Source: Redwood Research | Author: Ryan Greenblatt | Published: 2026-02-11

Summary

Greenblatt draws a crucial distinction that most AI discourse collapses: when inference costs go up, is it because models are doing harder tasks (which naturally take more tokens), or because models need disproportionately more compute per unit of capability? He uses a Pareto frontier framework plotting budget against “50% reliability time-horizon” — the task duration a human has a 50% chance of completing. In the linear regime, cost scales proportionally with task complexity, and the cost-as-percentage-of-human-labor stays flat. That’s not inference scaling — that’s just bigger jobs costing more, the way hiring a human for a week costs more than hiring one for an hour. True inference scaling is when you need disproportionately more compute to push the frontier further, creating diminishing returns that hit an economic ceiling.

Key Insight

The most important claim: most 2025 progress was frontier shifts in the linear regime, not unsustainable inference scaling. This matters enormously for forecasting. If progress is primarily “models can now do 4-hour tasks instead of 2-hour tasks, and it costs roughly twice as much,” that’s economically sustainable and will continue. If progress requires 10x compute for 2x capability, that’s a wall. Greenblatt argues we’re mostly in the first regime, with exceptions in narrow domains like math olympiad problems where extreme compute concentration works but doesn’t generalize.

Our Take

This connects directly to the monitoring range problem: in the linear regime, monitoring costs scale proportionally with capability, which means the monitoring range (how many capability levels ahead you can see) stays roughly constant. Under true inference scaling, the defender’s costs grow faster than the capability gains, which compresses the monitoring range. Greenblatt’s empirical claim — that we’re mostly in the linear regime — is optimistic for alignment: it means the monitoring problem isn’t getting exponentially harder per unit of capability gained. The math olympiad exception is interesting precisely because it’s the kind of narrow, verifiable domain where you’d expect inference scaling to work best and alignment concerns to matter least.

The piece is also useful for prediction markets. “Will inference costs make AI economically unviable by X date” markets should be trading lower if Greenblatt is right — the cost curve is linear in task complexity, not exponential in capability.