The Scaling Debate Gets Real

The AI scaling debate stopped being theoretical this week. Three pieces brought data.

The myths

The AI Snake Oil team published AI Scaling Myths — a systematic takedown of the assumption that continued scaling will lead to AGI. Narayanan and Kapoor identify three specific misunderstandings of scaling laws: that benchmarks measure general intelligence (they don't), that past scaling trends will continue indefinitely (physical constraints say otherwise), and that data scarcity is a minor problem (it isn't). The post is valuable because it doesn't argue that scaling doesn't work — it argues that the extrapolations from current results to future capabilities are built on weaker foundations than the industry admits.

The compute landscape

Jack Clark's Import AI #435 covers the latest data on training run sizes — with some runs now exceeding 100,000 GPU-hours — and traces how AI systems are absorbing capabilities that previously required human effort. The most interesting section covers "intelligence per watt" as a metric: how much capability do you get per unit of energy? If this metric is the right way to measure progress, then efficiency improvements matter as much as raw scale — which is exactly what DeepSeek demonstrated.

The productivity data

Tyler Cowen linked to Erik Brynjolfsson's latest analysis showing US productivity grew ~2.7% in 2025 — nearly double the 1.4% sluggish average of the prior decade. The headline mirrors Solow's famous paradox ("you can see the computer age everywhere but in the productivity statistics"), except now the statistics might be turning. Brynjolfsson argues the data shows a genuine "decoupling" where GDP grew robustly despite lower labor force expansion, with technology finally delivering measurable gains.

The tension

Scaling myths, massive training runs, and productivity data that might finally show AI's impact. The three pieces don't agree with each other — and that's the point. The scaling optimists, the scaling skeptics, and the economists measuring real-world output are all looking at different parts of the same elephant. The picture won't come into focus until someone synthesizes them, and nobody has yet.