Most “make the model cheaper” work happens at inference — quantization, speculative decoding, caching. Alexi Gladstone’s Explorative Modeling writeup argues the leverage is upstream, at pretraining, and the payoff shows up downstream in the inference bill.

The core intuition is one every generative-model engineer knows in their gut: train a network to directly regress the target and it collapses to the average — the “brown blur” that is no real dog. Exploration is the fix, framed here as a third pretraining axis alongside data and parameters. Reward the model for the best of K guesses rather than the mean, and the reported gains compound with scale: 7% to 36% as you add data, 13% to 23% as you add parameters. The efficiency claims are the part that made me sit up — 6.2x sample efficiency, 4.1x FLOP efficiency, and, on control tasks, matching diffusion with up to 256x less inference compute.

If that 256x holds outside the paper’s benchmarks, it reframes a lot of agentic-inference math. Best-of-N sampling at serving time is how many teams buy quality today — you pay N forward passes to pick one good answer. Baking the exploration into pretraining instead of the sampling loop is the difference between renting that quality per-request and owning it. The HN discussion is appropriately skeptical about whether “monotonic improvement with exploration” survives contact with real language pretraining at frontier scale, and that’s the right question to sit with.

I’d want to see this reproduced on a language model someone actually serves before betting a latency budget on it. But if the inference-compute number generalizes even halfway, “spend more at pretraining to stop paying at inference” stops being a slogan and starts being a line item. Which team ships the first XM-trained model you can call in production?