On March 19, 2026, Hacker News surfaced "NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute", which had 143 points and 29 comments when this crawl ran. In the linked Q Labs article, the team says an ensemble of 1.8B-parameter models, 18B parameters in total, trained on 100M tokens can match a standard language-model baseline that would normally need 1B tokens. That is the core of the 10x data-efficiency claim, and it is why the post drew attention: it pushes against the common assumption that better models require roughly proportional growth in both compute and data.

What the recipe actually includes

Ensembling, where multiple separately trained models are combined at inference time instead of relying on one model to absorb all capability.
Chain distillation, where each new model learns from the previous one rather than from a full teacher ensemble.
Heavy regularization, including unusually large weight decay, to make overparameterized models generalize better under tight data constraints.
Looping, which reuses part of the transformer repeatedly so the model spends more compute per prediction.

The broader argument is strategic. If high-quality text becomes the limiting resource, labs need ways to keep buying performance with compute after the easy data is gone. Q Labs presents Slowrun as evidence that fixed-data scaling can still work if the training recipe changes enough. The result is still an early research claim rather than a settled production recipe, but it clearly matters to anyone watching the next generation of scaling-law debates.

What Hacker News questioned

Several readers asked whether the gain survives deployment, because an ensemble that looks good during training can become expensive to serve.
Others wanted to know whether most of the benefit can eventually be compressed back into a single model, which would turn the result from an interesting benchmark into a usable systems recipe.
Some comments also stepped back and asked what a fair human baseline for data efficiency would even look like, given that humans arrive with massive evolutionary and experiential priors.

That makes the post important even if the 10x number does not transfer cleanly to every larger model stack. It reframes the bottleneck. Instead of asking only how to afford more training FLOPs, it asks what happens when the industry can afford more compute than genuinely new, clean, high-quality text. If that imbalance keeps growing, approaches like ensembling, distillation, and looped architectures may move from niche experiments into the main scaling conversation.

Sources: Q Labs article · Hacker News discussion

#ensembles

Hacker News Examines NanoGPT Slowrun's 10x Data-Efficiency Claim