Zheng Dong Wang's 2025 Letter
A deeply personal and well-informed perspective on the compute thesis of AI progress. Wang traces the arc from Moravec in 1976 through Sutton's Bitter Lesson to the present, arguing that compute is the fundamental driver and that the exponentials show no signs of stopping. The letter is part memoir, part manifesto, part travelogue -- and more honest about the social costs than most techno-optimist writing.
Key Quotes
"The enormous shortage of ability to compute is distorting our work, creating problems where there are none, making others impossibly difficult, and generally causing effort to be misdirected. Shouldn't this view be more widespread, if it is as obvious as I claim?" -- Hans Moravec, 1976
"General methods that leverage computation are ultimately the most effective, and by a large margin." -- Richard Sutton, "The Bitter Lesson"
"GPT-5 is wearing the skin of my dead friend."
"It's not different this time because it's always been different. There hasn't been any constant normal trend ever."
"Everyone is pulled forward by a sense of inevitability." -- Jack Clark
Key Themes
The compute thesis is the backbone: model performance improves predictably with compute (4-5x annual growth over 15 years), AI task horizons double every seven months (from 9 minutes to 4+ hours in two years), and Moore's Law continues to deliver. Wang's personal narrative -- from skeptic to believer after watching a 1000x compute scaling produce nearly unbelievable behavior -- gives the argument visceral force.
The Moravec quote from 1976 is stunning. Nearly fifty years before the current AI wave, he identified exactly the dynamic we're living through: insufficient compute distorts the entire field, making problems appear where there are none.
The second-order effects catalog is what distinguishes this from standard accelerationist writing: frozen labor markets, teen AI dependency (1/3 of American teens), bipartisan opposition to data centers, job displacement across software engineering, mathematics, and molecular biology. Wang doesn't dismiss these -- he acknowledges the gap between technological capability and social readiness.
Critical Analysis
The strongest part is the tension Wang holds between inevitability and agency. The compute curves are deterministic in retrospect but required "eleventh-hour toil of a bunch of nerds" in practice. This is a more sophisticated position than most AI commentary, which tends toward either pure determinism ("it's inevitable") or pure voluntarism ("we choose our future").
The weakest part is the prediction timeline: AGI in 8-15 years, no compute bottleneck through 2030. These are plausible but unfalsifiable in the near term, and the letter doesn't seriously engage with the counterarguments (data exhaustion, diminishing returns on scale, energy constraints).
The India travelogue section is an interesting digression -- UPI as financial inclusion success, the value of improvisation over optimization -- but it feels disconnected from the main argument. The London-as-post-AGI-city idea is more provocative: when utility-driven progress is automated, the cities that preserve anachronism and intellectual tradition become more valuable, not less.
This pairs well with 2025 in LLMs for a tactical view of the same year Wang is reflecting on philosophically.
Sources: Zheng Dong Wang's 2025 Letter Last updated: 2026-05-14