June 2024 News
June 2024 Gwern.net newsletter with links on TODO
June 2024’s Gwern.net newsletter is now out; previous, May 2024 (archives). This is a collation of links and summary of major changes, overlapping with my Changelog; brought to you by my donors on Patreon.
Writings
Links
AI
-
“Representation Engineering: A Top-Down Approach to AI Transparency”, et al 2023
-
“Champion-level drone racing using deep reinforcement learning”, et al 2023
-
“Creativity Has Left the Chat: The Price of Debiasing Language Models”, 2024 (documenting the severe mode collapse & loss of diversity in possible completions in RLHF-tuned LLMs—how much does this impede any attempt at search or planning using LLMs…?); “Consistency-diversity-realism Pareto fronts of conditional image generative models”, et al 2024; “Epistemic calibration and searching the space of truth”, Linus Lee (the boringness of Midjourney/DALL·E 3 vs DALL·E 2)
-
“In-context Reinforcement Learning with Algorithm Distillation”, et al 2022; “Supervised Pretraining Can Learn In-Context Reinforcement Learning”, et al 2023 (Decision Transformers are Bayesian meta-learners which do posterior sampling)
-
“Equitable Legal Remedies and the Existential Threat to Generative AI”, 2023 (argues US federal copyright courts have the power to order AI models outright erased & have done similar things before)
-
“Othello is solved”, 2023
-
“Robot developed by Ben Katz & Jared Di Carlo can solve Rubik’s Cube in record-breaking 0.38s” (the power of high-speed robotics & computer vision; unmodified, using a Lego robot instead: 1.8s; for comparison, the human all-time world record is 3.13s, or 2–8× slower.)
-
on Chinese AI: Where did all the Chinese Internet tokens go?; race-to-the-bottom price wars among Chinese LLMs; “China Is Losing the Chip War. Xi Jinping picked a fight over semiconductor technology—one he can’t win”; “For Chinese Students, the New Tactic Against AI Checks: More AI”/“Research misconduct in China: towards an institutional analysis”, 2024 (the fish rots from the head); “We simply need to turn the quantity of published articles into quality” —Zhang Hongjiang
-
Anthropic releases Claude-3.5-sonnet, Opus to follow
-
Ilya Sutskever launches ‘Safe Superintelligence’, a new startup to race for AGI by scaling LLMs
-
“AI Deception: A Survey of Examples, Risks, and Potential Solutions”, et al 2023; “Hoodwinked: Deception and Cooperation in a Text-Based Game for Language Models”, 2023; “Deception abilities emerged in large language models”, 2024 (LLMs given goals & inner-monologue increasingly can manipulate); “Reward hacking behavior can generalize across tasks”, Nishimura-et al 2024; “Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models”, et al 2024
-
“A Theory for Emergence of Complex Skills in Language Models”, 2023; “A Solvable Model of Neural Scaling Laws”, et al 2022
-
“Takeoff speeds presentation”, Tom Davidson 2023-09
-
“For months they toyed with ways to add more layers & still get accurate results. After a lot of trial & error, the researchers hit on system they dubbed ‘deep residual networks’” (the origins of algorithmic progress: cheap compute!)
-
“Robust agents learn causal world models”, 2024; “A Mechanistic Analysis of a Transformer Trained on a Symbolic Multi-Step Reasoning Task”, et al 2024 (Transformers can do internal planning in the forward pass)
-
“How AlphaFold Revolutionized Protein Science, but Didn’t End It”
-
“Probing the Decision Boundaries of In-Context Learning in Large Language Models”, et al 2024
-
“Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations”, et al 2024 (more power-law scaling of recommenders)
-
“A Benchmark for Learning to Translate a New Language from One Grammar Book”, et al 2023
-
“Market Concentration Implications of Foundation Models”, 2023 (are foundation models natural monopolies due to the enormous upfront capital costs?)
-
“Kasparov versus the World” (extreme diminishing returns to human coordination: 50k+ chess players + several IM/GMs couldn’t beat him!)
Genetics
Everything Is Heritable:
Recent Evolution:
Engineering:
Statistics/Meta-Science
-
“Will we continue scientific research?”, Alexander 1972 (on the future & morality of science)
-
Requests For Comment/Jagged_85: Evidence (If you ever feel confident about what you ‘know’ about the Islamic medieval period because you “just love reading Wikipedia”, you may want to know that in 201014ya, a crazy person spent 3 years falsifying hundreds of claims of Islamic invention, everything from “p-n junctions” to “capitalism”. He was banned in 201212ya, but it’s unclear how many falsehoods were left, whether he was using any other socks or anon IPs, why, what he has spent the past 12 years doing, or how far his claims have been carried by citogenesis or fellow Islamic ideologues. cf. Historia Augusta)
Politics/Religion
-
“Quantifying the potential persuasive returns to political microtargeting”, et al 2023 (microtargeting doesn’t work well: steep diminishing returns to personalization)
-
“The Beliefs of the Blob”, 2023 (how the US military-industrial complex sees the world & threats like China); Profile of the US State Department’s INR (a dissident intelligence agency whose predictions often differ from the rest of the Blob)
-
“Intergenerational Mobility in American History: Accounting for Race and Measurement Error”, 2023
-
“The unexpected origins of a modern finance tool: Discounting calculations are ubiquitous today—thanks partly to the English clergy who spread them amid turmoil in the 1600s, an MIT scholar shows” (2024; formal NPV estimates as mechanism for negotiation)
-
“Samsara”, Scott Alexander
Psychology/Biology
-
“Language is primarily a tool for communication rather than thought”, et al 2024 (still no complex cognition that requires thought, similar to 2021 failing to find any unique brain region critical to fluid intelligence. More evidence that while language is neither necessary nor sufficient for cognitive abilities, what it may be is critical for development: it is scaffolding. Language is efficient communication of knowledge which enables faster development. Anything learned with language can be learned without language, and once learned, language can be kicked away, like a ladder—but learning without language is slower, and will result in less per lifetime, and then what has been learned cannot be easily shared or preserved, so no accumulating ratchet either.)
-
“Goodbye, Ozempic: A new class of drugs is transforming obesity care. They are not all the same”; “‘Skinny jabs’: weight-loss drugs set for new boom as generic versions [of liraglutide] emerge; Alternatives to costly Wegovy and Saxenda will make such treatments more widely available worldwide”
-
“Social Status and Unethical Behavior: Two Replications of the Field Studies in et al 2012” , et al 2023 (failure to replicate claims the rich are more unethical)
-
“A Cause-of-Death Decomposition of Young Adult Excess Mortality”, et al 2018
-
“Ig Nobel Prize Winner Higashiyama Atsuki and the ‘Between-Legs Effect’ Mystery”
-
“The Cells That Eat, Regurgitate and Eat Your Tattoos Again” (et al 2018 ; permanent tattoo ink is fixed in place by waves of macrophages eating ink, storing, dying, and being replaced by more)
-
“True Porn Clerk Stories”, Ali 2002
-
Adragon De Mello (how to bring up genius)
Technology
-
The OceanGate disaster: how a charismatic high-tech startup CEO defied repeated correct predictions of disaster & created normalization of deviance by pushing to ship, inadequate testing, lying, firing dissenters, & gagging whistleblowers with NDAs (killing 5)
-
“The Turing-Complete User” design philosophy
-
“The Valve.Computer: A modern 8 bit design, built using 1950s thermionic valves”
-
“Repairing my mug with Kintsugi”, Fredrik Flornes 2024
Economics
-
“Microsoft Refused to Fix Flaw Years Before SolarWinds Hack” (normalization of deviance in large AI-related organizations: MS covered up the flaw that enabled Russia to extensively hack MS & the US federal government, pushed by Satya Nadella to sell cloud services)
-
“The Gender Gap in Confidence: Expected but Not [Fully] Accounted For”, 2024
-
“From Fire Hazards to Family Trees: The Sanborn Fire Insurance Maps”; “All That Is Solid Bursts into Flame: Capitalism and Fire in the Nineteenth-Century United States”, 2024
Philosophy
Fiction
-
“Can Poetry Matter? Poetry has vanished as a cultural force in America. If poets venture outside their confined world, they can work to make it essential once more”, Dana 1991
-
“The Golden Boy”, Gaiman 199331ya (rebuttal to “Watchmaker”)
-
“Extreme D&D DIY: adventures in hypergeometry, procedural generation, and software development (part 1)”, Said 2024 (modeling inter-dimensional invasion properly with higher-dimensional geometry for a D&D campaign)
-
Claire Jordan (“whitehound”/“Borolin”):
A blast from the old web: Claire Jordan is an elderly Scottish woman who has written extensively on fanfiction & related topics in British SF/F fandom after her degree in folklore & peripatetic UK career (running a Celtic/Wiccan shop, programming, evolution apologetics etc), since 198539ya (and at this domain since at least 200519ya). Among other things, she knew the man J. K Rowling had a crush on and based Severus Snape on, Scottish teacher/musician/activist John Nettleship. This gives her interesting insights into how the Harry Potter universe has evolved as fans create fanon, movies overwrite memories or insert assumptions of what the text must have meant, and J. K. Rowling keeps retconning or tweaking canon.
Miscellaneous
Books
Nonfiction:
-
Musui’s Story: The Autobiography of a Tokugawa Samurai, Katsu Kokichi, trans Craig 198836ya (David Kelly dramatic reading; a short but dense mid-life memoir by a boisterous ne’er-do-well samurai, who somehow survived his childhood and youth despite constantly getting into fights, running off, scamming people, falling into debt for parties & prostitutes. It is much like The Autobiography Of Benvenuto Cellini in that you are puzzled that he didn’t die, but then, you wouldn’t be reading his memoir if he had, would you? It is rich with incident (including a rather concerning number of serious injuries or infections of his testicles) and the texture of the Edo period life of a samurai from a moderately but not very well-off family, and like Ihara Saikaku’s Family Storehouse gives you a wonderful sense of the bustling metropolis and commerce of Tokugawa Japan as he wends his way through obsessively collecting swords (always carefully classifying them by length) and trying to avoid creditors or scolding family. You won’t necessarily get all that much out of reading the book compared to listening to David Kelly’s dramatic reading, but I still enjoyed it.)
Fiction:
Film/TV
Live-action:
Animated:
-
Frieren: Beyond Journey’s End (the based hit anime of the season, Sousou no Frieren lives up to its billing. The premise is unprepossessing: a near-immortal elf begins traveling after “the main story” has ended, as her companions start to die of old age, and begins to realize that the decade she spent with the party was the best decade of her life but she failed to realize it and threw the opportunity away, and now must come to grips with her grief and regrets. Hardly any anime has ever dealt with the themes of aging & loss as deftly as Frieren does, while still telling an entertaining story with solid action animation & a standout soundtrack. Like Odd Taxi, this will especially appeal to older or autistic fans, because younger fans will not have had the life experiences for the regrets to hit home. (The irony is not lost on me that, even as I watched Frieren, my cat of a decade had, unbeknownst to me, started to die of lung cancer.) I also think older techies & intellectuals will appreciate the theme of gradual human progress through science overturning historical laws, and the subtle pro-natalist viewpoint. Particularly for Western viewers, it features intriguingly un-Japanese worldbuilding—the demons & the Ubel character are unusually vivid depictions of psychopathy, with the demons seemingly ripped straight out of Blindsight. Anime tends to make almost a religion out of the theme of kokoro, and take it as axiomatic that communication is possible if only one makes enough effort and social harmony is always possible with enough effort, so I wonder how horrifying Japanese fans found the demons’ “blue-and-orange morality”, or if they were able to understand their dislike…?)
Music
MLP:
Doujin:
Misc: