In one sentence
AI progress is neither mystical nor guaranteed: it emerges from layers of empirical, compounding trends in computation, organizational capacity, and human ingenuity. These trends may fail, but their repeated persistence justifies taking transformative AI seriously—while preserving pluralism, human agency, and room for values that optimization cannot settle.
Overview
The essay begins by imagining what a perceptive observer in 2015 could have predicted about 2025. The result would include both astonishing AI capability and disappointing continuity: robots and debt, frontier research and public skepticism, automation and a frozen labor market. Wang uses this retrospective to argue that people systematically underestimate progress because they understand trends conceptually but fail to internalize what a new capability feels like in practice.
He then descends through a “scaling ladder”: AI task time horizons, model scaling laws, Moore’s Law, Sutton’s “Bitter Lesson,” organizational computation, and progress itself. At every level, he acknowledges reasons extrapolation might fail. Yet the same pattern recurs: compute expands, bottlenecks are overcome by new techniques, and systems acquire capabilities that seemed qualitatively out of reach. His confidence is therefore empirical and comparative, not deductive.
The final AI section argues that first-order effects—more capable, general, and efficient models—should dominate speculation until repeated evidence supports more elaborate second-order stories. Wang expects AI to become historically transformative, but says politics, economics, culture, and ethics have not yet seriously thought through the consequences. The appropriate response is not passive certainty but preparation, serious institutional thinking, and choosing work one would be proud to have done.
The personal half organizes the year around three loosely Isaiah Berlin–inspired themes: pluralism, marginalia, and unreasonable optimism. Travel, especially in India and China, supplies examples of institutions and people improvising around constraints. “Andor,” literature, and
Core ideas
Prediction requires historical humility
The 2015-to-2025 comparison shows that forecasting fails not only because people miss trends, but because they cannot emotionally or practically grasp what a trend’s consequences feel like once embedded in ordinary life.
Compute is the essay’s master variable
Wang treats time-horizon gains, scaling laws, Moore’s Law, the Bitter Lesson, and even organizational progress as nested versions of one idea: more computation, made possible by engineering and institutions, repeatedly unlocks capabilities.
The argument is empirical, not deterministic
Every trend has regime changes, bottlenecks, and human interventions. Wang explicitly says progress could stop tomorrow. His belief comes from the repeated fact that apparent bottlenecks have historically been overcome, making continued progress more plausible than immediate stagnation.
Firsthand surprise matters
Abstract charts are weak evidence psychologically. A better test is to choose a domain you know well, make predictions over time, and watch AI repeatedly outperform what expertise says should be possible. Direct encounters can update beliefs more effectively than general statistics.
AI’s first-order effects deserve priority
Before building elaborate theories about geopolitics, labor, or culture, establish the basic fact that models are becoming more capable, general, and efficient. Second-order speculation should be developed by serious experts outside AI, not treated as settled.
Progress does not answer how to live
AI may optimize or automate many activities, but it cannot determine which human values deserve priority. Berlinian pluralism—accepting that liberty, equality, justice, mercy, and other goods may be genuinely incommensurable—guards against one totalizing objective.
Marginalia and agency matter
History is made not only by famous leaders or headline technologies but by minor characters, administrators, engineers, companions, and countless local choices. Their actions can look like destiny in retrospect without being predetermined.
Optimism is an act of agency
“Unreasonable optimism” means continuing to build, choose, and care despite large forces that seem deterministic. The essay’s examples—UPI, Indian entrepreneurs, small businesses, and fictional rebels—show agency expressed through practical improvisation rather than grand certainty.
Practical takeaways
- When evaluating AI forecasts, separate the observed trend from the mechanism that might sustain it. Ask what has historically repaired bottlenecks, and what would falsify the extrapolation.
- Seek a personal benchmark: a task in your own area of expertise where you can make repeated predictions and observe model performance directly.
- Use a first-order-before-second-order rule: establish capability, adoption, cost, and reliability before making sweeping claims about social consequences.
- Do not let technological acceleration collapse your value system into one metric. Keep multiple legitimate ends in view, even when they conflict.
- Choose projects with short feedback loops and durable value; assume specific skills may depreciate quickly, but judgment, taste, relationships, and purpose still matter.
- Work intensely without treating exhaustion as proof of seriousness. Preserve slack for creativity, reflection, and unplanned discovery.
- Travel or study places long enough to encounter their default institutions—not merely their tourist surfaces. Short repeated visits can be more revealing than one hurried grand tour.
- Read specific background material before visiting a place, museum, or historical subject; context makes direct experience more intelligible.
Caveats and counterpoints
- The central case for continued AI progress relies heavily on extrapolating empirical trends whose underlying mechanisms remain only partly understood. The author acknowledges that scaling regimes can end and that progress has repeatedly depended on contingent human innovation.
- The essay is strongly optimistic about AI’s benefits and less developed on distributional harms, institutional failure, environmental costs, misuse, and how power might respond when capability is concentrated. It calls for second-order work more than it supplies a complete political program.
- Evidence from frontier labs, benchmarks, and Wang’s own research experience may not generalize cleanly to all economically valuable work. The essay notes concerns about task selection, reliability, contractor baselines, and benchmark limitations, but its conclusion still leans on frontier capability as a guide to broader transformation.
- The personal sections are intentionally impressionistic. Observations about India, China, cities, travel, and culture are useful as vignettes rather than systematic comparisons.
- The article is dated December 30, 2025, and presents a snapshot of a rapidly changing AI landscape; specific model names, forecasts, and institutional situations should not be treated as timeless.
Questions worth revisiting
- Which of Wang’s nested trends is the most defensible: hardware compute, model scaling, task time horizons, organizational learning, or progress itself?
- What evidence would count as substantial and repeated evidence that the compute thesis has failed?
- If AI capability becomes abundant, which human activities should remain deliberately slow, inefficient, or non-optimized?
- How can pluralist values be represented in institutions that must still make concrete choices and tradeoffs?
- Does concentration of compute increase or weaken state power, and what determines the outcome?
- What would a serious second-order analysis of AI look like from labor economics, democratic politics, education, or culture?
- How should an individual choose work when both capabilities and career paths may depreciate quickly?
Return to this when…
Return to this letter when you need a compact reminder of the strongest optimistic case for AI progress, especially the distinction between conceptual awareness and lived recognition. Also revisit the closing themes—pluralism, marginalia, and unreasonable optimism—when technological forecasts begin to feel like arguments for passivity or inevitability.