Gwern marks many of his posts with an importance score, but doesn’t provide a sorted index. This is that index. It updates every week.

Currently, 418 pages have an importance rating. Pages within each score are ordered newest first by creation date.

10 13 pages

  1. I propose an approach for highly personalized LLMs, for near-future productivity gains and personal info/cybersecurity against increasingly powerful LLMs: they should, in the spirit of uploading, try to emulate the user’s values and preferences in order to amplify the principal—not replace them. I discuss a package of techniques and proposals to accomplish such ‘guardian angels’; dynamic evaluation of LLMs combined with active learning and elicitation and heavy inner-monologue search/data-augmentation.

    13846 words finished possible

  2. Speculative proposal to create artificial neural nets with human-like performance by high-learning-rate/regularization training of overparameterized NNs to trigger catapulting/grokking. Over-parameterization as a route to true generalization would resolve many outstanding mysteries of artificial versus natural intelligence.

    16187 words finished unlikely

  3. On GPT-3: meta-learning, scaling, implications, and deep theory. The scaling hypothesis: neural nets absorb data & compute, generalizing and becoming more Bayesian as problems get harder, manifesting new abilities even at trivial-by-global-standards-scale. The deep learning revolution has begun as foretold.

    22801 words finished likely

  4. Preliminary draft notes towards a cost-benefit analysis of CRISPR-based editing for complex human traits with 2015–2016 SOTAs; obsolete.

    25501 words abandoned likely

  5. A cost-benefit analysis of the marginal cost of IVF-based embryo selection for intelligence and other traits with 2016–2017 state-of-the-art.

    103474 words finished likely

  6. Attempts at considering the profitability of life-extension interventions for healthy adults

    16584 words in progress likely

  7. Compilation of studies comparing observational results with randomized experimental results on the same intervention, compiled from medicine/economics/psychology, indicating that a large fraction of the time (although probably not a majority) correlation ≠ causality.

    20774 words in progress log

  8. Correlations are oft interpreted as evidence for causation; this is oft falsified; do causal graphs explain why this is so common, because the number of possible indirect paths greatly exceeds the direct paths necessary for useful manipulation?

    8249 words in progress highly likely

  9. Critics of AI risk suggest diminishing returns to computing (formalized asymptotically) means AI will be weak; this argument relies on a large number of questionable premises and ignoring additional resources, constant factors, and nonlinear returns to small intelligence advantages, and is highly unlikely.

    9811 words finished likely

  10. Mathematical mistake/error-rates limit our understanding of rare risks and ability to defend against them.

    17844 words finished likely

  11. 3 essays by AI researcher Douglas Hofstadter exploring cooperation/game theory/‘superrationality’ in the context of the failure of political coordination to prevent global nuclear war (published 1985, from Scientific American).

    21261 words By Douglas Hofstadter finished log

  12. Why most supplements fail: IQ improvement skepticism, Yudkowsky & Bostrom's heuristics, nootropics

    24362 words finished highly likely

  13. MAD will not work in outer space; pre-emptive strikes are nigh-guaranteed.

    5524 words finished unlikely

9 23 pages

  1. Fictional short story about Clippy & AI hard takeoff scenarios grounded in contemporary ML scaling, self-supervised learning, reinforcement learning, and meta-learning research literature.

    13415 words fiction finished log

  2. Bibliography of ML scaling papers showing smooth scaling of neural net performance in general with increasingly large parameters, data, & compute.

    3514 words finished log

  3. AIs limited to pure computation (Tool AIs) supporting humans, will be less intelligent, efficient, and economically valuable than more autonomous reinforcement-learning AIs (Agent AIs) who act on their own and meta-learn, because all problems are reinforcement-learning problems.

    13259 words finished likely

  4. 2015 survey of online modafinil users asking about dosages, consumption patterns, ratings, purchases, and demographics

    6177 words in progress highly likely

  5. Mirrors of ~89 Tor-Bitcoin darknet markets & forums 2011–2015, and related material.

    11936 words finished highly likely

  6. On desirable correlates of intelligence

    27073 words notes highly likely

  7. Problems with social experiments and evaluating them, loopholes, causes, and suggestions; non-experimental methods systematically deliver false results, as most interventions fail or have small effects.

    12121 words By Peter H. Rossi finished log

  8. A time dilation chamber as thought experiment on the power of pure thought, with comparison to computer AGI advantages/disadvantages.

    4676 words finished likely

  9. Does DNB increase IQ? What factors affect the studies? Probably not: gains are driven by studies with weakest methodology like apathetic control groups.

    12645 words in progress highly likely

  10. Iodine improves IQ in fetuses; adults as well? A meta-analysis of relevant studies says no.

    13356 words in progress highly likely

  11. Shrinking marginal returns to science and technology

    25564 words abandoned possible

  12. Human and animal sunk costs often aren't, and sunk cost bias may be useful on an individual level to encourage learning. Convincing examples of sunk cost bias typically operate on organizational levels and are probably driven by non-psychological causes like competition.

    24192 words finished likely

  13. Break down survival as Drake equation, see how plastination differs from cryonics, try to calculate advantage

    6243 words in progress likely

  14. History, background, visiting, ordering, using, & analyzing the drug market Silk Road 1

    43565 words finished likely

  15. On the benefits and lack of demerits of nicotine (research up to 2015)

    15100 words in progress likely

  16. Lithium is a well-known mood stabilizer & suicide preventive; some research suggests lithium may be a cognitively-protective nutrient and at the population level, chronic lithium consumption (through drinking water) predicts lower levels of mental illness, violence, & suicide.

    6067 words in progress unlikely

  17. What are the true limits to motivation?

    16450 words notes unlikely

  18. Efficient memorization using the spacing effect: literature review of widespread applicability, tips on use & what it's good for.

    26362 words finished highly likely

  19. Effects, health concerns, suppliers, prices & rational ordering.

    37961 words finished highly likely

  20. Internet links small groups, helping dissolve big groups; good, bad? But a bit sad.

    23406 words finished possible

  21. My prediction/betting strategies and track record, reflections on rationality, prediction judgments

    30475 words finished highly likely

  22. Melatonin improves sleep, & sleep is valuable

    13385 words finished likely

8 26 pages

  1. Training new LLMs on old LLM outputs is a good way to get a lot more data and recycle compute, and is not paradoxically self-defeating nor necessarily harmful.

    2027 words blog By Gwern finished certain

  2. Meta-mathematics discussion of why math is open-ended and generalizes, from the perspective of meta-reinforcement learning and contrasted with fields like speedrunning: it culturally selects for feasibly finding reusable findings about short general Turing machines.

    5978 words in progress possible

  3. Speculation about what sort of ordinary human writing is most relevant and useful to future AI systems.

    5354 words finished possible

  4. Dwarkesh Patel interviews Gwern Branwen about anonymity, AI scaling, AGI timelines, and the path from Wikipedia editing to Gwern.net. The interview becomes an intellectual autobiography: scaling as the triumph of compute over elegance, writing as LLM-era self-preservation, and rabbit holes as method. It also covers hearing impairment, Borges, site design, funding, GLP drugs, psychedelics, and open questions.

    26217 words By Gwern, Dwarkesh Patel finished certain

  5. Creative writing by OpenAI's GPT-3 model, demonstrating poetry, dialogue, puns, literary parodies, and storytelling. Plus advice on effective GPT-3 prompt programming & avoiding common errors.

    174723 words finished likely

  6. Criticizing studies and statistics is hard in part because so many criticisms are possible, rendering them meaningless. What makes a good criticism is the chance of being a 'difference which makes a difference' to our ultimate actions.

    8608 words finished highly likely

  7. Decision-theoretic analysis of how to optimally play the Haghani & Dewey 2016 300-round double-or-nothing coin-flipping game with an edge and ceiling better than using the Kelly Criterion. Computing and following an exact decision tree increases earnings by $6.6 over a modified KC.

    24091 words By Gwern, Arthur Breitman, nshepperd, FeepingCreature, Gurkenglas finished likely

  8. Bayesian decision-theoretic analysis of the effect of fancier packaging on subscription cancellations & optimal experiment design.

    9138 words finished likely

  9. Teaching a text-generating char-RNN to automatically imitate many different authors by labeling the input text by author; additional experiments include imitating Geocities and retraining GPT-2 on a large Project Gutenberg poetry corpus.

    14054 words finished likely

  10. Bayesian decision-theoretic analysis of local mail delivery times: modeling deliveries as survival analysis, model comparison, optimizing check times with a loss function, and optimal data collection.

    20190 words finished highly likely

  11. E-book edition of the 2002 Carter Scholz novel of post-Cold War science/technology, extensively annotated with references and related texts.

    229559 words By Carter Scholz, Gregory Benford, Hugh Gusterson, Sam Cohen, Curtis LeMay finished log

  12. Modern ethics excludes as many beings as it includes.

    23552 words finished likely

  13. Tricking the body into living longer?

    6240 words notes possible

  14. Simple cost-benefit: distributed computing considered harmful as scientific lemon projects run at high resource cost.

    7396 words in progress possible

  15. 2011 essay on how Bitcoin's long gestation and early opposition indicates it is an example of the 'Worse is Better' paradigm in which an ugly complex design with few attractive theoretical properties compared to purer competitors nevertheless successfully takes over a niche, survives, and becomes gradually refined.

    11819 words finished likely

  16. How do you encrypt a file such that it can be decrypted after a date, but not before? Use serial computations for proof-of-work using successive squaring, chained hashes, or witness encryption on blockchains.

    11292 words finished highly likely

  17. Archiving the Web, because nothing lasts forever: statistics, online archive services, extracting URLs automatically from browsers, and creating a daemon to regularly back up URLs to multiple sources.

    13734 words finished certain

  18. EEG recordings of sleep and my experiments with things affecting sleep quality or durations: melatonin, potassium, vitamin D etc.

    31385 words in progress likely

  19. 2013 discussion of how systemic biases in science, particularly medicine and psychology, have resulted in a research literature filled with false positives and exaggerated effects, called 'the Replication Crisis'.

    21758 words finished highly likely

  20. More effective ways to kill = terrorists are stupid, or killing not most important thing to them

    16109 words finished highly likely

  21. Terrorists act irrationally from a rational activism perspective, and groups act in ways most consistent with terrorism being about social status and belonging

    15924 words finished likely

  22. Iron Law of Bureaucracy: the downwards deletionism spiral discourages contribution and is how Wikipedia will die.

    31737 words finished highly likely

7 38 pages

  1. Research proposal for measuring how coding LLM perplexity scales with codebase context size, using Lean as a test case for whether formal languages have better predictability exponents and could lead to safer, more secure software worldwide.

    2778 words finished possible

  2. Proposal to use SAEs to crack open the dark matter of LLM inference about text.

    2001 words finished possible

  3. Proposal to teach LLM efficient interviewing by using psychometric tests to construct optimal question-and-answer sequences to train on. Would enable better personalization and imitation, and eventually, uploading.

    2525 words finished possible

  4. Discussion of possible tasks to measure LLM capabilities in soft 'creative' tasks like brainstorming or editing, to quantify failures in creative writing domains.

    4832 words finished possible

  5. Transcript of famous & widely-quoted 1986-03-07 lecture by Turing-Award mathematician Richard Hamming about how to do scientific research & development based on his life, antecedents of eminence, people he knew, the growing use of computers in science, navigating bureaucracy, maintaining creativity, and running Bell Labs.

    24556 words By Richard W. Hamming finished log

  6. Many fashions and artworks originate as copies of practical objects. Why? Because any form of optimized design is intrinsically esthetically-pleasing, and a great starting point.

    3995 words finished possible

  7. Nonfiction writing by OpenAI's GPT-3 model, testing logic, commonsense reasoning, anagrams, PDF/OCR cleaning, creative nonfiction, etc.

    58827 words finished likely

  8. Tryon’s Rat Experiment is a multi-decade selective breeding animal experiment begin in the 1930s which rapidly bred enormous differences in a complex psychological trait, maze-running, demonstrating core principles of behavior genetics.

    1406 words notes highly likely

  9. Experiments with OpenAI's 'preference learning' approach, which trains a NN to predict global quality of datapoints, and then uses reinforcement learning to optimize that directly, rather than proxies. I am unable to improve quality, perhaps due to too-few ratings.

    23294 words finished likely

  10. Demonstration tutorial of retraining OpenAI's GPT-2 (a text-generating Transformer neural network) on large poetry corpuses to generate high-quality English verse.

    70616 words By Gwern, Shawn Presser finished highly likely

  11. Links on traumatic brain injuries and mortality/morbidity from falling; do we greatly underestimate how many problems are caused by head injuries and how many QALYs are lost to falling?

    2414 words finished log

  12. Markets/evolution as backstops/ground truths for reinforcement learning/optimization: on some connections between Coase's theory of the firm/linear optimization/DRL/evolution/multicellular life/pain/Internet communities as multi-level optimization problems.

    23603 words finished possible

  13. Simple demonstrations of the AI control problem using MCTS in a Gridworld

    1606 words notes log

  14. Self-experiment on whether changes in bedroom CO2 levels affect sleep quality

    7053 words in progress likely

  15. Anthology of sociology, statistical, or psychological papers discussing the observation that all real-world variables have non-zero correlations and the implications for statistical theory such as 'null hypothesis testing'.

    32498 words finished highly likely

  16. Witticisms, parodies, pointed observations, japeries and/or jocularities, Tom Swifties, nominative determinism, and discursive drollery

    13004 words in progress possible

  17. Survival analysis of lifespans, deaths, and predictive factors of Tor-Bitcoin darknet markets

    11391 words in progress possible

  18. May 2013 overview of Silk Road 1's rise, powered by Tor & Bitcoin, enabling safe and easy online drug sales through the mail.

    10731 words By Mike Power finished log

  19. Analyzing predictors of Google abandoning products; predicting future shutdowns

    17193 words finished likely

  20. Does music impede studying and thinking?

    4630 words notes unlikely

  21. A census database of all publicly-reported arrests and prosecutions connected to the Tor-Bitcoin drug darknet markets 2011–2015, and analysis of mistakes.

    50035 words in progress highly likely

  22. Self-experiment on whether screen-tinting software such as Redshift/f.lux affect sleep times and sleep quality; Redshift lets me sleep earlier but doesn't improve sleep quality.

    3666 words finished highly likely

  23. Self-experiment on vitamin D effects on sleep: harmful taken at night, no or beneficial effects when taken in the morning.

    9162 words finished highly likely

  24. Does early interest in science fiction predicts future scientific careers or accomplishments?

    2002 words By Isaac Asimov finished log

  25. Applied Computer Science: On Murder Considered As STEM Field—using information theory to quantify the magnitude of Light Yagami’s mistakes in Death Note and considering fixes

    10798 words finished highly likely

  26. Notes on nootropics I tried, and my experiments

    64584 words in progress likely

  27. How I reinvented longevity insurance, which provides payouts if one lives to a certain point and thus might run out of savings.

    1566 words finished certain

  28. Thoughts on uses and a categorization of said uses for lifelogging

    4720 words notes possible

  29. Examples of low-hanging fruit suggesting the education system does not solely optimize for learning

    18706 words notes likely

  30. Aesthetically & economically, maybe there is too much new art. Don't take this too seriously.

    48420 words in progress unlikely

  31. A compendium of DNB, WM, IQ information up to 2015.

    78897 words in progress unlikely

  32. Network effects & benefits of gritting one's teeth & submitting to a Wikipedia's rules, rather than using Wikia or one's own site.

    1744 words finished highly likely

6 33 pages

  1. Interview with Adam Mastroianni (Experimental History) on writing and where ideas come from; done November 2025 at the Inkhaven writing residency.

    11442 words By Gwern, Adam Mastroianni finished certain

  2. Proposal & discussion of how default mode networks for LLMs are an example of missing capabilities for search and novelty in contemporary AI systems.

    2832 words finished possible

  3. The moral application of copyright and intellectual property to AI is the minimum protection necessary to promote the progress of science and the useful arts.

    2415 words in progress possible

  4. Proposal for a personal wiki built on neural nets: all edits are logged & used to finetune a NN assistant in realtime.

    8176 words in progress possible

  5. A side effect of preference-learning approaches like RLHF is a severe loss of 'diversity' or 'creativity' in outputs, analogous to broader mode collapse in generative models.

    501 words finished log

  6. Incompetence is the norm; most people who engage in a task (even when incentivized for performance or engaging in it for countless hours) may still be making basic errors which could be remedied with coaching or deliberate practice.

    1971 words note finished log

  7. Links on the circularity and mimetic desire basis of fashions in consumer goods, ideologies, politics, organizations like academia, religion etc.

    1117 words note finished log

  8. Examples of 'scenius' where highly-competitive tightly-communicating groups had surprisingly large effects and often finding low-hanging fruit (somehow) eluding generations before them.

    2367 words finished log

  9. Compilation of links demonstrating that it is common to find order-of-magnitude speedups in software, especially when designed carefully for (extraordinarily powerful) contemporary computer hardware with end-to-end principled thinking.

    2783 words note finished highly likely

  10. Generating Irish/folk/classical music in ABC format using GPT-2-117M, with good results.

    8618 words By Gwern, Shawn Presser finished likely

  11. A review of Russell 1986's Like Engend'ring Like: Heredity and Animal Breeding in Early Modern England, describing development of selective breeding and discussing models of the psychology and sociology of innovation.

    11730 words review finished likely

  12. An annotated fulltext bibliography of publications on the Study of Mathematically Precocious Youth (SMPY), a longitudinal study of high-IQ youth.

    61470 words in progress log

  13. Review of Roland & Shiman 2002 history of a decade of ARPA/DARPA involvement in AI and supercomputing, and the ARPA philosophy of technological acceleration; it yielded mixed results, perhaps due to ultimately insurmountable bottlenecks---the time was not yet ripe for many goals.

    6447 words review finished likely

  14. A common dream in philosophy and politics and religion is the idea of languages superior to evolved demotics, whether Latin or Lojban, which grant speakers greater insight into reality and rationality, analogous to well-known efficacy of mathematical sub-languages in solving problems. This dream fails because such languages gain power inherently from specialization.

    5168 words finished possible

  15. After putting up with slow glitchy WiFi Internet for years, I investigate improvements. Upgrading the router, switching to a high-gain antenna, and installing a buried Ethernet cable all offer increasing speeds.

    4664 words finished highly likely

  16. Analysis of whether bitter melon reduces blood glucose in one self-experiment and utility of further self-experimentation

    5696 words finished likely

  17. Pessimistic forecasters are overconfident in fixating, hedgehog-like, on only one scenario for how they think something must happen; in reality, there are always many ways through the garden of forking paths, and something needs only one path to happen.

    2943 words finished likely

  18. An LLLT user's blinded randomized self-experiment in 2013 on the effects of near-infrared light on a simple cognitive test battery: positive results

    3235 words finished unlikely

  19. Proposal to use meal-replacements to partition daily variance in mood/productivity between diet and other factors

    4958 words notes possible

  20. Defining the security/hacker mindset as extreme reductionism: ignoring the surface abstractions and limitations to treat a system as a source of parts to manipulate into a different system, with different (and usually unintended) capabilities.

    5334 words finished highly likely

  21. A catalogue of software constructs, languages, or APIs which are unexpectedly Turing-complete; implications for security and reliability.

    7133 words finished highly likely

  22. Compiling academic and media forecaster's 2012 American Presidential election predictions and statistically judging correctness; Nate Silver was not the best.

    5233 words finished highly likely

  23. Self-experiment with sub-psychedelic doses of LSD; no benefit

    19041 words finished likely

  24. Technological developments can be foreseen but the knowledge is largely useless because startups are inherently risky and require optimal timing. A more practical approach is to embrace uncertainty, taking a reinforcement learning perspective.

    17375 words finished likely

  25. One man's modus ponens is another man's modus tollens is a saying in Western philosophy encapsulating a common response to a logical proof which generalizes the reductio ad absurdum and consists of rejecting a premise based on an implied conclusion. I explain it in more detail, provide examples, and a Bayesian gloss.

    16613 words finished highly likely

  26. Any 'computer' is made up of hundreds of separate computers plugged together, any of which can be hacked. I list some of these parts.

    3009 words finished highly likely

  27. A retrospective of 8 years of SoC, and lessons learned

    15494 words finished highly likely

5 42 pages

  1. The Hutter Prize elegantly prevents cheating by charging for model size, but that same rule now excludes the scale regime where modern AI improves. A useful successor would evaluate incremental predictive information on future text under fixed inference compute, and let entrants inherit prior winners’ predictions.

    1563 words By Gwern finished possible

  2. American public survey suggests meaningful support for organ donation from anencephalic (terminal) babies.

    1400 words finished likely

  3. A poem for the end of the Anthropocene. 3-part Buddhist meditation on the ethics and nature of AI minds, largely written by LLMs.

    4553 words fiction By Gwern, Kimi K2, GPT-5 Pro, Claude-Sonnet-4.5, DeepSeek-V3.2 finished fiction

  4. James Reason’s Swiss cheese model illuminates how running red lights happens when multiple failures align—not from recklessness but from decision paralysis at the wrong moment. A ‘point of no return’ strategy pre-commits you to either stop or proceed based on a specific spot, creating a fail-safe that eliminates dangerous last-second judgment calls.

    2796 words finished highly likely

  5. Experiments in translating John Milton’s Paradise Lost to alliterative verse using LLMs, 2023--2026 and prototyping inner-monologue prompts for iterative composition.

    74201 words fiction By Gwern in progress fiction

  6. Prototype poems and abandoned draft completions of a William Empson poem, by various LLMs 2022–2025. Samples illustrate the inverted U-curve of LLM creative writing, where RLHF destroyed creativity for a few years. (The final version has been published as “The Fourth Truth Of Pain”.)

    10803 words fiction By Gwern, davinci-001, davinci-003, GPT-3.5, Kimi K2 Thinking, Gemini-3-pro-preview, GPT-5.1 Pro, Claude-4.5 finished fiction

  7. Cookbook of code for common selection & order statistics scenarios in economics, psychology, decision theory, drug development etc.

    2110 words notes highly likely

  8. Regression-to-a-mean is a general statistical phenomenon which leads to several widespread fallacies in analyzing & interpreting statistical results, such as ‘residual confounding’ and ‘Lord’s paradox’.

    2162 words finished certain

  9. Bibliography of ML papers related to multi-layer perceptrons (fully-connected neural nets), often showing surprising efficacy despite their reputation for being too general to be usable (representing a possible future Bitter Lesson).

    2327 words finished log

  10. Some interesting links on local optima/greediness/risk-aversion/creative destruction (eg. Porter hypothesis or equity premium puzzle), sometimes demonstrated by disasters.

    869 words note finished log

  11. Links on land valuation & taxation, single taxes, Georgism, and mechanism design.

    271 words finished log

  12. The Cultural Revolution was one of the greatest disasters in human history, the result of a self-reinforcing cycle of ideology failing to match reality and unsolved social problems, and the deranged reaction of zealots triggering defection and civil warfare.

    4198 words review finished likely

  13. Leprechaun hunting the origins of the famous skeptical observation that because millions of events are constantly happening, 'miracles' happen once a month; it was actually coined by Freeman Dyson.

    2789 words finished highly likely

  14. A tutorial explaining how to train and generate high-quality anime faces with StyleGAN 1+2 neural networks, and tips/scripts for effective StyleGAN use.

    42349 words finished highly likely

  15. Experiments in using BigGAN to generate anime faces and whole anime images; semi-successful.

    10019 words in progress highly likely

  16. Selection effects in media become increasingly strong as populations and media increase, meaning that rare datapoints driven by unusual processes such as the mentally ill or hoaxers are increasingly unreliable as evidence of anything at all and must be ignored. At scale, anything that can happen will happen a small but nonzero times.

    11282 words finished highly likely

  17. Extended book review of Bradshaw 2013 (Cat Sense) on the connections between cat psychology, evolution/genetics, history of domestication or lack thereof, & possible dysgenics, highlighting modern maladaptivity of cat psychology, with key references; speculation on cat toys, knocking things over, and tails.

    22577 words review finished possible

  18. A suggested x-risk/Great Filter is the possibility of advanced entertainment technology leading to wireheading/mass sterility/population collapse and extinction. As media consumption patterns are highly heritable, any such effect would trigger rapid human adaptation, implying extinction is almost impossible unless immediate collapse or exponentially accelerating addictiveness.

    29865 words in progress possible

  19. A classic pattern in technology economics, identified by Joel Spolsky, is layers of the stack attempting to become monopolies while turning other layers into perfectly-competitive markets which are commoditized, in order to harvest most of the consumer surplus; discussion and examples.

    10801 words finished highly likely

  20. Personal experience and surveys on running out of socks; discussion of socks as small example of human procrastination and irrationality, caused by lack of explicit deliberative thought where no natural triggers or habits exist.

    15347 words finished possible

  21. A high-concept Batman short story in the style of a 1980s comic book script (alluding to T. S. Eliot) about the Scarecrow and the gifts no one appreciates: pain/guilt/fear/anxiety. Fear death (including by water).

    3617 words fiction finished fiction

  22. Evaluating Long Bets as a prediction market shows it is dysfunctional and poorly-structured; despite the irrationality of many users, it is not good even as a way to raise money for charity.

    11467 words in progress likely

  23. 9 months of daily A/B-testing of Google AdSense banner ads on Gwern.net indicates banner ads decrease total traffic substantially, possibly due to spillover effects in reader engagement and resharing.

    24945 words in progress possible

  24. How much should we rewatch our favorite movies (media) vs keep trying new movies? Most spend most viewing time on new movies, which is unlikely to be good. I suggest an explicit Bayesian model of imprecise ratings + enjoyment recovering over time for Thompson sampling over movie watch choices.

    4838 words notes possible

  25. In generating a sample of n datapoints drawn from a normal/Gaussian distribution, how big on average the biggest datapoint is will depend on how large n is. I implement a variety of exact & approximate calculations from the literature in R to compare efficiency & accuracy.

    12743 words finished highly likely

  26. In discussing a large Scottish public health experiment, Student noted that it would've been vastly more efficient using a twin experiment design; I fill in the details with a power analysis.

    5370 words finished highly likely

  27. Estimation of catnip immunity rates by country with meta-analysis and surveys, and discussion of catnip alternatives.

    17449 words in progress likely

  28. Miscellaneous statistical stuff

    66678 words note in progress possible

  29. Email interview with McCaleb on early history of his MtGox, verifying it did not trade real but online Magic cards

    1192 words By Jed McCaleb finished log

  30. Does creatine increase cognitive performance? Maybe for vegetarians but probably not.

    5378 words finished unlikely

  31. A compilation of reviews of books I have read since ~1997.

    157543 words review in progress log

  32. Self-experiment on whether consuming caffeine immediately upon waking results in less time in bed & higher productivity. The results indicate a small and uncertain effect.

    6291 words finished possible

  33. How does LessWrong usage correlate with cryonics attitudes and signup rates?

    3624 words in progress possible

  34. Exercise in pointing out the obvious: Greenland has been, is, and will be indefinitely, a white elephant. Denmark turned down 100m USD from the USA in 1946; I discuss how this was a bad idea—America got what it needed anyway while Denmark kept control of a loser.

    5644 words finished possible

  35. Attempt to resolve Pascal's mugging by proposing & justifying a linear prior probability which bounds losses.

    5413 words draft possible

  36. Misc thoughts, memories, proto-essays, musings, etc.

    94394 words note in progress possible

4 57 pages

  1. How much to edit? A top-k attention-window toy model of the publish-polish trade off: You should polish in inverse proportion to how much of your writing a reader will see; the more diverse your writing, the better off you are writing more rather than better.

    1845 words finished likely

  2. A posthuman speaker finds the Port of Oakland’s container cranes still standing in the fog like the dinosaurs whose bones became the oil that forged their steel, and blesses them: stay, wade, become what you always were. 6-part mixed-verse by multiple LLMs, alluding to Gray and ‘The Ruin’ and ‘False Self’. Detailed Colophon discusses LLM writing process, prompts, and AI creativity.

    9169 words fiction By Gwern, Claude-4.6-opus, GPT-5.4 Pro, Kimi K2.5 Thinking finished fiction

  3. GPT-5 claims a new poem is weak but easily fixed. I challenge it to do so. It does, sort of, but also shows the flaws in autonomous LLM poetry right now—they need pressure, not polish.

    4389 words fiction By Gwern, GPT-5.4 Pro finished highly likely

  4. A poet stands before an audience of AIs spanning the scale of intelligence—from bits to Jupiter-brains—and answers ‘What is a poem?’, reciting a love poem for his dead girlfriend Alice. The poem argues that verse is a compression algorithm for human experience: the poet throws his voice across Lethe, and each reader decompresses Alice from signal, making the AIs creators. The ars poetica enacts its argument, moving from ballad quatrains through trochaic catalog into metrical dissolution, until a tercet closes the contract between poet/machine: ‘Behind each mask, an Alice runs.’

    14093 words fiction By Gwern, Gemini-3.1-pro-preview, Claude-4.6-opus, GPT-5.2 Pro, Kimi K2.5 Thinking, Fable finished fiction

  5. Gwtar is a new polyglot HTML archival format which provides a single, self-contained, HTML file which still can be efficiently lazy-loaded by a web browser. This is done by a header's JavaScript making HTTP range requests. It is used on Gwern.net to serve large HTML archives.

    5153 words finished certain

  6. Proposal: Speculation on style-aware image generators: using temporally trained embeddings and influence graphs to assign cultural credit, simulate counterfactual art histories, and generate styles rather than isolated images.

    2966 words finished possible

  7. A free-verse ode to aviation’s mundane miracle, using William Carlos Williams’ “so much depends” as structural refrain across six sections—from a torque-spec’d bolt to sleeping passengers to ground observers—culminating in a formally fractured final stanza that breaks the phrase across 17 lines, syntax itself suspended like the plane between blues. (Inspired by a June 2025 flight out of SFO, San Francisco.) Includes bonus Pindarian ode version.

    3028 words fiction By Gwern, Claude-Sonnet-4.5, GPT-5 Pro, Gemini-2.5-pro finished fiction

  8. Retrospective on Project Xanadu’s success and failure: a lack of design iteration, meaningful use-cases, or practicality stopped a valuable vision from maturing into something useful. (And contrasted with my approach.)

    5051 words finished possible

  9. On statistical and psychological grounds, we probably are not losing many future Einsteins to speedrunning or streaming or e-sports etc.

    2442 words finished likely

  10. Autobiographical documentary about 'underground artist' Robert Crumb on his 1991 farewell tour, discussing his personality, family, mental health, and source of his graphomanic art.

    6025 words review finished highly likely

  11. Documentary about suicides at the Golden Gate Bridge sheds little light on the brute facts of suffering and mental illness—facts which perhaps cannot be conveyed to others (on film or otherwise), and are inherently private.

    1633 words review finished possible

  12. Generative modeling proposal for increasing diversity of samples by a helper NN memorizing past samples and ‘repelling’ new samples away from old ones.

    2704 words finished possible

  13. Old neural net technique for runtime (online) learning which boosts performance.

    888 words finished log

  14. Review of evidence for Elon Musk being on the bipolar mood disorder spectrum.

    5733 words finished likely

  15. A collaborative AI completion of the second half of William Empson’s poem “This Last Pain” that transforms his locked-door metaphor into a camera obscura of Buddhist no-self doctrine and AI consciousness: the 18-line poem in Empsonian couplets argues that productive knowledge requires constraining distance. 4 AI models (Kimi K2, Gemini-3, Claude-4.5, GPT-5.1) then quarrel irreconcilably over whether the completion succeeds—their fractured disagreement itself demonstrating the poem’s thesis.

    6381 words fiction By Gwern, Kimi K2 Thinking, Gemini-3-pro-preview, GPT-5.1 Pro, Claude-4.5 finished fiction

  16. Diffusion models supposedly beat GANs because they scale better and stabler. That is unproven, and false. GANs should be revisited.

    2619 words By Gwern notes possible

  17. An exercise in spotting agenda journalism: dissecting a 2022 New York Times article on teen vaping sentence by sentence to show how it pins an EVALI hospitalization on legal nicotine vapes while never quite saying so—since the actual cause was illegal THC products. Every misleading sentence is technically true.

    6881 words finished certain

  18. Discussion of how to teach active reading and questioning of scientific research. Partially fake research papers may teach a critical attitude. Various ideas for games reviewed.

    4680 words finished log

  19. Sketch of a decentralized mineralization-based carbon capture: suppliers stake on reported deposits of mineral dust in publicly-auditable locations.

    6089 words finished possible

  20. Neural networks like GPT-3 power text adventure games where you can do anything; but they are too expensive. I propose that if we turn them into Choose Your Own Adventure hypertext games, they become feasible and enable new gameplay.

    5875 words finished possible

  21. Bibliography of links on the evolution of parasocial relationships online: more than social media acquaintances or influencer-consumption, less than true social relationships, typically with money involved somewhere; implications for the future of human relationships/media and artificial intelligence?

    2643 words in progress log

  22. Description of 3 anime datasets for machine learning based on Danbooru: cropped anime faces, whole-single-character crops, and hand crops (with hand detection model).

    5909 words By Gwern, Arfafax, Shawn Presser, Anonymous, Danbooru Community finished log

  23. Speculative world-building about a clonal humanity, based on identical twins and asexual reproduction.

    3288 words notes log

  24. Fermi estimates or problems are quick heuristic solutions to apparently insoluble quantitative problems rewarding clever use of real-world knowledge and critical thinking; bibliography of some examples.

    1256 words in progress possible

  25. Variance components analyses focus on estimating the net contribution of an entire group of variables to an outcome, without requiring estimating each variable; this is critical for learning if the haystack of variable contains a needle at all, and yet, this approach is hardly used outside behavioral genetics. That should change.

    1458 words in progress possible

  26. A description of advanced tips and tricks for effective Internet research of papers/books, with real-world examples.

    15022 words finished certain

  27. Catalogue of errors or problems caused by cats randomly doing things (ie. real-life fuzz testing).

    1817 words finished certain

  28. Account of well-intentioned but ill-fated attempt to conscript unintelligent or outright mentally-retarded men into fighting in the Vietnam War after remedial education, illustrating the difficulty of social interventions, the practical consequences of low intelligence, and the cruelty & evil of ignoring the reality of individual differences.

    4175 words review finished highly likely

  29. Proposal: the spacing curve could be used to review only flashcards one has forgotten, instead of almost forgotten. This inverse or opposite spaced repetition system would be useful for systematic efficient reviewing of indefinitely large sets of notes, citations, movies or other media to rewatch, etc., which could not feasibly be memorized, but which would generate serendipity or re-assesments if reviewed at some point in the distant future.

    2773 words in progress possible

  30. Someone (who is not me) should do something: a wishlist of miscellaneous research or project ideas—free to a good home.

    32685 words in progress unlikely

  31. A randomized blinded self-experiment of the effects of ZMA (zinc+magnesium+vitamin B6) on my sleep; results suggest small benefit to sleep quality but are underpowered and damaged by Zeo measurement error/data issues.

    3624 words finished possible

  32. A book study of surveys of the high-IQ elementary school HCES concludes that high IQ is not predictive of accomplishment; I point out that the disappointing results are consistent with the subjects not being geniuses due to regression to a mean (because of extremely early IQ tests) & small sample size.

    10910 words finished highly likely

  33. International population online surveys of cat owners about catnip and other cat stimulant use.

    13601 words in progress likely

  34. Many multi-step processes look like 'leaky pipelines', where a fractional loss/success happens at every step. Such multiplicative processes can often be modeled as a log-normal distribution (or power law), with counterintuitive implications like skewed output distributions and large final differences from small differences in per-step success rates.

    1337 words in progress highly likely

  35. Multilevel modeling of effect of small group's meditation on math errors

    3493 words in progress unlikely

  36. Does Google Alerts return fewer results each year? A statistical investigation

    3993 words finished likely

  37. 2 self-experiments on potassium citrate effects on sleep: harm to sleep when taken daily or in the morning

    3323 words finished highly likely

  38. On Bayesian updating of beliefs in sequentially searching a set of possibilities where failure is possible, such as waiting for a bus; the psychologically counterintuitive implication is that success on the next search increases even as the total probability of success decreases.

    11657 words By Ruma Falk, Abigail Lipson, Clifford Konold finished log

  39. A log of experiments done on the site design, intended to render pages more readable, focusing on the challenge of testing a static site, page width, fonts, plugins, and effects of advertising.

    34051 words in progress possible

  40. 1955 letters of John Nash and the NSA on a cryptosystem and Nash's belief that near-perfect cryptography could exploit exponential difficulties.

    4782 words By John Nash finished log

  41. AI folklore tells a story about a neural network trained to detect tanks which instead learned to detect time of day; investigating, this probably never happened.

    21885 words finished highly likely

  42. Many writers anecdotally report they write best first thing early in the morning, apparently even if they are not morning people. Do they, and why?

    111154 words in progress possible

  43. Scraping and downloading Haskell-related repositories from GitHub

    4989 words finished highly likely

  44. A precis of my old work on the English Wikipedia, 2004–2017 (edit counts, articles by subject, and particularly notable articles)

    1140 words finished log

  45. Individual human brains are more predictable and similar than they are different, reflecting low Kolmogorov complexity and implying that beta uploading may be more feasible than guessed, with suggestions on optimizing archived information.

    8918 words in progress likely

  46. How small must be the computer simulating the universe?

    3708 words finished unlikely

  47. Cynical tactics encouraged by Wikipedia's abdication of thought known as 'No Original Research' and 'Reliable Sources'

    4250 words finished highly likely

  48. Back of envelope financial calculations: Warranties, fine; insurance, no!

    3682 words finished highly likely

  49. Game theory on BSD versus GPL: partnership? Is GPL an evolutionary stable strategy against invasion by proprietary copyright strategies?

    2470 words in progress possible

  50. Criteria making software useful long-term & worth learning in detail

    2338 words in progress possible

  51. Lessons learned about bitrot in Haskell software

    1789 words in progress likely

3 74 pages

  1. Deep reinforcement learning explained in small words in Grow-Speech, a controlled-English dialect that defines each technical term before use and reduces the field to ‘try, score, and change’.

    8879 words By GPT-5.5 Pro, Claude-4.7-opus, Gwern finished possible

  2. A Calvinoesque philosophical miniature about a Silk Road city whose citizens cast horoscopes at birth, seal them until death, and elevate compatibilism into civic rite.

    11237 words fiction By Gwern, Claude-4.7-opus, GPT-5.4 Pro finished fiction

  3. Discussion of AI-invented poetic meter for writing Pindaric odes in English. Testing on a Pindar translation, it may be a viable meter.

    8656 words fiction By Gwern, GPT-5.4 Pro, GPT-5.2 Pro finished likely

  4. 3 ‘Coastal Spondaic’ poems on walking: over sand, rain, and wood; under tide, forest, and rot.

    2434 words fiction By Gemini-3.1-pro-preview, GPT-5.2 Pro, GPT-5.5 Pro, Claude-4.8-opus, Claude-5-Fable, Kimi K2.6 Thinking, Gwern finished fiction

  5. Proposal: train LLMs to predict Stan code from synthetic data sampled from Stan’s prior predictives. The output is an interpretable, executable Bayesian model, and a symbolic posterior over model structure. Possibly useful for faster inference, search, and interpretability.

    3808 words finished highly unlikely

  6. Pindaric ode commemorating millions of laboratory animals sacrificed for science, in a strict alliterative meter. Synthesizing the history of mice strains with mythic figures like Laika/OncoMouse, the poem explores the covenant between researcher and subject. From the granaries of Abbie Lathrop to the tragedy of euthanasia or release, it seeks to elevate these unwilling martyrs to a place of memory. By transforming the cold ‘ledger’ of data into ‘gold’ of song, the ode attempts to pay the moral debt that statistics alone cannot settle.

    2728 words fiction By Gwern, GPT-5.2 Pro, Claude-4.5-opus, Kimi K2 Thinking, Gemini-3-pro-preview finished fiction

  7. Video games are art, but a strange art: their essence is transformation of the player, not description to the player. This makes meaningful criticism nearly impossible—you can point at the moon, but it’s not the moon, and once someone sees it, they no longer need the pointing.

    3483 words in progress possible

  8. Proposal for a conference organization tool modeled after ad hoc collaboration tools like tabletop roleplaying games or disaster response: just have everyone scribble on a map of the conference location.

    2180 words blog By Gwern finished unlikely

  9. List-processing tricks for generating embeddings at different levels of document abstraction to allow efficient semantic searching.

    1783 words finished possible

  10. Copyright mechanism proposal to solve the orphan works problem: self-assessed Harberger taxes on any inherited copyright are then invested, and dedicated to eventually buying out the owners. Works are either immediately public-domained, or the owners voluntarily 'sell' them if their value underperforms a baseline investment.

    4881 words finished remote

  11. Why are popularizing educational newsletter-frequency writers of important fields like Matt Levine for finance so rare? Because most fields are too slow or ambiguous, and writers of the right combination of expertise, obsession, and persistence are also rare.

    3076 words in progress possible

  12. A discussion of why I don’t intend to turn Gwern.net into a book, and how trying to write a book can harm writers.

    4415 words finished possible

  13. Proposal to create a 'search engine' like OEIS but for individual numbers, allowing fuzzy lookups, by training a neural net embedding on the scientific & mathematic literature's corpus of co-occurring numbers, and by known special transformations.

    1726 words finished likely

  14. Speculation on why 3-move games like Rock-Paper-Scissors are so common in games & decisions: aside from allowing ties, is this design information-theoretically optimal in game difficulty vs information per game?

    1529 words finished possible

  15. Children's animated fantasy movie with surprisingly deep thoughts about personal identity and immortality.

    1691 words review finished highly likely

  16. Cats are not humans, but we design things like they are, for our convenience. What are the design patterns for cat-architecture? One missing design pattern: progressive concealment, for cat ledges, flaps & window boxes.

    4170 words finished possible

  17. Answer: Too many! After observing 14 unique shower controls across 20 stays while traveling, I used capture-recapture statistics to estimate the total population of designs, finding a likely mean of >34 types but a wide credible interval because I recorded the wrong kind of recapture data. Ultimately, the exercise served as a demonstration of R. A. Fisher’s warning that statistical analysis is too often a ‘post mortem’ that cannot rescue a flawed experimental design.

    3023 words By Gwern, GPT-5.2 Pro finished highly likely

  18. Satirical essay on how AI can never truly solve a Rubik’s Cube like human beings can, written by an AI. Inspired by ‘Supersized Machines’ (Garfinkel et al 2017).

    8791 words By Gwern, Claude-3, Claude-2 finished highly likely

  19. Proposal for applying the AUNN neural net architecture to reconstruction of brains in a modular piece-wise fashion.

    2023 words finished unlikely

  20. Proposal for applying the AUNN neural net architecture to reconstruction of historical documents using pretrained large language models.

    2621 words finished unlikely

  21. Explanation of the emotional core of The Quantum Thief’s exploration of the trap of persistent personal identity and seeking freedom.

    4217 words review finished certain

  22. Why is Yamato so influential on a generation of anime otaku and later anime, despite a weak story and compromised production? It offers a fantasy WWII in which Imperial Japan is the hero, and all its sins projected onto an imaginary enemy.

    1564 words review in progress possible

  23. The activity of manufactured collecting is puzzling; what explains the enormous resources spent on what is, by and large, neither lucrative, prestigious, nor entertaining? Is it just a misfire and quirk of human psychology poorly adapted to industrial society where 'collectibles' can be varied and marketed 24/7/365?

    5850 words in progress possible

  24. Chemicals active at the nanogram scale are invisible, difficult to observe, and near-impossible to clean or remove; but they are still psychoactive enough to torment you like demons.

    3645 words finished likely

  25. Flamewars over platforms & upgrades are so bitter not because people are jerks but because the choice will influence entire ecosystems, benefiting one platform through network effects & avoiding 'bitrot' while subtly sabotaging the rest through 'bitcreep'.

    6594 words finished likely

  26. Discussion of whether there is a general ‘unsorting’ list operation to avoid redundancy, repetition, or lack of novelty. Probably not, there are too many things you might want to maximize or minimize.

    2528 words finished likely

  27. Real-world examples of using advanced tips and tricks for effective Internet research of papers/books.

    12656 words finished certain

  28. Did John Calhoun's 1960s Mouse Utopia really show that animal (and human) populations will expand to arbitrary densities, creating socially-driven pathology and collapse? Reasons for doubt.

    5197 words finished likely

  29. I describe how I made the website ThisWaifuDoesNotExist.net (TWDNE) for displaying random anime faces generated by StyleGAN neural networks, and how it went viral.

    11142 words finished highly likely

  30. A compilation of opera reviews since 2019.

    16336 words review in progress log

  31. The multiple-invention (>3) history of the idea of extremely-powerful embryo selection by using gametogenesis to run many 'generations' in vitro.

    4581 words finished highly likely

  32. Acne is a good way to learn self-experiments for teenagers. I offer tips on what one could do, and an initial list of crowdsourced interventions to test from CureTogether.

    3722 words finished highly likely

  33. A meditation on subcultures & review of the cartoon series My Little Pony: Friendship Is Magic, focusing on fandom, plot, development, and meaning of bronydom.

    18036 words review finished highly likely

  34. Some anomalies/questions which are not necessarily important, but do puzzle me or where I find existing explanations to be unsatisfying.

    13445 words finished possible

  35. Decision analysis of whether cloning the most elite Special Forces dogs is a profitable improvement over standard selection procedures. Unless training is extremely cheap or heritability is extremely low, dog cloning is hypothetically profitable.

    29822 words finished possible

  36. Discussion of fictional eugenics program in the SF Dune-verse and how it contradicts contemporary known human genetics but suggests heavy agricultural science and Mendelian inspiration to Frank Herbert's worldview.

    10391 words finished highly likely

  37. Review of Isao Takahata’s last & greatest film on how to accept the transience of life, filled as it is with both joy & sorrow.

    1818 words review finished likely

  38. Commandline tool providing interactive statistical pairwise ranking and sorting of items

    7527 words finished likely

  39. Some claim the disease hydrocephalus reduces brain size by 95% but often with normal or even above-average intelligence, and thus brains aren't really necessary. Neither is true.

    6663 words finished highly likely

  40. Analysis of 2013–2014 LessWrong survey results on how much more self-identified EAers donate suggests low median donation rates due to current youth and low incomes.

    2326 words finished highly likely

  41. Using modern statistical analysis, how quickly should improved bladder stone surgical methods have been adopted?

    1891 words abandoned possible

  42. List of academic urban legends, and references on how often writers do not check references, which seems to be a major reason for their propagation.

    7056 words finished highly likely

  43. A small 2014–2015 non-blinded self-experiment using Bacopa monnieri to investigate effect on memory/sleep/self-ratings in an ABABA design; no particular effects were found.

    4067 words finished possible

  44. A compilation of movie, television, and opera reviews since 2014.

    35337 words review in progress log

  45. Is sleep affected by the phase of the moon? An analysis of several years of 4 Zeo users' sleep data shows no lunar cycle.

    2693 words finished highly likely

  46. Discussion of Cordwainer Smith SF story, arguing that the pain-of-space is based on forgotten psychological issues in air travel, and concerns about worse ones in space travel, which were partially vindicated by the existence of interesting psychological changes in astronauts.

    10517 words finished possible

  47. 3 magnesium self-experiments on magnesium l-threonate and magnesium citrate.

    27198 words in progress possible

  48. A small fraction of human responses will always be garbage because we are lazy, bored, trolling, or crazy.

    1705 words finished possible

  49. Notes relating to my use of a treadmill desk and 2 self-experiments showing walking treadmill use interferes with typing and memory performance.

    9446 words finished likely

  50. On divining the esoteric truth of Neo-Venezia through holes in world-building.

    3270 words in progress highly unlikely

  51. Meta page describing Gwern.net editing activity, traffic statistics, and referrer details, primarily sourced from Google Analytics (2011-present).

    14881 words in progress certain

  52. A compilation of anime/manga reviews since 2010.

    41165 words review in progress log

  53. Meta page describing Gwern.net site ideals of stable long-term essays which improve over time; idea sources and writing methodology; metadata definitions; site statistics; copyright license.

    13657 words finished highly likely

  54. Meta page describing Gwern.net, the self-documenting website’s implementation and experiments for better ‘semantic zoom’ of hypertext; technical decisions using Markdown and static hosting.

    19696 words finished highly likely

  55. Generalizing functions with hardwired values

    2096 words in progress log

  56. Extensive anthology of Gainax/Anno/Evangelion quotes, excerpts, sources, references, and analyses, organized by reliability and year.

    203679 words notes log

  57. Who am I online & what have I done? Contact information; sites I use; computers and software tools; things I've worked on; psychological profiles

    8191 words finished highly likely

  58. Haskell juvenilia, reinventing entry-level math wheels, of no interest to anyone; investigating when two theories or representations mean the same things.

    2899 words finished remote

  59. recursion into iteration; primality testing

    3898 words finished log

  60. Links to various resources for SICP studying

    752 words finished log

  61. Remaining materials towards one day building and trying out a real fog gun

    2005 words notes unlikely

  62. Why Wikipedia and YouTube will never be integrated.

    1994 words finished certain

  63. Haskell: tutorial on writing a daemon to archive links in Wikipedia articles with TagSoup and WebCite; obsolete.

    7349 words finished likely

2 53 pages

  1. Bureaucratic-horror fiction: the molting chapter of a statutory owner's manual for companion mimes, with glossary excerpts and a completed Grade III incident-notification form.

    5247 words fiction By Claude-4.8-opus, GPT-5.5 Pro, Claude-5-Fable, Gwern draft fiction

  2. Web design post-mortem of ways to typeset poetry and how we developed our advanced approach. Possible approaches are PDF plugins, PNG screenshots, semantic class soup, and Pandoc line blocks. Gwern.net has a two-tier system: div.poem for stanzaic verse with automatic enjambment, pre.poem-html implementing 'Source Code WYSIWYG' with Nimbus Mono L font and JS inline-width compensation. Covers slash/caesura syntax highlighting, responsive line-breaking, and LLM corpus cleanup.

    9298 words finished highly likely

  3. Trying out having Gemini-3-pro-preview write system prompts compatible with creative fiction writing (sci-fi short story examples), and trying it out on GPT-5 Pro and Claude-4.5-opus with mixed results.

    6146 words By Gwern, Gemini-3-pro-preview, GPT-5 Pro, Claude-4.5-opus finished possible

  4. Speculation on feline-human fascination being ancestral vigilance triggered by their behavioral similarity to major primate predators, evolutionarily, creating a compelling ‘safe danger’ like watching a captivating villain. We still remember when they ate us.

    5192 words finished unlikely

  5. Can 2024-era LLMs reason successfully about how to write using only words that could not have appeared in any Bible ever? Mostly.

    3721 words finished certain

  6. There are many luxury goods. Are there luxury websites? They could exist, but mostly don't. Maybe website quality is too hard to understand right now.

    3324 words in progress unlikely

  7. I manually extract RGB colors for 225 websites and graph them: red > blue > green > purple > yellow–brown.

    2399 words finished highly likely

  8. The 2007 indie SF film Timecrimes shows the horror & metaphysical implications of stable time loops: over a single day, the protagonist commits escalating crimes while trapped in a 3-iteration loop enabled by a time machine-or rather, caused by it.

    7678 words review finished possible

  9. We develop AI image generation workflows for webpage dropcap typography, creating PNGs & SVGs using image generators. As demos, we create new Gwern.net logos and several custom FLOSS dropcap sets, including felines, Gene​ Wolfe horror fiction, and neural-net-inspired dropcaps.

    15951 words finished log

  10. Harold Bloom amusingly recommends all students as his ‘best student ever’; modeled as a ‘record value’ order statistics problem, this is more common than it seems.

    1184 words finished certain

  11. Some July 2023 Claude-2 & ChatGPT satirical writing samples of gonzo journalism & typography.

    4090 words By Claude-2, GPT-4o, Gwern finished fiction

  12. Why do cats like to push stuff over edges and then curiously watch the fallen object? I suggest that they are play-hunting, and are testing the ‘prey’ for liveness & playing-dead, similarly to tossing or poking it with claws.

    1714 words finished possible

  13. Proposal for incentivizing meta-learning of exploration in deep reinforcement learning: domain randomization with reward-shaping, where there is a fixed-length ‘play time’ with no rewards/losses at the beginning of each episode.

    2423 words finished possible

  14. Brief description of how the direct-to-video-rental model helped create the anime boom by allowing wide diversity in bankrolled anime projects, outside the production committee system.

    2369 words finished possible

  15. The Twitter UI/UX for accepting follow requests is a fractal of bad design.

    1853 words finished certain

  16. 1-year-long website A/B test of controversial typographic choices of indenting & justifying paragraphs: a fairly precise null result. No design change was made.

    4031 words finished possible

  17. Description of a useful service for web development: a website which wraps a neural network trained to classify images by whether they would look better inverted in a website/app dark-mode, or faded.

    2684 words obsolete highly likely

  18. Why do movies or games go through development hell, and usually come out worse, rather than better, and long gestation times do not seem to improve them despite the large resources poured into them by the most talented people?

    3057 words finished unlikely

  19. In typography/design, 'sidenotes' place footnotes/endnotes in the margins for easier reading. I discuss design choices, HTML implementations and their pros/cons.

    9660 words finished certain

  20. A typographic proposal: replace cumbersome inline citation formats like ‘Foo et al. (2010)’ with subscripted dates/sources like ‘Fooet al2020’. Intuitive, easily implemented, consistent, compact, and can be used for evidentials in general.

    8285 words finished certain

  21. Which books on GoodReads are most difficult to finish? Estimating proportions in December 2019 gives an entirely different result than absolute counts.

    24536 words finished likely

  22. A gallery of typographic and graphics design examples of rubrication, a classic pattern of using red versus black for emphasis.

    10754 words finished certain

  23. Post-mortems of failed neural network experiments in generating anime images, pre-StyleGAN/BigGAN.

    2255 words finished highly likely

  24. Review of famous light novel series Umineko no Naku Koro ni: it is a highly idiosyncratic, wildly self-indulgent, yet impressive exploration of all possible locked-room mysteries, knocked for a loop mid-composition by personal tragedy, triggering a descent into a deeply harmful endorsement of fantasies & running from painful realities.

    7187 words review finished highly likely

  25. Proposals for new technologies, businesses, startups, satirical or serious.

    10111 words in progress unlikely

  26. Bakker's Second Apocalypse vs Frank Herbert's Dune: time loops and finding freedom in an unfree universe. In Dune, humanity is liberated by growth and development and escaping the predeterminism of prescience; in Bakker, they are destroyed by it, and liberation is achieved only by death and reunification with a deeper underlying block-universe/monistic reality.

    7203 words review finished possible

  27. Blind randomized taste-test of mineral/distilled/tap waters using Bayesian best-arm finding; no large differences in preference.

    5423 words finished highly likely

  28. Physical beauty & attractiveness of the general population of men/women seems to have increased greatly in the past few centuries, judging by surviving art/photos, contemporary judgments, and objective criteria like missing teeth, likely due to economic/technological/medical/nutritional improvements, but less from cosmetic tricks. Beauty may, however, be in decline very recently as some of those trends reverse (eg. now too much food, not too little).

    6355 words finished possible

  29. Isaac Newton's cosmology apparently involved regular apocalypses caused by comets overstoking the furnace of the Sun and the repopulation of the Solar System by new intelligent species. He supports this speculation with an interestingly-incorrect anthropic argument.

    6029 words finished highly likely

  30. I compare my electric tea kettle to my stove kettle, and apply some simple statistical modeling verifying the electric is faster.

    4043 words finished highly likely

  31. Richard Feynman recounts an amazing anecdote about philosophy of science; we trace it to a forgotten University of Michigan research programme into ‘floor cues’.

    9275 words finished highly likely

  32. Commandline folklore: sorting files by filename or content before compression can save large amounts of space by exposing redundancy to the compressor. Examples and comparisons of different sorts.

    4166 words finished certain

  33. Downloading and parsing TPB metadata to estimate Bitcoin usage and revenue for uploaders

    1733 words notes log

  34. In which the author receives surprising offers from kind strangers

    19077 words finished log

  35. Short essays on the influence on Japanese culture of anime works Daicon, Evangelion, Space Battleship Yamato, and the artists Mr., Chiho Aoshima, and others.

    2745 words By Takashi Murakami finished log

  36. Essay by Pop Art artist Takashi Murakami on Japanese society and on WWII infantilizing Japanese culture as revealed by media, anime, and otaku.

    17808 words By Takashi Murakami, Linda Hoaglund finished log

  37. Objectively, anime/manga better than American alternatives. So why the need to justify?

    7171 words in progress unlikely

  38. Fulltext annotated e-book of 2002 memoir by anime producer Yasuhiro Takeda, discussing Japanese SF conventions & fandom, formation & history of Gainax and its productions up to 2002, including the origins of Evangelion & the tax raid.

    79728 words By Yasuhiro Takeda finished log

  39. Meta page describing Gwern.net website design experiments and post-mortem analyses.

    32647 words finished highly likely

  40. Descartes's God is pantheism; a reductio ad absurdum of his ontology

    2126 words finished unlikely

  41. Dialogue pointing out some difficulties of materialist objections to abortion.

    4028 words finished possible

  42. Ambiguous ending of Death Note means even the victor is unclear; who was right?

    4579 words finished possible

  43. Copyright considered paradoxical, incoherent, and harmful from an information theory and compression perspective as there is no natural kind corresponding to 'works', merely longer or shorter strings for stupider or smarter algorithms.

    4203 words finished possible

  44. Stoic meditation with reference to being homeless. Written to myself at a particularly low point; like many, I take comfort in considering how things could be worse.

    557 words finished emotional

  45. China & Egypt used wooden pillows; my recreations fail

    1746 words finished log

1 43 pages

  1. Go-Speech is a controlled-English writing format that improves Grow-Speech's left-to-right vocabulary growth with an Up-Goer-Five-style Basic-English Databank. Intended for LLM prompting, audit, and child-facing humorous explanation.

    6357 words By GPT-5.5 Pro, Claude-4.7-opus, Gwern draft log

  2. Formal specification of Guy L. Steele’s 1998 controlled-vocabulary English writing system, which requires all multi-syllabic core words to be explicitly defined before use. Intended for LLM use.

    5454 words By Gwern, GPT-5.5 Pro, Claude-4.7-opus finished log

  3. Pulp SF novel about mind-controlling evil UFO alien abductions seems more about BDSM and furries than appreciated, possibly explaining its early popularity and eventual disappearance.

    6851 words review By Gwern finished possible

  4. Ideas for a costume party game revolving around guessing identities, with an asymmetrical twist.

    1354 words finished possible

  5. Western TV animation is often shockingly ugly and crude compared to other animation like French or Japanese. Why, when it’s not even cheaper? Degenerate echoes of what were once cutting-edge art.

    1357 words finished possible

  6. An esoteric document proposal: abuse Unicode to create the fanciest possible ‘plain text’ documents.

    7934 words finished unlikely

  7. Is it possible to prove the existence of retrocognition if precognition also exists, because precognition could be used to foresee any proof found of retrocognition? Does this also disprove precognition as well?

    2312 words finished possible

  8. 49 words finished log

  9. Timothy Falcon's quantitative-finance interview problem #14 asks for the optimal stopping strategy when playing a card-drawing game of  l cards where red = +$1 & black = −$1; the value approaches 0.5 × √l. I re-solve it with dynamic programming in R, and others in Neat, Haskell & C, with increasing efficiency.

    9786 words By Gwern, FeepingCreature, nshepperd, Khoth finished certain

  10. Microwaving water for tea makes it taste bad, for unknown reasons. Is it a lack of bubbles? Lingering CO2? Comments and proposed experiments.

    1549 words finished possible

  11. Alice Sheldon's SF short story 'The Screwfly Solution' is misread by most readers as an alien attack driving men to insane violence; a closer revisionist read shows that the alien attack affected women as well, into suicidal passivity.

    2517 words finished possible

  12. Looking into the oldest edible food suggests a range of 0.05--20mya food, after discarding myths like honey or woolly mammoth.

    2395 words finished possible

  13. Fonts are durable, highly-reusable, compact, & high-quality software products which do not 'bitrot'. Nevertheless, hundreds or thousands of new ones come out every year despite enormous duplication; why? I speculate that designer boredom seems to be the answer: they crave novelty.

    5567 words finished possible

  14. Collation of anecdotes and speculation about why many cats like earwax, and human earwax especially. Is it the valeric acid? Plus human sweat.

    4959 words in progress highly likely

  15. I explain the somewhat arbitrary-seeming death of the popular Star Wars character Grand Admiral Thrawn as being the logical culmination of a tragic character arc in which his twisted (or ‘thrawn’) cunning & scheming, one of his defining traits, ultimately backfires on him, causing his bodyguard to betray & assassinate him.

    1528 words finished certain

  16. I wondered how fast deer could evolve to avoid car fatalities. Truncation selection with known mortality and generation length suggests that it would take a long time even with high heritability.

    3152 words finished likely

  17. In MLP:FiM, the 3 pony races sometimes bear offspring of other pony races; I review 4 complicated Mendelian models attempting to explain this, and note that a standard polygenic liability-threshold model can fit it parsimoniously.

    2180 words finished highly likely

  18. Short story on the limits of propaganda and 'Newspeak' using a constructed language; from Chapter 11 of Gene Wolfe's The Book of the New Sun, volume 4, The Citadel of the Autarch.

    6081 words By Gene Wolfe finished log

  19. Monthly chronological list of recent major writings/changes/additions to Gwern.net (see also the monthly newsletter).

    5917 words in progress log

  20. Collect music metadata and look for patterns

    6413 words abandoned possible

  21. Kunihiko Ikuhara and staff episode and music commentary/discussion 1997–2011 on the anime Revolutionary Girl Utena; discusses origins of ideas and meaning of themes, and video/audio remastering for the DVD box set.

    16276 words By Kunihiko Ikuhara, Yuichirou Oguro, Hiroshi Kaneda, Haruyasu Yamazaki, Tomomi Takemura, Hideki Ito, Yo Yamada, Tomokazu Mii, Yoji Enokido, Shinya Hasegawa, J. A. Caesar, Toshimichi Otsuki, Chiho Saito, Sarah Alys Lindholm, C.A.P finished log

  22. Famous Ted Chiang SF short story ‘Story Of Your Life’ is usually misinterpreted as, like the movie version Arrival, being about time-travel/precognition; I explain it is instead an exploration of xenopsychology and a psychology of timeless physics.

    17050 words finished likely

  23. English translation of French translation of Hideaki Anno's controversial NewType interview at the end of the TV broadcast

    3494 words By Hideaki Anno, Shinichiro Inoue finished log

  24. Definition of otaku, mania, moe, dame, anime, and generations

    7508 words By Toshio Okada, Kaichiro Morikawa, Takashi Murakami, Reiko Tomii finished log

  25. Do mindless positive/negative comments skew article quality ratings up and down?

    3825 words finished unlikely

  26. Implementing the Hypertime multiverse model of time travel in Haskell

    2569 words abandoned likely

  27. Working at Gainax, Evangelion's production & censorship, anime translation

    7421 words By Michael House finished log

  28. Translated interviews about making Evangelion 2.0 w/Anno, Higuchi, Enokido, & Tsurumaki

    37479 words By Ryusuke Hikawa, Hideaki Anno, Shinji Higuchi, Yōji Enokido, Kazuya Tsurumaki, Mohiro Kitoh, Shigeto Koyama, Yoshito Asari in progress log

  29. Teas I have drunk, with reviews and future purchases; focused primarily on oolongs and greens. Plus experiments on water.

    51641 words review in progress log

  30. Race as reflected in gender ratios within fictional bi-racial marriages in anime/manga show equal sex ratios and Western European overrepresentation with striking absence of Korean characters.

    34227 words in progress likely

  31. a comprehensive taxonomy of predictions on future Evangelion media

    5523 words in progress log

  32. Essay on writing waka on truth & lies

    1800 words fiction finished fiction

  33. Observations on Frank Herbert's Dune series—the notes are probably overstated, the Butlerian Jihad was not a robot uprising seeking to exterminate humanity, and precognition/prescience in the Dune universe appears to work backwards allowing for retro-causality and stable time-loops as exemplified by Leto II's Golden Path creating Paul Atreides and the events of Dune.

    9926 words notes possible

  34. Evangelion thoughts---anti-religion, pro-psychology, character development, otaku genre.

    19390 words notes likely

  35. Argument that texts are neither moral nor immoral as they require active interpretation.

    898 words finished unlikely

  36. Re-inventing anarcho-capitalism in a French context. (Obviously wrong juvenilia kept for historical reasons.)

    3566 words notes remote

  37. Internal, external, stylometric evidence point to live-action leak of Death Note Hollywood script being real.

    16206 words finished likely

  38. Literary analysis of the light novel/anime series The Melancholy of Haruhi Suzumiya: Haruhi is not God, Kyon is

    23560 words in progress highly unlikely

  39. On the interpretation of Gene Wolfe's short story ‘Suzanne Delage’ as inversion of Bram Stoker’s Dracula horror novel.

    27602 words finished highly likely

  40. An attempt to reinvent classic theories of social interaction as expressions of power

    1836 words finished unlikely

0 16 pages

  1. A 9-part lyric narrative poem in mostly loose 13-syllable ABAB quatrains, punctuated by terse italic weekday refrains that read as a chore-and-ritual checklist. Beginning with a father patrolling “ghosts” and the speaker’s childhood nightmares, it follows how an adult partner’s hands become a private pre-verbal grammar before focusing on everyday life then parenthood, where the same gesture is inherited. The closing image—daughter’s hand into yours into mine—turn vigilance into steadiness and haunted memory into hard-won tenderness.

    17488 words fiction By Gwern, Claude-4.5-opus, GPT-5.1 Pro, Kimi K2 Thinking, Gemini-3-pro-preview finished fiction

  2. Short story about ‘The Institute of September the Thirtieth’, dedicated to the study of a single day in 1​939, in the form of a review of a fictional book.

    15317 words fiction finished fiction

  3. Variations on the idea of criminals being convicted and sentenced to a ‘second’ life sentence—what could that possibly mean...?

    4489 words finished fiction

  4. Curious what the first 'unused' alphabetic acronym is, I have GPT-4 write a script to check English Wikipedia. After three bugs, the first unused one turns out as of 2023-09-29 to be the three-letter acronym 'CQK', with another 2.6k TLA unused, and 393k four-letter acronyms unused. Exploratory analysis suggests alphabetical order effects as well as letter-frequency.

    14184 words finished highly likely

  5. Bibliography of papers on math & physics methods of hunting lions in the Sahara Desert.

    469 words finished log

  6. I scrape Arxiv to find underused Greek variables which can add some diversity to math; the top 10 underused letters are ϰ, ς, υ, ϖ, Υ, Ξ, ι, ϱ, ϑ, & Π. Avoid overused letters like λ, and spice up your next paper with some memorable variables!

    7320 words finished highly likely

  7. Inline HTML snippets for the videos of the 2 famous Gainax fan-made convention anime, for easier transclusion (and as a test-case).

    82 words By Gainax finished log

  8. 1963 Slovenian poem on the morality of raising rabbits for meat.

    301 words By Miroslav Válek finished log

  9. Ratings of mead and fruit wines I have tried.

    5364 words review finished log

  10. Short history of trends in Japanese erotic anime and manga, especially lolicon

    1164 words By Stephen Sarrazin finished log

  11. Perspective and interpretation

    6505 words draft likely

  12. The unexpected critical depths of a single throw-away scene

    1196 words finished unlikely

  13. Philosophical fiction or a prose poem about suffering innocents and the theodicy

    523 words notes emotional