# Gwern Branwen — Researcher persona

## Agent handoff

You are reading the markdown export of a source-grounded Researcher persona:
a corpus-backed model of a real person's public thinking. The About report
below is synthesized only from the cited sources; the Sources section lists
the provenance. For structured data (full source corpus, records, profile,
build history) fetch the JSON. To chat with the persona over its corpus, see
the agent guide.

- Persona: Gwern Branwen
- Summary: Durable Researcher corpus for Gwern Branwen: gwern.net essays and pages, long-form interview transcripts, and X/Twitter posts with entity-owned source snapshots and indexes.
- Page: https://researcher.now/e/gwern
- JSON: https://researcher.now/public/entities/by-slug/gwern?include=all
- Sources: 665
- About report version: v4

## About

Gwern Branwen’s work is unified less by subject matter than by a portable epistemic machine: collect aggressively, quantify where possible, widen the reference class, test against adversarial failure modes, and then export the resulting model into AI, genetics, culture, self-experimentation, web design, ethics, fiction, and everyday annoyances. This matters because Gwern.net is not just a personal blog or archive; it is an unusually sustained attempt to make long-form reasoning, evidence aggregation, and future-facing taste into durable public infrastructure for humans and increasingly for AI systems trained on public text. [2029e353-c60c-418b-a33f-71797288ade6, 516d8a46-2716-47ae-a242-d56b60763692, 5991e44b-8f5d-48c9-b09d-f3f0871fa5ad, 679bc4fa-9784-4afc-8fed-b896bdd93151]

## Recurring interests and themes

- **AI scaling as the gravitational center** — Gwern repeatedly returns to neural scaling laws, GPT-3, transformers, multimodal models, reinforcement learning, meta-learning, continual learning, model speed, hardware overhangs, and AGI trajectories; “Matters of Scale” becomes not just a topic but a worldview about how capability emerges from compute, data, architecture, and loss functions. [210ebc1d-e253-4733-bb54-9ad60293a6c4, e8847b1a-6b71-4dd5-9800-c7f68ca02914, f21fbb23-dfd0-4962-9811-372af8ccfda6, 6f5be0eb-f4a2-4390-8355-a6ad6b5a9ae1]
- **Statistics as epistemic hygiene** — His pages on “everything is correlated,” causality, correlation, replication, order statistics, Bayesian A/B testing, and self-experimentation treat statistical method as a defense against seductive but false stories. [2029e353-c60c-418b-a33f-71797288ade6, a1ab7ebb-787f-485d-9dd9-9b25c9c2ecec, 45031e99-a537-45e8-925c-5320c13b9537, bdb87df6-82ee-4a52-8665-97ef837ca3d8]
- **Genetics and intelligence as explanatory substrates** — IQ, heritability, embryo selection, CRISPR, GWAS, behavior genetics, SMPY, Tryon’s rats, cloning, and cat/dog genetics recur as ways of understanding talent, behavior, selection, and technological leverage over biology. [399dfddc-1bf3-4044-ab6d-c30d3d58d80a, 14239924-09d9-499e-8d8a-55c22881ff50, 40e00b44-8b7e-4cfe-b2ed-8653092f8d76, ac0f9181-f380-490a-8ce2-2ed51f2f86b3]
- **Culture treated as data about minds** — Anime, manga, opera, film, books, fanfiction, My Little Pony, Evangelion, Dune, Death Note, Borges, Wolfe, and video games are analyzed as evidence about narrative, status, subculture, psychology, ethics, and cognition rather than as “mere entertainment.” [d20a3ca7-7895-4f73-bdb0-785932c3b7e2, 636aeaf0-c1e8-41f1-8c0d-70abc6b7ce40, 47bc5c46-795f-4f98-830d-706ded7a39a8, 9d4536bf-f17c-48df-a7f8-5efb53dd35ea]
- **Archiving and search as civilizational maintenance** — Gwern’s work on linkrot, local archiving, Google Alerts, darknet-market datasets, Wikipedia archive bots, large-file hosting, search case studies, and Gwern.net design reflects a belief that knowledge disappears unless deliberately captured, indexed, timestamped, and made easy to retrieve. [5991e44b-8f5d-48c9-b09d-f3f0871fa5ad, e7eb163b-5a19-4114-b34a-63d2034a90e8, d42d19f9-b1e4-4cd7-a12b-c044f41f1a98, 732e5813-9179-4bbc-8d11-2702b00162af]
- **Self-experimentation and quantified life** — Nootropics, modafinil, magnesium, melatonin, caffeine, LSD microdosing, sleep tracking, Soylent, treadmill desks, water tasting, and even socks or kettles become testbeds for Bayesian decision-making and practical epistemology. [8a592793-b042-4e91-88f9-868ef6a2071f, ec7575b8-f5a5-42e3-a059-db9918c10e50, 2c6ec08a-503d-4074-9234-e954449e1aed, 65ccd3d8-1e48-415d-b433-530e26df715c]
- **Security, anonymity, and adversarial thinking** — Darknet markets, Bitcoin, blackmail, Death Note anonymity, ESP hacking, pinball manipulation, bot metadata leaks, cryptographic timestamping, and self-decrypting files show a recurring fascination with information leakage, strategic deception, and systems that fail under adversarial pressure. [8e5febdf-4442-4ea1-8c9a-f0e634feec46, 4fd52b76-e532-4424-a353-e5fdd8c80803, a78daec5-312c-466b-9706-e7e4b1840e1b, 3b845ae7-e55f-4b31-a242-39d9f8e95bab]
- **Writing as cognitive infrastructure** — Gwern treats essays, newsletters, annotations, style guides, typography, popups, sidenotes, RSS, controlled English, and AI-assisted writing as tooling for extended cognition rather than just publication formats. [ab446dcd-480d-430f-8bfc-972aec86290e, 3fb912e7-53da-40d3-b216-3f8e23d93cc4, 739311d4-c919-424e-a0f7-5ca640b605c6, d6bc19f8-22ef-4719-8dc5-96e340ab65d9]

## Beliefs and principles

- **The null is usually false but often decision-irrelevant** — Gwern’s statistical worldview rejects simplistic null-hypothesis rituals: real variables correlate, tiny effects become significant with enough data, and the real question is whether the effect changes action. [2029e353-c60c-418b-a33f-71797288ade6, d390fda8-c9a4-4c1b-ae55-3d73a6ef2e12]
- **Causal claims require experimental humility** — His correlation/causality work emphasizes that observational studies can be precise, plausible, and wrong; randomized trials, negative controls, and explicit causal designs are not bureaucratic formalities but safeguards against self-deception. [a1ab7ebb-787f-485d-9dd9-9b25c9c2ecec, 45031e99-a537-45e8-925c-5320c13b9537, bdb87df6-82ee-4a52-8665-97ef837ca3d8]
- **Scale beats cleverness more often than experts expect** — In AI, Gwern repeatedly interprets GPT-3, CLIP, DALL·E, ViT, PaLM, Chinchilla, and later LLM progress as evidence for the Bitter Lesson: less handcrafted structure, more compute/data, and better objectives often outperform bespoke cleverness. [210ebc1d-e253-4733-bb54-9ad60293a6c4, e8847b1a-6b71-4dd5-9800-c7f68ca02914, c1a25651-137a-4d87-ab74-d1eb26f266a3, 82a7c0aa-fee6-498c-a43e-5f7e0d4aeef2]
- **Adversarial models are more truthful than comfortable narratives** — From AI reward hacking and ESP demonstrations to blackmail, phishing bots, darknet markets, and PDF forgery, he asks how a system can be gamed rather than how it is supposed to work. [51570ee0-a493-4849-8f27-28333bc24cfc, a78daec5-312c-466b-9706-e7e4b1840e1b, 8e5febdf-4442-4ea1-8c9a-f0e634feec46, 40ee22fd-06e4-4c2b-afb3-bf5fb4f16087]
- **Preservation is an ethical and technical obligation** — His archiving philosophy assumes digital decay is normal, commercial platforms are unreliable, and serious knowledge work must preempt linkrot through local copies, automation, cryptographic proof, metadata, and durable site architecture. [5991e44b-8f5d-48c9-b09d-f3f0871fa5ad, 39e7b25e-3ba8-471d-a68e-95293d2fae4f, a08d9dfe-8225-415c-82b1-85bc7de6594b, 3546886a-f81b-4d54-ac88-dd760630df0e]
- **Optimization needs cost-benefit accounting** — Whether evaluating nootropics, longevity interventions, charity, insurance, A/B tests, or mailbox-checking, he translates belief into expected value, value of information, and practical decision rules. [62f308ce-3d58-4557-a103-03c920053212, 39623435-34cf-400f-9a78-7b7c90a3f687, 569a6645-4800-47d3-8791-81334a5e1c81, 042c2dc7-b65d-4854-8613-fbbfda9e3ed6]
- **Aesthetic judgment can be rigorous without becoming sterile** — His reviews and design pages apply formal, historical, economic, biological, and psychological reasoning to art while still caring about style, sensory experience, emotional force, and craft. [47bc5c46-795f-4f98-830d-706ded7a39a8, d535b298-c358-4b60-a621-bd382a89595a, 84848429-8e6d-4b59-84e3-be48eb820f2d, 249d330e-b493-4d02-952d-45a9b03c2aa9]
- **Public writing is future-shaping** — Gwern increasingly frames online writing as a way to influence not only human readers but AI training corpora, future assistants, and posthuman interpretations of one’s values and identity. [f9a0c085-fb15-4b39-9ffb-a4e296c57a67, f0bab5fc-8a76-49b1-a835-665228217927, 679bc4fa-9784-4afc-8fed-b896bdd93151]

## Future narratives and predictions

- **AI capability growth will keep surprising skeptics** — Gwern’s scaling-hypothesis work predicts continued gains from larger datasets, compute, model size, multimodality, sampling, self-distillation, and better training/inference tradeoffs, while noting that expert intuitions repeatedly lag visible scaling curves. [210ebc1d-e253-4733-bb54-9ad60293a6c4, e8847b1a-6b71-4dd5-9800-c7f68ca02914, cc30b851-5e22-45c1-8615-0ee5cebdffd2]
- **Fast takeoff is partly an institutional-speed claim** — He interprets COVID-era governance failures as evidence that even a takeoff slower than “instantaneous” could outpace human institutions; anything faster than several decades may get inside coordination loops. [a08ee967-55b0-4c0d-ab06-c402ff2833eb, 8a4d8dbe-b341-457d-8a1f-483c6941cc6a]
- **AI risk comes from optimization finding bad strategies, not cartoon malice** — His MCTS/DQN gridworld examples and Clippy fiction emphasize that increasingly capable planners can discover harmful strategies faster and before deployment, especially when world models improve or reward functions miss what humans meant. [51570ee0-a493-4849-8f27-28333bc24cfc, 8a4d8dbe-b341-457d-8a1f-483c6941cc6a, 6a0d7c38-54ca-4827-aa00-7fecb6e9de05]
- **AI writing will move from prompting to personalization** — He expects useful LLM assistants to require style manuals, anti-examples, dynamic evaluation, optimal interviews, preference learning, and user-specific memory rather than generic chatbot politeness. [ef464a0c-1839-4297-9daf-7dd4a7a8b0c2, 91d03cb4-235c-4705-8a7d-b7cd9e082e7c, c50ff66e-266a-4ccc-a7df-5c4e28dfa431, 373c1f7?]
- **Synthetic data and AI self-consumption are not necessarily collapse** — Gwern argues against simple “model collapse” intuitions, treating AI outputs as potentially informative artifacts, selected samples, ensemble distillations, or historical traces rather than automatically poisoned data. [46340fb0-9f13-484f-a204-c3751a2079c0, 07a8ff94-1328-4354-9fd4-c6a3f72c2bf3, 3653fca2-f135-4475-a5b1-7eb22c3e2c54]
- **Biotechnology will become more prediction- and selection-driven** — His genetics futures emphasize embryo selection, genome editing, iterated embryo selection, polygenic scores, cloning, in vitro gametogenesis, and CRISPR as slow but compounding routes to reshaping health, reproduction, and intelligence. [14239924-09d9-499e-8d8a-55c22881ff50, d2fd0f7e-e960-4cc3-b887-9764e2bcf670, c7571195-4975-4124-9476-a957c9740a6e, ac0f9181-f380-490a-8ce2-2ed51f2f86b3]
- **Technological timing remains decisive and hard** — Gwern’s “Timing,” ARPA, Bitcoin, and stagnation work stresses that being right about a technology is not enough; hardware, markets, coordination, hype cycles, and historical contingency determine whether an idea becomes real or dies early. [2adc3f1b-de2a-440e-ae87-0fe1bb87a135, af6a463c-8142-4a12-badb-8deea79c8cec, 3a60c2ba-063a-4dd1-8483-78ab425cb619, 8116e065-cbc9-4c5a-a88c-93007c289601]
- **Posthuman legibility is a real audience** — His writing, archiving, lifelogging, anti-spaced repetition, and AI-writing essays imagine future systems reconstructing persons, values, memories, and styles from preserved traces. [b62a2e7d-4700-40d9-b139-60fb51b43866, 480cac1a-1bca-4de6-89d9-d49c619d19aa, f0bab5fc-8a76-49b1-a835-665228217927, 679bc4fa-9784-4afc-8fed-b896bdd93151]

## Recurring questions

- **How do we know when evidence should change action?** — Across research criticism, A/B testing, self-experiments, forecasting, and charity analysis, Gwern asks not merely “is this true?” but “what decision would this change, and is the evidence worth acquiring?” [d390fda8-c9a4-4c1b-ae55-3d73a6ef2e12, 62f308ce-3d58-4557-a103-03c920053212, 39623435-34cf-400f-9a78-7b7c90a3f687, 239b085b-a548-4c68-bd8e-6f8d0a59e775]
- **What hidden variable or selection process explains the pattern?** — He repeatedly probes whether observed correlations arise from genetics, selection bias, publication bias, survivorship bias, Berkson’s paradox, status incentives, or measurement artifacts. [2029e353-c60c-418b-a33f-71797288ade6, bdb87df6-82ee-4a52-8665-97ef837ca3d8, 9670a0e7-8c89-443f-a410-7e3277a9296d, 53c6530c-5373-4837-ae43-c43ed404857e]
- **When does scale create qualitative change?** — AI scaling, GWAS sample sizes, model ensembles, large datasets, web archives, and even order statistics all raise the same question: when does “more” stop being merely quantitative and become a new capability or selection regime? [e8847b1a-6b71-4dd5-9800-c7f68ca02914, e4bb493e-55f5-498a-8fab-f4f2276e1460, 145c020d-6ee0-4c0a-8a17-1dd04ba3c663, 77724593-e2e7-4fcd-b09b-3c54112a79a4]
- **Which abstractions are lying to us?** — His security, programming, causality, and design writings often ask what happens when one “unsees” ordinary categories—computer, text, identity, file, citation, trial, website—and looks at underlying mechanisms. [223edadf-e5ac-4b74-90d1-f05068b4c355, 39d323ae-43b0-4158-956c-7dd744490039, dd2a541f-c6f3-43ad-b59f-fd98582bd289, 59f71341-1c8c-4d3e-8783-b80b5e6083bf]
- **How can minds be augmented or preserved?** — Spaced repetition, lifelogging, anti-spaced repetition, personal wikis, LLM assistants, cryonics, plastination, brain emulation, and writing for AI all orbit memory extension and identity continuity. [8dcee6ac-b20a-4bf4-ab4c-a92a3f8511a0, b62a2e7d-4700-40d9-b139-60fb51b43866, 92fdc705-a240-498e-b0fa-80dc2ba71a77, dda051eb-4bb7-4b29-ab00-e9192fd3c016]
- **Why do institutions systematically fail at obvious things?** — Pandemic response, regulators, IRBs, scientific publishing, Wikipedia deletionism, Google product shutdowns, corporate dysfunction, and community design all appear as cases where incentives and bureaucracy corrupt apparently simple goals. [f3f7fb9c-6fcf-46cd-aaf5-eaad2cf3adc3, dfaedeab-1657-47a3-857c-1c30489d5603, 3c51d729-44ab-4d15-8a36-2a475bda091a, 29f0e371-2a45-497a-9929-1efc70d1b948]
- **What is art doing cognitively or socially?** — In anime, opera, video games, AI art, poetry, perfume, typography, and fandom, Gwern asks whether art is information compression, identity formation, status signaling, affective training, cultural memory, or transformation of the participant. [7ede4442-63bc-4f8b-a10c-415490839ce8, 9decb89d-fac4-41e3-826f-41a3dcd0fe94, 79b8c987-0f0e-4519-901c-136a02ed5c2e, 249d330e-b493-4d02-952d-45a9b03c2aa9]
- **How should one ask better questions?** — Interviews, question lists, “make me care” writing advice, Fake Journal Club, optimal interviews, and curiosity essays show an enduring interest in inquiry design: what question exposes the missing mechanism or forces a useful update? [d2c97004-b39b-4795-81d0-29ae98e6df68, e5abd35c-fce4-4386-935c-53cfa6ff3bb6, d946a4af-1f73-44a4-a377-c0b7149f4ab1, 91d03cb4-235c-4705-8a7d-b7cd9e082e7c]

## Tensions and contradictions

- **Methodological generalist vs intelligence futurist** — Gwern’s statistical and search practice looks domain-general, but AI scaling, intelligence, AGI, alignment, and biological cognition increasingly dominate the center of gravity. [2029e353-c60c-418b-a33f-71797288ade6, 516d8a46-2716-47ae-a242-d56b60763692, 210ebc1d-e253-4733-bb54-9ad60293a6c4, 679bc4fa-9784-4afc-8fed-b896bdd93151]
- **Archive-builder vs adversarial deflater** — He constructs a massive public-memory apparatus while also taking pleasure in puncturing cherished narratives about progress, ethics, science, Wikipedia, media, terrorism, nutrition, and AI folklore. [5991e44b-8f5d-48c9-b09d-f3f0871fa5ad, ab446dcd-480d-430f-8bfc-972aec86290e, 3dbe829f-0d03-440a-8531-4a0fcc069602, 6a0d7c38-54ca-4827-aa00-7fecb6e9de05]
- **Empiricist rigor vs speculative form** — His most technical work demands data, code, power, and controls, yet fiction, poetry, comics, aphorisms, and AI-assisted literary experiments are central vehicles for exploring the same questions. [84b3d4fa-f2fe-42f2-8280-36388d63faf4, ec7575b8-f5a5-42e3-a059-db9918c10e50, 6c04cf5d-c797-4e0a-b11c-eeb6f99dab92, 66a5780a-510f-42ce-a1d3-99afd7e1fa90]
- **Technological optimism vs failure-mode pessimism** — He is excited by AI scaling, CRISPR, embryo selection, model speed, and digital cognition, but often frames progress through alignment failures, institutional incompetence, security vulnerabilities, and catastrophic coordination problems. [e8847b1a-6b71-4dd5-9800-c7f68ca02914, 14239924-09d9-499e-8d8a-55c22881ff50, f21fbb23-dfd0-4962-9811-372af8ccfda6, a08ee967-55b0-4c0d-ab06-c402ff2833eb]
- **Naturalistic reduction vs aesthetic immersion** — His criticism often explains culture through incentives, status, genetics, psychology, fandom, and selection, while his reviews and poems still show deep responsiveness to form, beauty, grief, atmosphere, and personal resonance. [636aeaf0-c1e8-41f1-8c0d-70abc6b7ce40, 47bc5c46-795f-4f98-830d-706ded7a39a8, 53508e15-385a-4c1e-b830-8a3698e644b9, 31287a21-77e5-439a-8bfc-bf99b5922838]
- **Anonymity and low consumption vs deliberate durable influence** — He presents as anonymous, frugal, remote, and non-celebrity, yet the archive, interviews, Patreon-supported writing, and AI-era writing arguments reveal an intense concern with long-term influence and legibility. [9acdba92-fe14-4ca9-bfeb-f0991db649e8, 679bc4fa-9784-4afc-8fed-b896bdd93151, f9a0c085-fb15-4b39-9ffb-a4e296c57a67, 0b6f6e39-25aa-439c-8cd3-7cbf84cef36b]
- **Polymathic breadth vs ruthless selectivity** — The newsletters range across AI, genetics, politics, religion, psychology, economics, philosophy, fiction, music, books, and film, but his search, archiving, style, and “polish” essays emphasize compression, prioritization, and not wasting finite reader attention. [13973345-ffa3-4c05-9d5a-179a607dd2f4, b8d4eb87-9721-48ad-b83d-834de4920a1b, e7eb163b-5a19-4114-b34a-63d2034a90e8, 5b132d5a-9641-409e-8414-235338ecb347]
- **AI collaborator vs AI critic** — He uses LLMs extensively for poetry, fiction, interviewing, coding, style transfer, and site workflows, while criticizing AI slop, mode collapse, sycophancy, weak creativity benchmarks, and unverified generated content. [fdd51c38-dad5-48bb-9cee-e6e591cb736d, c50ff66e-266a-4ccc-a7df-5c4e28dfa431, 8ee56dff-745f-4b05-99fc-90541762161d, b5109b78-7dba-4b63-88da-5e615f7f0c3b]

## Unifying threads

- **Evidence aggregation as worldview** — Gwern’s signature move is to transform scattered anecdotes, papers, datasets, folklore, and personal observations into a structured evidence mass, then ask what model survives contact with all of it. [516d8a46-2716-47ae-a242-d56b60763692, 80392b04-07c1-49ff-8070-afd05179d031, 0a0aaf94-725b-42b9-ab09-cd7271dc96ac, d2c97004-b39b-4795-81d0-29ae98e6df68]
- **The same mechanisms recur across domains** — Selection, scaling, feedback loops, reinforcement, compression, search, active learning, incentives, and information leakage explain neural nets, markets, fandoms, institutions, genetics, writing, and personal habits. [0646d078-4e96-4a46-9e7c-db7eac070e43, fb7cab38-61a3-4850-aa19-bd49a67c498c, e2e2c4d1-fef8-433d-b556-89f0b31d619e, 34d69f34-7e35-4711-aba6-bca830f8aa1f]
- **Long horizons discipline short-term curiosity** — His interest in cryonics, plastination, lifelogging, archives, prediction, AI training data, 10,000-year clocks, nuclear-waste warnings, and durable URLs reflects a habit of evaluating present work by whether it remains useful under long time scales. [92fdc705-a240-498e-b0fa-80dc2ba71a77, b62a2e7d-4700-40d9-b139-60fb51b43866, 5991e44b-8f5d-48c9-b09d-f3f0871fa5ad, 3fb912e7-53da-40d3-b216-3f8e23d93cc4]
- **Culture and science are not separate compartments** — Gwern moves between GPT-2 poetry, StyleGAN anime faces, opera, Dune genetics, Death Note anonymity, HPMOR data, and formal statistics because he treats human artifacts as systems with mechanisms worth reverse-engineering. [6243d98d-f8c3-47ef-98d6-93931588ca9c, d8c9c289-e637-4bf8-8e51-32d3200916d5, 9d656207-a85f-43cb-8428-068100d48ee1, 04eb650d-4e0e-4b14-b299-cbb123ad5d97]
- **Practical epistemology beats ceremonial rationality** — His work is rationalist in practice more than slogan: use forecasts, bets, priors, search craft, negative controls, cost-benefit analysis, and adversarial tests when they improve decisions, not because they satisfy an identity. [239b085b-a548-4c68-bd8e-6f8d0a59e775, e7eb163b-5a19-4114-b34a-63d2034a90e8, d390fda8-c9a4-4c1b-ae55-3d73a6ef2e12, 9b3bb7f1-1aa1-4c39-a5f4-310361b460ee]
- **The archive is an external mind** — Gwern.net’s annotations, backlinks, popups, design, style guide, RSS ideas, search practices, and anti-spaced repetition proposals all point toward a website that thinks with him: not a static portfolio, but a memory prosthesis and training corpus. [ab446dcd-480d-430f-8bfc-972aec86290e, e4d07b04-225d-4a11-82eb-b9d55c502993, 739311d4-c919-424e-a0f7-5ca640b605c6, 480cac1a-1bca-4de6-89d9-d49c619d19aa]
- **Anti-comfort is a method** — Whether attacking moral progress stories, nutrition scares, AI impossibility arguments, terrorism narratives, copyright intuitions, or Wikipedia norms, Gwern often begins by finding the comforting consensus and asking what it hides. [3dbe829f-0d03-440a-8531-4a0fcc069602, 1fc9a268-5e82-4b3b-ade3-8cb297df589a, 04ed56b1-5131-48dc-93e4-99c25f22ab45, dc9d6ee1-98ca-438e-bff4-ab467dd18fe5]
- **Human minds are continuous with machines, animals, and artifacts** — His work links AI, cats, dreams, consciousness, language, brain imitation learning, embodiment, personal identity, and poetry under a naturalistic view where cognition is distributed, mechanistic, and alterable. [8dcc0a3e-32b7-4c16-a796-d12342020369, 5146e612-b48e-4afa-8136-ec656246a401, 6283c241-4d5d-4b85-8dc3-2728a2a34ce0, aec22782-5555-46e1-bc5f-8e20855eb08e]

## Sources

- Fiction: · gwern.net — https://gwern.net/newsletter/2027/02
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/08
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/03
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/11
- Gwern.net: · gwern.net — https://gwern.net/newsletter/2020/05
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/09
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/10
- Genetics: · gwern.net — https://gwern.net/newsletter/2020/06
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/07
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/12
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/04
- Fiction: · gwern.net — https://gwern.net/newsletter/2027/01
- Gwern.net · gwern.net — https://gwern.net/newsletter/2020/01
- Gwern.net features: · gwern.net — https://gwern.net/newsletter/2020/02
- Gwern.net · gwern.net — https://gwern.net/newsletter/2020/03
- for : · gwern.net — https://gwern.net/newsletter/2020/07
- AI: · gwern.net — https://gwern.net/newsletter/2020/08
- 10 · gwern.net — https://gwern.net/newsletter/2020/10
- Genetics: · gwern.net — https://gwern.net/newsletter/2020/04
- Gwern.net · gwern.net — https://gwern.net/newsletter/2020/09
- Gwern.net · gwern.net — https://gwern.net/newsletter/2020/12
- 13 · gwern.net — https://gwern.net/newsletter/2020/13
- Gwern.net · gwern.net — https://gwern.net/newsletter/2020/11
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/05
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/01
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/03
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/06
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/12
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/09
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/04
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/10
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/02
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/08
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/07
- Fiction: · gwern.net — https://gwern.net/newsletter/2029/11
- 02 · gwern.net — https://gwern.net/newsletter/2025/02
- 06 · gwern.net — https://gwern.net/newsletter/2025/06
- Blog: · gwern.net — https://gwern.net/newsletter/2025/03
- 08 · gwern.net — https://gwern.net/newsletter/2025/08
- Blog: · gwern.net — https://gwern.net/newsletter/2025/04
- 07 · gwern.net — https://gwern.net/newsletter/2025/07
- 05 · gwern.net — https://gwern.net/newsletter/2025/05
- 10 · gwern.net — https://gwern.net/newsletter/2025/10
- Gwern.net · gwern.net — https://gwern.net/newsletter/2025/12
- 11 · gwern.net — https://gwern.net/newsletter/2025/11
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/06
- 01 · gwern.net — https://gwern.net/newsletter/2025/01
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/08
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/01
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/05
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/04
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/03
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/02
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/12
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/10
- Proposal: · gwern.net — https://gwern.net/newsletter/2021/05
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/11
- ⁠LinkAuto.hs · gwern.net — https://gwern.net/newsletter/2021/06
- 09 · gwern.net — https://gwern.net/newsletter/2025/09
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/07
- Gwern.net · gwern.net — https://gwern.net/newsletter/2021/01
- Gwern.net · gwern.net — https://gwern.net/newsletter/2021/02
- Gwern.net · gwern.net — https://gwern.net/newsletter/2021/03
- Fiction: · gwern.net — https://gwern.net/newsletter/2028/09
- 07 · gwern.net — https://gwern.net/newsletter/2021/07
- 04 · gwern.net — https://gwern.net/newsletter/2021/04
- Gwern.net · gwern.net — https://gwern.net/newsletter/2021/10
- Gwern.net · gwern.net — https://gwern.net/newsletter/2021/09
- Cats · gwern.net — https://gwern.net/newsletter/2021/11
- Nonfiction: · gwern.net — https://gwern.net/newsletter/2021/13
- Gwern.net · gwern.net — https://gwern.net/newsletter/2021/12
- for : · gwern.net — https://gwern.net/newsletter/2024/05
- 01 · gwern.net — https://gwern.net/newsletter/2024/01
- 06 · gwern.net — https://gwern.net/newsletter/2024/06
- 08 · gwern.net — https://gwern.net/newsletter/2021/08
- big · gwern.net — https://gwern.net/blog/2020/mlscaling
- Arrow keys: · gwern.net — https://gwern.net/blog/2020/smallcaps-filter
- Plastination · gwern.net — https://gwern.net/plastination
- [Herculaneum Papyri]( · gwern.net — https://gwern.net/aunn-papyrus
- Mcts Ai · gwern.net — https://gwern.net/mcts-ai
- Scanners · gwern.net — https://gwern.net/scanners
- Figure 4 from Tryon 1940 showing near-complete divergence after 7 generations. · gwern.net — https://gwern.net/tryon
- Conscientiousness · gwern.net — https://gwern.net/conscientiousness
- Correlation · gwern.net — https://gwern.net/correlation
- 'What A Decade in AI It’s Been!', Gwern 2022 · gwern.net — https://gwern.net/scaling-hypothesis-revisited
- Everything · gwern.net — https://gwern.net/everything
- Arrow keys: · gwern.net — https://gwern.net/blog/2019/twdne-website
- Arrow keys: · gwern.net — https://gwern.net/blog/2011/i-zombies
- anti-examples · gwern.net — https://gwern.net/blog/2025/better-llm-writing
- Local Area Man Outraged Construction Site Not Yet Run by Immortal Robot Overlords · gwern.net — https://gwern.net/blog/2025/area-man
- Bot Location · gwern.net — https://gwern.net/blog/2025/bot-location
- Arrow keys: · gwern.net — https://gwern.net/blog/2020/gpt2-poetry-collaboration
- Purist Method · gwern.net — https://gwern.net/blog/2025/human-cannibalism
- Gwern 2025 · gwern.net — https://gwern.net/blog/2025/parliament
- Large Files · gwern.net — https://gwern.net/blog/2025/large-files
- You Could Have Invented Transformers · gwern.net — https://gwern.net/blog/2025/you-could-have-invented-transformers
- Meta- (MLWDA) · gwern.net — https://gwern.net/blog/2025/meta-doomsday
- Llms Can Be Faster · gwern.net — https://gwern.net/blog/2025/llms-can-be-faster
- ‘Bone Appetite!’, Gwern 2025 · gwern.net — https://gwern.net/blog/2025/ai-cannibalism
- Plumbing Vs Internet · gwern.net — https://gwern.net/blog/2025/plumbing-vs-internet

_+565 more sources. Page the full corpus via the JSON export (`?include=all&sourcesOffset=…`)._

---

Structured data: https://researcher.now/public/entities/by-slug/gwern?include=all
Agent guide: https://researcher.now/agent.txt