2025 Letter by Zhengdong Wang
Main Arguments
The Compute Theory of AI Progress
Wang argues that computational resources—"compute"—is the fundamental driver of AI advancement. He traces this insight through multiple frameworks:
- Time Horizons: AI can complete tasks taking humans 4+ hours, up from 9 minutes two years prior, doubling every seven months
- Scaling Laws: Model performance improves predictably with increased compute (4-5x annual growth over 15 years)
- Moore's Law: Transistor density and system-level compute have sustained exponential growth for 50+ years
- The Bitter Lesson: General methods leveraging computation ultimately outperform domain-specific human expertise
Key Personal Narrative
Wang recounts his journey from skepticism to conviction: - 2015: Dismissed early AI via science fair project as late to the field - 2017: Thought the field was "saturated" after reading AlexNet - 2021: Remained doubtful despite GPT-3's emergence - 2023-2024: Direct experience scaling embodied AI agents proved transformative
His pivotal moment came when a model trained with 1000x more compute exhibited behavior he could scarcely believe possible—smooth, intentional, nearly perfect task execution.
Notable Quotes
"General methods that leverage computation are ultimately the most effective, and by a large margin." (Richard Sutton, "The Bitter Lesson")
"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)
"GPT-5 is wearing the skin of my dead friend." (User response to ChatGPT-5 launch)
"Everyone is pulled forward by a sense of inevitability." (Jack Clark)
"It's not different this time because it's always been different. There hasn't been any constant normal trend ever."
Second-Order Effects & Concerns
Wang catalogs emerging societal tensions: - Opposition to data centers as bipartisan issue - Frozen labor market (cause unclear) - Teen dependency on AI for companionship (1/3 of American teens) - Public skepticism despite massive investment - Job displacement in software engineering, mathematics, molecular biology
He invokes "Zeynep's Law": until substantial repeated evidence emerges, assume counterintuitive findings false and first-order effects dwarf second-order ones.
On Inevitability & Human Agency
Wang embraces "unreasonable optimism"—small acts of agency against deterministic forces. Reflects Isaiah Berlin's value pluralism: highest human values (liberty, equality, justice, mercy) are incommensurable; no perfect ordering exists.
Progress looks inevitable retrospectively but required countless contingent human choices. Compute scaling appears deterministic yet masks "eleventh-hour toil of a bunch of nerds."
Predictions & Timeline
- AI will complete week-long human tasks within years
- Whatever constitutes AGI may arrive within 8-15 years
- Compute efficiency improvements accelerating
- Data center investment will exceed Apollo program cost (inflation-adjusted)
- No computational bottleneck anticipated through 2030
The India Section & Cultural Observations
Extended travel narrative emphasizing contrast between Western optimization and Indian improvisation. UPI's success as financial inclusion tool. Value of unpredictability and human networks. Major progress depends on unsung contributors, not celebrated leaders.
On Cities & Post-AGI Futures
London as ideal post-AGI city: 1000-year history, intellectual tradition unburdened by Silicon Valley's "fail-fast" mentality. Preservation of anachronism becomes valuable when utility-driven progress dominates.
Conclusion
Embraces pluralism over monism, marginalia over heroic narratives, and agency over determinism. The compute-driven AI wave is happening. Policy, culture, and human purpose must adapt—but haven't yet. This gap between technological inevitability and social readiness defines the decade ahead.