2025 letter 🔗 **Source:** [2025 letter](https://zhengdongwang.com/2025/12/30/2025-letter.html) > [!note]+ Document Notes > ```yaml > tags: [AI, compute, progress] > author: Zhengdong Wang > source: zhengdongwang.com > reference: [https://zhengdongwang.com/2025/12/30/2025-letter.html] > date created: 2026-01-07 14:32 > ``` > > # 🏷️ Title: 2025 letter > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - Zhengdong Wang, “2025 letter” (personal annual letter on AI, travel, and pluralism). ## 📝 Summary - What happened: Wang reflects on the past decade of AI progress, argues for “compute” as the central driver, describes his own research at DeepMind, and mixes in travel and philosophical reflections. - Why it matters: It frames AI as a general-purpose technology on the cusp of transformative impact, yet still “wildly early.” - Key stakeholders: AI labs, policymakers, researchers, educators, broader knowledge workers. - Risks and tradeoffs: Over/underestimating AI timelines, centralization of compute, neglect of humanities and pluralism. - Bottom line: AI’s first-order impacts via scaling compute are inevitable enough to plan around; second-order social responses are underdeveloped and urgently need work. ## 📖 Main Content - Core arguments: - Backcasting from 2015–2025 shows how hard it is to predict AI; we systematically underestimate it. - A ladder of empirical trends—task “time horizons,” scaling laws, Moore’s Law, Sutton’s “Bitter Lesson,” and ultimately long-run progress—suggests that more compute reliably yields more capable models, even though each trend has caveats. - The author’s own trajectory (science fairs, college, DeepMind work on embodied agents and computer-use agents) repeatedly demonstrates that simply scaling compute by large factors produces qualitatively surprising behavior. - AI is “moving history” like the pandemic: widely signaled before its full physical impact, hitting domains (homework, coding, research) in waves. - First-order effects (cheaper, more general, more efficient intelligence) will dominate; second-order effects (policy, geopolitics, culture, meaning) are much less thought through yet. - Philosophically, Isaiah Berlin–style pluralism, attention to marginalia (supporting characters and institutions), and “unreasonable optimism” are offered as guides for living and working in an AI-saturated era. - Context & terms: - “Compute”: total computational resources devoted to training and running models; framed as the main axis along which capabilities scale. - “Scaling laws”: empirical relationships where model performance improves predictably with more data, parameters, and compute. - “Time horizons”: measures of how long a coherent task AIs can perform (e.g., coding for hours) is increasing. - The piece references AlphaGo, GPT-4/5, Gemini, DeepSeek, METR, and others to anchor claims. - Uncertainties: - The author repeatedly notes that all these curves are empirical, not theoretically guaranteed, and could plateau. - The exact societal impacts (labor, geopolitics, governance) are treated as open questions that require domain experts outside AI to seriously engage. ### 🗒️ Notes - Opens with a Peter Thiel–Elon Musk anecdote about “a billion humanoid robots” and the gap between rhetoric and fully “thought through” futures. - Uses 2015 vignettes (AlphaGo’s early win, Ukraine war, Falcon 9 landing, OpenAI’s founding) to show how much changed by 2025 and how easy it was to miss. - Introduces a stack of trends: - Task time horizons (e.g., METR’s plot on how long models can sustain useful work). - Scaling laws tying compute to performance. - Moore’s Law reinterpreted as “total system compute” at the data-center level. - Sutton’s “Bitter Lesson” and similar earlier insights (e.g., Hans Moravec) arguing that methods leveraging computation beat hand-crafted expertise. - Kurzweil’s long-run view of increasing information organization and “everything’s computer.” - Personal narrative: - High-school science fairs in Arizona; early coding and avoiding biological lab competition. - Discovery of “machine learning” via AlexNet and big astronomical datasets. - College history major, inspired by the AlphaGo documentary’s ethos. - Joining DeepMind (2021), work on reinforcement learning and embodied simulations; a key experiment scales compute 1000× and dramatically improves performance. - Experiencing a later agent (built on Gemini-scale models) that performs physical tasks so smoothly that he first assumes a bug, then realizes it simply crossed a qualitative threshold. - Argument style: - For each trend, he lists reasons not to trust extrapolation (domain bias, measurement noise, empirical nature, lack of underlying theory), then notes that the trend kept going anyway. - This repeated pattern leads him to cautiously accept extrapolation: not as inevitability, but as “it would be stranger if it stopped now.” - AI as moving history: - Pandemic analogy: public knowledge preceded everyday experience; only later did “the whole world share the same story.” - AI waves: high-school homework, then software engineering, then early inroads into research mathematics and molecular biology. - Anecdotes about interns treated like “Perplexity Pro bots” and someone joking that an AI (o3 Pro) could have predicted a bad date. - Timelines and magnitude: - AGI timeframe estimates converge to “within our lifetimes,” with some skeptics still saying 8–15 years. - He notes rising investment, talent inflows, and lack of hard scaling bottlenecks (including power) in the near term. - The exact definition of AGI matters less to him than the recognition that AI will be the “technology of the century.” - First- vs second-order: - First-order: rapidly falling cost of intelligence, more generality, and higher efficiency; labs explicitly design each generation to be cheaper/faster at similar or better capability. - Second-order: messy, sometimes contradictory policy slogans (race with China yet pause; AI-driven UBI yet acceleration; catering to entertainment vs paternalistic content controls). - Invokes Zeynep Tufekci’s heuristic: assume counterintuitive findings are false and second-order effects smaller than first-order until strong evidence says otherwise. - Governance and geopolitics: - AI 2027 tabletop: what matters at “takeoff” is global compute share, not just chip design or production. - London finance firms acknowledge AI’s potential but primarily fear being overregulated relative to foreign competitors. - A D.C. anecdote features a politician invoking “alignment and alternative interpretability methods” in quasi-religious tones. - Humanities and meaning: - Cites Jonathan Malesic on the humanities’ role in teaching us “how we want to live,” via slow, often invisible work that deepens others’ lives. - Harvey Lederman’s worry about being among “the last” humans to do philosophical exploration before automation takes over; appreciates living in a time with purpose. - Jasha Sohl-Dickstein’s advice to young investigators: embrace leverage, fast projects, robustness to the Bitter Lesson, and doing work you’ll be proud of. - Life choices: - Skills and GPUs depreciate quickly; if AI will overturn many career paths, one might as well pursue what one really wants (e.g., a “big group project with friends” like Sergey Brin). - Author is joining DeepMind’s “Post-AGI” team, underscoring his belief that we’re still early. - Philosophical frame: - Draws on Isaiah Berlin’s value pluralism: high values are often incommensurable, yet not anything-goes. - Emphasizes holding contradictory truths: work hard yet preserve slack; be both arrogant and humble as a scientist; care about everything and “nothing really matters” at the same time. - Personal/travel: - Extensive travels (Europe, Asia, North America) reinforce themes of pluralism and marginalia; India in particular embodies “neti” (“not this, not that”) as a third way distinct from U.S. and Chinese models. - Highlights India’s UPI as massive financial-inclusion infrastructure and an example of optimistic state-led building. - Describes grassroots data-collection efforts for underrepresented languages and the dignity shift when people get paid to teach AI their mother tongues. - Culture: - Star Wars: Andor is praised as a show about “marginalia” — the supporting characters behind galactic events — mirroring the letter’s appreciation of unsung actors in history and organizations. - Notes the absurd but relatable corporate parallels in imperial hierarchy scenes. - Productivity: - Argues against extreme “996” work cultures; suggests ~60 hours/week and “facing the right direction” as more important. - Time, not sheer effort, is rarely the fundamental bottleneck; direction and commitments matter more. - Ends with a light household recommendation: a long wooden shoehorn as a tiny, joyful productivity aid. ## 📚 References - Richard Sutton, “The Bitter Lesson” - Hans Moravec, “Bombast” - Ray Kurzweil, *The Singularity Is Near* - METR (Model Evaluation and Threat Research) time-horizon work - Papers on scaling laws for Transformers (e.g., Kaplan et al. 2020) - AlphaGo/AlphaZero, GPT-4/5, Gemini, DeepSeek, ARC-AGI, Humanity’s Last Exam, FrontierMath - Jonathan Malesic, Harvey Lederman, Jasha Sohl-Dickstein writings/talks (alluded to) - Isaiah Berlin’s essays on value pluralism and history - Tony Gilroy’s *Andor* (TV series) - India’s Unified Payments Interface (UPI) public documentation ## 🗣️ Paraphrase - Paragraph A: The author looks back from 2025 to 2015, showing how difficult it would have been to foresee today’s AI achievements and using that difficulty to motivate humility about future predictions. - Paragraph B: He builds an argument that a tower of empirical trends, all centered on compute, is enough to justify expecting much more powerful AI soon, then broadens into reflections on how societies and individuals should respond. ## 📰 Newsletter AI is no longer future hype — it’s the new baseline. In his “2025 letter,” Google DeepMind researcher Zhengdong Wang walks through a decade of scaling compute, from AlexNet to GPT-5, and argues that we’re still “wildly early.” For WordPress builders, that means treating AI as core infra: smarter code-assist, content workflows, and accessibility tooling. As you plan your next plugins, themes, or learning paths, design for a world where the cost of intelligence keeps dropping — and human judgment, community, and governance matter more than ever. ## ✅ Action Items #### DevRel - Map how AI-code assistants and agents change your target developers’ workflows; update demos and tutorials accordingly. - Build short, tightly scoped AI examples (e.g., “compute-aware” performance checks, content refactoring) rather than sprawling showcases. - Incorporate discussions of scaling laws and compute constraints into talks to ground expectations. - Collect structured feedback from developers on where AI tools actually help vs. hinder. #### OSPO - Add AI/compute considerations to open source governance: model weights, data licensing, and infra access. - Define policies for contributions that integrate proprietary AI APIs into open codebases. - Track dependencies on large AI providers; maintain an inventory of critical models and services. #### Community - Host discussions on how AI reshapes contributor roles (docs, code, design, translation). - Offer mentorship on using AI as an assistant, not a replacement, for learning and contribution. - Document “AI etiquette” for community spaces (e.g., marking AI-generated content, expectations of human review). #### WordPress - Core: explore AI-augmented triage and patch review that respects maintainers’ authority. - Playground: experiment with AI agents that can reproduce bugs and propose fixes in-browser. - Training: build learning paths that combine classical dev skills with AI tooling literacy. - Documentation: maintain guidance on when and how to safely use AI-generated code and copy. - Accessibility: evaluate AI tools to assist with alt-text, headings, and contrast checks, while keeping humans in the loop. #### The WPOCC - Support fellowships for maintainers exploring responsible AI in WordPress ecosystems. - Fund experiments in AI-assisted governance tools (e.g., agenda drafting, summary bots) with transparent evaluation. - Encourage cross-project conversations on compute access, sustainability, and shared infra. ### FAIR - Encourage local AI tooling that runs on contributor hardware where possible, reducing centralization risk. - Publish clear governance artifacts around AI funding, infra partnerships, and sunset clauses. - Treat AI-related policies as living documents, updated as empirical evidence accumulates. ## 🧪 Hypotheses - H1: In the near term, increased compute remains the dominant driver of frontier-model capability gains. - H2: AI’s first-order economic effects (cheaper intelligence, broader automation) will overwhelmingly outweigh second-order cultural effects in the 2020s. - H3: Domains with clear benchmarks (coding, games, robotics tasks) will experience capability shocks earlier than less formalized fields. - H4: Societies that treat AI as infrastructure (like India did with UPI for payments) will capture more inclusive benefits. ### 📊 Methodology - For H1: Collect model-capability benchmarks across generations; regress performance on estimated training compute, controlling for data/architecture changes. - For H2: Compare macro indicators (productivity, wages, sectoral shifts) with measures of cultural/psychological impact (surveys, media) over time. - For H3: Track adoption and performance changes in benchmark-heavy vs benchmark-poor professions, using longitudinal labor and productivity data. - For H4: Case-study comparative work across countries that build public digital/AI infrastructure versus primarily private-platform approaches. ### ❓ Questions - How far can compute-driven scaling go before hitting hard physical or economic limits? - What institutional forms are best suited to govern “general-purpose intelligence” as infrastructure? - How should individuals plan careers when skills and tools depreciate at GPU-like speeds? - What does a pluralist, non-monist ethics of AI deployment look like in practice? ### 🔮 Future Research - Theoretical accounts of why scaling laws hold as robustly as they do. - Detailed power, cooling, and land-use modeling for multi-GW AI data centers. - Longitudinal studies of AI’s impact on specific professions (law, medicine, education, philosophy). - Frameworks for measuring and governing “global compute share” in geopolitics. ### 💡 Implications - Practice: Expect rapid, repeated capability overhangs; design organizations to adapt quickly. - Policy: Focus on compute, data, and infra governance rather than just outputs and content. - Security: Treat AI labs and large clusters as strategic infrastructure with associated safeguards. - Education: Emphasize judgment, ethics, and pluralism over purely procedural skills. - Funding: Shift some research support toward second-order societal impacts, not just capabilities. ### 🧠 Learning Styles Assessment - Reading level estimate: roughly college-level (Flesch-Kincaid suggests ~12–14th grade due to long sentences and technical vocabulary). #### 👀 Visual learners - Present: Conceptual descriptions of graphs (time horizons, scaling curves), but few actual visuals. - Improve: Add charts of compute vs performance, timelines of AI milestones, diagrams of the “ladder” from tasks → scaling laws → Moore’s Law → progress. #### 🔊 Aural learners - Present: Narrative style that would convert well to audio, but no built-in audio. - Improve: Provide a 5–10 minute audio reading of the letter; podcast-style discussion of its themes. #### 🗣️ Verbal learners - Present: Rich metaphors (pandemic, waves, marginalia), references, and quotations; strong narrative voice. - Improve: Add a short glossary for non-expert readers (compute, scaling laws, time horizons). #### 🧩 Kinesthetic learners - Present: Encouragement to “pick your own test” problem to experience AI’s shock. - Improve: Offer concrete exercises (e.g., try a specific coding or research task with a model, record expectations vs results). #### 🧠 Logical learners - Present: Layered argument from empirical trends up to philosophical claims; uses examples and counterpoints. - Improve: Include explicit tables summarizing trends, assumptions, and possible failure modes. #### 🧑‍🤝‍🧑 Social learners - Present: Many anecdotes of conversations, workshops, and trips; invites discussion. - Improve: Add reflection prompts suitable for reading groups or seminars. #### 🧘‍♀️ Solitary learners - Present: Long-form text conducive to reflective, solo reading. - Improve: Provide guided journaling questions at the end of each major section. #### 🌿 Naturalistic learners - Present: Occasional environmental metaphors (waves, landscapes) and travel imagery. - Improve: Tie compute/power discussion to environmental impacts and sustainability tradeoffs. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Mostly plain HTML text on a blog; likely reasonable headings. - Improve: Ensure clear H1–H3 hierarchy, language attributes, and descriptive link text. #### Keyboard navigation - Present: Standard article likely keyboard-accessible. - Improve: Verify skip-to-content links and visible focus outlines. #### Contrast and readability - Present: Typical blog styling, probably adequate but unverified. - Improve: Check color contrast ratios, line height, and font size for long-form reading comfort. #### Text resizing and reflow - Present: Likely responsive layout. - Improve: Test at 200% zoom; minimize horizontal scrolling and layout shifts. #### Media and captions - Present: Mainly text with some images; alt text unknown. - Improve: Add descriptive alt text to images; avoid auto-playing media if any. #### Forms and interactions - Present: Minimal or no complex forms. - Improve: If subscription/comments exist, ensure properly labeled inputs and accessible error messages. #### Standards reference - Map any gaps to WCAG 2.2 AA, especially 1.4.x (perceivable), 2.1.x (keyboard), 2.4.x (navigable), and 3.3.x (input assistance). ### 🌍 DEIB Assessment #### Cultural breadth - Present: Wide geographic and cultural exposure (India, Malaysia, Europe, U.S.); reflects on different governance styles. - Improve: Include more voices quoted directly from non-Western authors and practitioners. #### Inclusive language - Present: Generally neutral, reflective tone; limited jargon explained by context. - Improve: Add short explanations for highly technical or insider AI terms. #### Neurodiversity support - Present: Long, coherent sections; predictable annual-letter structure. - Improve: More subheadings, summaries, and bullet lists to break up dense text. #### Representation - Present: Mentions a range of colleagues, thinkers, and cultures. - Improve: Make diversity of sources more explicit (gender, region, discipline). #### Psychological safety - Present: Reflective and non-dogmatic; acknowledges uncertainty. - Improve: Offer clearer content warnings only if future letters cover more distressing topics (e.g., catastrophic risk scenarios). ### ✅ Logical Validity #### Strong reasoning - Uses inductive reasoning: multiple empirical trends (time horizons, scaling laws, Moore’s Law) are shown to hold despite valid objections, supporting cautious extrapolation. - Employs analogy (pandemic) to illustrate lags between conceptual awareness and felt impact, clarifying how AI might unfold socially. #### Fallacies - Possible survivorship bias: focuses on successful trends and breakthroughs; less on failures or dead ends. - Could be read as an appeal to authority via anecdotes involving prominent labs and figures; mitigated by explicit caveats. #### Tone of voice - Reflective, personal, cautiously optimistic. - Alternates between technical explanation and literary, almost lyrical narrative. - Frequently self-deprecating (“I’m so late to this AI thing”) while clearly committed. ### 🛡️ DARVO Analysis - Deny: The author does not deny risks; instead, he stresses that AI “will be a big deal” and invites more serious second-order analysis. - Attack: No attacks on critics; even skeptical voices (e.g., Gary Marcus) are presented respectfully. - Reverse victim/offender: No attempt to portray powerful actors (labs, investors) as victims; instead, he foregrounds responsibility and pluralist thinking. ## 🎯 Conclusion AI’s trajectory, viewed through compute and scaling, is both empirically grounded and historically contingent—but all signs suggest we’re very early in a transformative arc. The author urges readers to take first-order impacts seriously, invest more thought in second-order societal responses, and live with pluralism and optimism: pick meaningful problems, accept rapid change, and build the institutions and cultures that can channel unprecedented computational power well. ## 🧱 Refactor Appendix (append after 🎯 Conclusion) ### 📚 Literature Split Pad - Purpose: Stage Konik CEE units that become per-source Literature notes. - Split level: **H3**. - Include verbatim quotes under 25 words and a [view highlight] link. ### [H1] Compute scaling drives AI capability > [!quote] [2025 letter](https://zhengdongwang.com/2025/12/30/2025-letter.html) by [[Zhengdong Wang]] via [[zhengdongwang.com]]. Published 2025-12-30. Accessed 2026-01-07. > “Compute compounds, and its results confound.” Two lines: context and use. Cite data if relevant. [view highlight](https://zhengdongwang.com/2025/12/30/2025-letter.html#H1) #xref #articleseed ID:rw_[MISSING] ### [H2] AI is moving history like a pandemic > [!quote] [2025 letter](https://zhengdongwang.com/2025/12/30/2025-letter.html) > “AI is also moving history. That it will be a big deal will also be public information long before it is physically felt.” Two lines: context and use. [view highlight](https://zhengdongwang.com/2025/12/30/2025-letter.html#H2) #xref ID:rw_[MISSING] --- ### 🔖 Atomic Split Pad (Zettel Candidates) - Purpose: Promote reusable claims into Atomic notes. - Split level: **H4**. - Each item: one idea, present tense, specific, testable. #### [Z1] Compute remains main driver of frontier AI - Frontier-model performance in 2020s strongly correlates with training compute, within given architectures. - Pointer to letter and scaling-law papers. - Links: [[AI-Scaling]] [[Compute-Infrastructure]] #### [Z2] AI’s first-order impacts dwarf second-order stories - In early deployment phases, cheaper intelligence and automation dominate over cultural narratives. - Pointer to discussion of Zeynep’s Law and first/second-order effects. - Links: [[AI-Societal-Impact]] [[Technology-Of-The-Century]] --- ### ✏️ Fleeting Queue (Extract by Selection) - Purpose: Capture quick actions or sparks for Fleeting notes. - Extraction: **Select one line** → “Refactor: Extract selection to new note (content only).” - Keep each line short and self-contained. - [F1] Try my own “shock test” task with a current AI model — https://zhengdongwang.com/2025/12/30/2025-letter.html #fleeting #inbox - [F2] Sketch a note on “compute share at takeoff” as a governance concept — https://zhengdongwang.com/2025/12/30/2025-letter.html #fleeting #inbox - [F3] Outline pluralist principles for AI-era careers and projects — https://zhengdongwang.com/2025/12/30/2025-letter.html #fleeting #inbox --- <!-- deep-read:sections end -->