Thesis
His research paper on clarity and judgment.
Upstream of Everything
Clarity, Deflation, and the Race to Build Minds That Judge for Us
Every decision is made on a model of the world, and a model that is wrong in a way that matters aims every later choice at a fiction.
The deflation argument is Jeff Booth's. The paper cites his book The Price of Tomorrow (2020): technology is inherently deflationary. Jeff Booth is family, and a big reason behind all of James's learning.
Hatched bar: 2004 study, later summary. Bracket: that study's own values, 131 to 172 s. Dashed line: the 2016 median, 40.2 s.
Redrawn from Figure 1 of the paper, page 4. Data: González and Mark (2004), Mark et al. (2012), Mark et al. (2016) and Mills (2023).
Show data
| Year | Seconds | How it was measured |
|---|---|---|
| 2004 | about 150 | Observation with stopwatches. The 2004 paper reports 131 s on any device or paper and 172 s on the PC. About 150 s is Mark's later summary. |
| 2012 | 75.5 | Window logging, 13 participants. |
| 2016 | 47.0 | Screen logging, 40 participants over two work weeks. Median 40.2 s. |
Log scale. The three prices the paper names. Its figure draws every step between them. $60 to $0.07 is 857 times lower in 2.9 years.
Redrawn from Figure 8 of the paper, page 21. Data: Cottier et al. (2025), Epoch AI, licensed under CC BY 4.0.
Show data
| Date | US$ | Cheapest model at this level |
|---|---|---|
| November 2021 | $60 | GPT-3. The cheapest model scoring at least 43.9% on MMLU, the GPT-3 level. |
| March 2023 | $2 | GPT-3.5 Turbo. The cheapest model at the GPT-3 level in March 2023. |
| October 2024 | $0.07 | Gemini 1.5 Flash-8B. The cheapest model at the GPT-3 level in October 2024. |
The argument in twelve steps
Introduction: Clarity as the Upstream Condition
This paper follows that claim from the human observer to the machines now being built to judge for us.
The Bounded Observer: Attention, Memory, and the Limits of Judgment
Section 2 describes the limits of the human judge.
Evolution: The Slow Algorithm That Built the Bottleneck
Section 3 sets out the three conditions for selection and the slow pace at which they produced the brain and the eye.
Technology as Accelerated Selection: The Deflationary Engine
Section 4 treats technology as the same process on a faster substrate and develops the argument that it drives prices down.
Machine Learning: Selection on a Faster Substrate
Section 5 explains how models are trained, why scale worked, which multipliers account for recent progress, and why fluency, however cheap, is a different thing from clarity.
From Narrow Tools to Drop-In Workers: AGI and Its Bottlenecks
Section 6 adopts an operational definition of general intelligence and examines six bottlenecks that stand in its way.
Superintelligence and the Limits of Forecasting
Section 7 asks what follows if the loop runs past human level and how little that forecast specifies.
Consequences: Labor, Security, and Science
Section 8 compares estimates of economic impact and considers military and scientific effects.
The Industrialization of Intelligence
Section 9 follows the capital and checks projected build-outs against what has been built.
Alignment: Who Writes the Loss Function
Section 10 takes up the problem of steering systems smarter than their overseers and sets out a taxonomy of risk.
Can the Transition Be Positive-Sum?
Section 11 lays out six scenarios.
Conclusion
Section 12 returns to clarity.