The idea sounds counterintuitive: if we keep increasing data and compute, shouldn't models become more and more general? The answer Goldfeder, Wyder, LeCun and Shwartz-Ziv proposed in 2026 is more forceful: under real-world constraints, specialization isn't a strategic choice — it's a mathematical and empirical consequence.
The mathematical prediction: No Free Lunch and finite resources
The No Free Lunch theorem says something simple and hard to swallow: there is no algorithm that wins on every possible problem. Averaged over all conceivable problems, all algorithms tie. What does that mean in practice? Improving on one set of tasks implies giving up performance on others.
Add finiteness: compute, data and time are limited. If you split resources among a hundred tasks, the slice per task shrinks. The arithmetic is straightforward: more coverage means less depth per task. That's why the strategy that maximizes performance under real conditions is usually to concentrate resources on a well-defined objective.
