Bound

Argued. Not a forecast. Contract 0.1, issued in publication v0.2. Open forecasts: 0.

A bound is a ceiling that no scheduled observation fully settles. Each claim below names the evidence that would withdraw it. None of them is a date.

Device efficiency

On published CMOS models, further energy-efficiency gains are large but not open-ended. Epoch’s estimate puts the H100 near 1.4 × 10¹² FLOP/J, and a 16-bit physical model near a geometric mean of 2.9 × 10¹⁴ FLOP/J. That is about a 200-fold gap. Their distribution says there is a 50 percent chance gains cease before roughly that 200-fold improvement.

Source: Epoch AI, Limits to the energy efficiency of CMOS microprocessors, 15 December 2023. 14 December 2023

Withdrawn if a measured, cited device exceeds that 16-bit mean by an order of magnitude, or if the model’s stated assumptions are replaced by a later Epoch estimate.

A training-run wall, not a ceiling

Moving data, not the Landauer limit, is the nearer constraint on one training run. Epoch’s model finds that a run past about 2 × 10²⁸ FLOP in three months becomes infeasible at good utilization with current techniques. A more aggressive batch-size assumption pushes that fall in utilization out to about 3 × 10³⁰ FLOP. This is an engineering bound. It is not the bound on what a later stack can do.

Source: Epoch AI, Data movement bottlenecks to large-scale model training, 2 November 2024.

Withdrawn for the stated technique if a documented run past 2 × 10²⁸ FLOP finishes in three months at utilization comparable to current frontier runs.

What the ceiling is not

Nothing in these bounds says a system cannot do most checkable cognitive work faster than a person, in parallel. They also do not say it can know what an experiment has not yet measured. Sample efficiency is not implied by scale. A pretraining loss curve is not a bound on fluid intelligence.

Withdrawn if a pre-registered check shows a system acquiring a novel skill, on a held-out task a person finds easy, at a data and compute cost within an order of magnitude of a skilled person. No such instrument is named in version 0.2. The five domain lines on the first screen stay empty.

Issued 9 October 2026. Version 0.2. Not a correction of the first screen.