Exponential realist · the future section

The cost curves

The modules speak in one voice, the analytical observer. This page is the one place where another voice, the exponential realist, speaks at length. Its rule is simple: no claim about the future without a measured curve, a time window and a source behind it. It sorts curves by how much history they have, follows each one only as far as that history allows, and says plainly where it stops.

Curves with a long record

These have ten to thirty years of consistent data behind them. Extrapolating a few years ahead is reasonable, as long as the reason the curve falls is still at work.

Curves too young to extrapolate

These are real and steep, but short. Two years of data can show a direction; they cannot show a learning rate. The honest move is to report them, name what drives them, and decline to run them forward.

Exponential realist Price of a fixed level of LLM capability Nov 2022 – Oct 2024

Running a model that performs at the level of GPT-3.5 became more than 280 times cheaper in under two years. That is roughly a 19-fold fall per year. Hardware improves only about 1.3-fold a year, so most of the fall has to come from the models and the software that serves them. It is the curve behind Module 14's discussion of AI costs and Module 15's cheaper simulations of language-model agents. Other measurements of the same thing land lower: a 2024 analysis by the venture firm a16z put the fall at roughly tenfold a year. Somewhere between ten and twenty is the honest range. Two caveats keep it honest too: it measures standing still at a fixed level of capability, not the price of the newest models; and two years is too short to call it a law.

Curves that do not exist yet

Some forecasts in circulation rest on things that have not been built: general intelligence on a date, fusion on the grid, self-improving systems. There is no cost curve for a technology that does not exist, so this section does not extrapolate them. Module 14 sets out the arguments about artificial general intelligence; the realist's contribution is to note that none of them can be read off a measured trend.

What follows, if the curves hold

If batteries and solar keep their learning rates, the cost of clean electricity keeps falling, and the binding limit on computing shifts from the price of chips to the speed of building grids. If the price of a fixed level of AI capability keeps falling at anything like its recent pace (which the section above declined to promise), today's frontier becomes an everyday tool within a few years, and the experiments this site describes, from simulated economies to populations of language-model agents, move from research budgets to ordinary ones.

Every one of those sentences carries an if. Learning curves have held for decades because the conditions behind them held: growing markets, competing manufacturers, open trade. Remove those, and a curve can stall. The realist's forecast is therefore conditional by design: here is the trend, here is why it has held, and here is what would break it.

Module 14: AI Futures · Module 15: The Future of Agent-Based Modeling · How this was made