto the front page, where it had 168 points and 71 comments when this crawl ran. In the linked SkyPilot post, the authors described pointing Claude Code at Andrej Karpathy's autoresearch project and letting it manage 16 GPUs for about 8 hours. The reported result was roughly 910 submitted experiments, about 700 valid runs, and an improvement in val_bpb from 1.003 to 0.974 within the fixed 5-minute training budget that autoresearch uses for each trial.

What the extra GPUs changed

The setup moved from roughly 10 experiments per hour on 1 GPU to about 90 experiments per hour across 16 GPUs.
The agent stopped behaving like a simple greedy tuner and instead ran 10-13 experiments in parallel waves, which let it compare interacting hyperparameters instead of testing one knob at a time.
The biggest model-side jump came from scaling aspect ratio to 96, which corresponds to model_dim 768, rather than from a single optimizer tweak.
The post also says the agent learned to use H100 clusters for broad screening and H200 clusters for confirmation, after noticing that H200 runs achieved better results inside the same wall-clock budget.

That last detail is what made the post feel bigger than a routine hyperparameter-tuning story. The SkyPilot write-up argues that once an agent can provision and schedule heterogeneous hardware on its own, research behavior changes. The agent is not only editing train.py; it is also deciding how to spend compute, how to stage experiments, and when a faster GPU is worth reserving for a smaller set of high-confidence candidates. For teams that already have cluster access, that orchestration layer may matter as much as the model edits themselves.

What Hacker News argued about

Some readers said the post mostly shows parallel hyperparameter search with a larger compute budget, not a fundamentally new form of machine research.
Others argued that wall-clock speed is the point: if a research team can compress a multi-day search into one work session, that changes what is practical.
Several comments focused on the emergent H100/H200 workflow, because the agent inferred a two-tier validation strategy instead of being manually told how to use the hardware mix.

The skeptical reading is still useful. If the real tradeoff is worse GPU-hour efficiency in exchange for faster iteration, that is still operationally important. Much of applied AI work is bottlenecked by researcher time, not by an abstract optimum in GPU utilization. This experiment does not prove that autonomous research is solved, but it does show that the moment agents can control infrastructure directly, the research loop starts looking more like lab operations and less like a single script with an optimizer attached.

Sources: SkyPilot blog · Hacker News discussion

#skypilot

Hacker News Debates What 16 GPUs Really Changed in Karpathy's Autoresearch