Gwern Branwen writes and researches under a pseudonym. Gwern.net hosts his essays, research notes, reviews, and self-experiments across AI, psychology, statistics, and other subjects. — Gwern — About Me.
On AI and Technology
Gwern has written extensively on machine-learning scaling, including the argument that additional compute, data, and model scale can unlock capabilities. — Gwern — The Scaling Hypothesis.
Ideas:
- Automating from the bottom up: Gwern speculates that AI may replace lower-level company work first, leaving human executives in supervisory roles longer. — Dwarkesh Patel Interview.
- Intelligence as search: Gwern proposes a high-level model of intelligence as search over possible programs, rather than a single hand-designed algorithm. — Dwarkesh Patel Interview.
- Diverse AI minds: Gwern says different language models can have discernibly different cognitive styles, while qualifying that this comparison sets capability aside. — Dwarkesh Patel Interview.
- Write for future models: Gwern argues that writing can shape future AI systems because published text becomes part of the material models learn to predict. — Dwarkesh Patel Interview.
- Sketch ideas before AGI: Gwern suggests specifying valuable projects and why they matter now, even if more capable future systems might implement them later. — Dwarkesh Patel Interview.
- Taste remains personal: Even if AI can execute tasks, Gwern argues it cannot decide what an individual personally values or enjoys on that individual’s behalf. — Dwarkesh Patel Interview.
- Human vision with AI execution: Gwern imagines a company in which a human sets a vision while AI systems carry out much of the work beneath that role. — Dwarkesh Patel Interview.
- No single master algorithm: Gwern doubts that intelligence depends on one special algorithm; he instead emphasizes many learned special cases and computation. — Dwarkesh Patel Interview.
Further lessons:
- Scaling can produce new abilities: Gwern’s scaling hypothesis holds that larger neural networks trained with more data and compute can acquire capabilities not obvious from smaller systems. — Gwern — The Scaling Hypothesis.
- The bitter lesson and compute: Gwern connects his scaling argument to Rich Sutton’s “Bitter Lesson”: general methods that exploit growing computation can outperform hand-crafted approaches. — Gwern — The Scaling Hypothesis.
- Test generative writing empirically: Gwern’s GPT-3 experiments probe where generated poetry and fiction can work and where human curation remains important. — Gwern — GPT-3 Creative Fiction.
- Data diversity matters: Gwern argues that scale includes not only model size and compute but also the amount and diversity of training data; the original Pile attribution is removed. — Gwern — The Scaling Hypothesis.
- Curate AI-safety work: Gwern maintains a bibliography of research on AI safety and alignment; this supports a narrower claim than saying he developed an alignment theory. — Gwern — AI Safety Bibliography.
On Self-Experimentation and Learning
Gwern has published self-experiments, including blinded comparisons and analyses of their results. — Gwern — Self-Blinded Water Test.
Ideas:
- Test claims personally: Gwern’s self-blinded water experiment illustrates his practice of testing a claim for himself rather than relying only on received opinion. — Gwern — Self-Blinded Water Test.
- Biography can be misleading: Gwern observes that readers often want biographical details that are less predictive than validated psychological measures. — Gwern — About Me.
Further lessons:
- Use spaced repetition for memory: Gwern reviews spaced-repetition systems and the evidence that appropriately spaced reviews can improve long-term recall. — Gwern — Spaced Repetition.
- Blind self-experiments where possible: Gwern’s water-taste experiment shows how blinding and randomization can separate perceived effects from expectation. — Gwern — Self-Blinded Water Test.
- Question far transfer from dual n-back: Gwern’s meta-analysis finds that apparent IQ gains in dual n-back studies are strongly associated with weak controls, leaving little evidence for improved underlying fluid intelligence. — Gwern — Dual n-Back Meta-Analysis.
- Make self-tracking testable: Gwern publishes protocols and data from personal experiments, illustrating a quantified-self approach that can be examined by others. — Gwern — Self-Blinded Water Test.
- Distinguish lithium correlations from proof: Gwern reviews observational research associating trace lithium in drinking water with population-level outcomes, while treating causal interpretation as uncertain. — Gwern — Lithium in Drinking Water.
On Psychology and Society
Gwern also writes about psychological biases, internet subcultures, and the behavior of social groups. — Gwern — Melancholy of Subculture.
Ideas:
- Anonymity buys a hearing: Gwern argues that pseudonymity can make readers engage with writing before assigning its author a social category. — Dwarkesh Patel Interview.
- Make useful work reusable: Gwern values factoring useful functionality into a library that others actually adopt, even when this brings little personal credit. — Gwern — About Me.
- Draw on others’ work: Gwern argues that borrowing and improving prior solutions is often better than reinventing them for credit. — Gwern — About Me.
Further lessons:
- Connection can fragment culture: Gwern argues that the internet helps narrow subcultures find one another while potentially weakening shared culture and leaving their members isolated from wider society. — Gwern — Melancholy of Subculture.
- Sunk costs are not always simple: Gwern argues that some behavior labelled sunk-cost bias can act as a learning heuristic, while clear examples of fallacy may arise at organizational levels. — Gwern — Are Sunk Costs Fallacies?.
- Look beyond terror as the motive: Gwern argues that terrorist activity can be better explained by status, belonging, and internal group incentives than by the goal of maximizing public fear. — Gwern — Terrorism Is Not About Terror.
- Ask what collecting satisfies: Gwern examines collecting as a distinct, sometimes puzzling behavior and considers possible psychological and social explanations. — Gwern — The Collecting Mindset.
- Parasocial ties are a changing category: Gwern curates research on online parasocial relationships that fall between passive audience contact and reciprocal friendship. — Gwern — Parasocial Relationships Online.
On Writing and Research
Gwern explains his writing and website practices in first-person essays about his work and the site’s design. — Gwern — About This Website.
Ideas:
- Prefer useful incremental work: Gwern credits Wikipedia’s inclusionist culture for an ethos of reusing imperfect work and making steady small improvements over chasing credit. — Gwern — About Me.
- Record subjective judgment: Gwern says he wants to write down preferences, desires, evaluations, and judgments that a model cannot choose for him. — Dwarkesh Patel Interview.
- Pseudonymity reduces exposure: Gwern says pseudonymity has helped protect him from harassment while also letting readers encounter the work before judging his identity. — Dwarkesh Patel Interview.
Further lessons:
- Let essays accrete over time: Gwern describes long essays as the result of accumulating notes and noticing patterns over time, rather than a rejection of all short-form writing. — Dwarkesh Patel Interview.
- Make sources inspectable: Gwern’s site uses citations, bibliographies, and link previews so readers can follow the trail behind an essay. — Gwern — Website Design.
- Publish the website’s machinery: Gwern documents the implementation of Gwern.net and links its source code, making much of the site’s design inspectable. — Gwern — Website Design.
- Use reviews as research: Gwern publishes detailed reviews that combine description, criticism, and related research, as illustrated by his My Little Pony essay. — Gwern — MLP: Immanentizing the Equestrian.
Miscellaneous Insights
A selection of other interesting ideas from across his work.
Ideas:
- Intelligence is not always adaptive: Gwern argues that general intelligence is only one possible evolutionary solution and may not pay for every organism. — Dwarkesh Patel Interview.
- Fluid intelligence is distributed: Gwern argues against locating fluid intelligence in a single brain region, using that as part of his case against a simple intelligence mechanism. — Dwarkesh Patel Interview.
Further lessons:
- Study darknet markets empirically: Gwern analyzed Silk Road’s design, incentives, and practical operation rather than treating darknet markets only as an abstract policy debate. — Gwern — Silk Road.
- Fashion ideas move in cycles: Gwern’s fashion note collects evidence and models for cycles in consumer taste and ideas; it does not claim he built the mathematical models himself. — Gwern — Fashion Cycles.
- Small groups can matter disproportionately: Gwern gathers examples of tightly communicating, highly competitive groups that achieved outsized advances. — Gwern — Unreasonable Small Groups.
- Analyze popular culture seriously: Gwern’s My Little Pony review treats the show and its fandom as subjects for literary and cultural analysis. — Gwern — MLP: Immanentizing the Equestrian.
- Investigate changes in beauty: Gwern examines whether physical attractiveness has changed over time using historical art, photographs, and other imperfect evidence. — Gwern — Progress in Beauty.
- Keep older research accessible: Gwern maintains an archive of historical papers and linked sources alongside new research, making older material easier to revisit. — Gwern — Document Archive.
- Compare correlation with experiments: Gwern compiles studies that compare observational findings with randomized results for the same intervention, testing when correlations track causal effects. — Gwern — How Often Does Correlation = Causality?.
- Learn useful capabilities end to end: Gwern argues that useful capabilities such as curiosity need not be hand-built into AI agents if they can emerge through optimizing training rewards. — Gwern — AIs May Not Need Sleep or Play.
Learn more: