Normal Science

Advocates Can Influence LLM Values By Editing Wikipedia by Jonah Woodward

This ar­ti­cle is a sum­mary of an origi­nal study: Brazilek, J., Navas, M., & Gnauck, A. (2026). Small ed­its, large mod­els: How Wikipe­dia ad­vo­cacy shapes LLM val­ues. Zen­odo. https://​​doi.org/​​10.5281/​​zen­odo.19981454 We’d like to thank Alexa Gnauck and Maria Navas for their in­dis­pens­able con­tri­bu­tions to this re­search. TL;DRWikipe­dia ar­ti­cles are weighted highly among LLM train­ing data. Thus, we show that it’s pos­si­ble for ad­vo­cacy or­gani­sa­tions to in­fluence what L...

Model evals are eating AI safety hiring (and why the “AI safety” title is disappearing) by AI Safety Careers

I’ve been spend­ing a lot of time dig­ging through lab ca­reer pages, think tank sites, and re­search orgs lately for a small pro­ject I’m run­ning. Look­ing at the raw data from the last few months, there’s a re­ally ob­vi­ous shift in what teams are ac­tu­ally hiring for right now: model eval­u­a­tions, red team­ing, and pre­pared­ness are ex­plod­ing.It seems like labs and gov­ern­ments (like the var­i­ous AI Safety In­sti­tutes) are hit­ting the ex­act same wall at the same time. Every­one i...

Cluelessness: Summary of the argument, why it matters, and counterarguments by Anthony DiGiovanni 🔸

I’d like to elicit di­rect, pro­duc­tive cri­tiques of the ar­gu­ment for clue­less­ness from my se­quence on “un­aware­ness”, which I’ll call the un­aware­ness ar­gu­ment. To that end, this post will: break down the un­aware­ness ar­gu­ment at a high level;ex­plain why the EA com­mu­nity should care about this ar­gu­ment; andsum­ma­rize the an­gles for cri­tiquing the ar­gu­ment that I ex­pect to be most pro­duc­tive, and how the se­quence re­sponds to ex­ist­ing cri­tiques. Ar­gu­ment breakdow...

We can guess what intergalactic war would look like. And strangely, it matters. by 80000_Hours

By Robert Wiblin | Watch on Youtube | Listen on Spo­tify | Read transcriptIn­ter­galac­tic war is prob­a­bly billions of years away — yet physics can already tell us how it ends. And strangely that con­clu­sion is rele­vant to de­ci­sions peo­ple have to make to­day.In this video, Rob Wiblin walks through a fas­ci­nat­ing anal­y­sis from re­searcher Beren Millidge that uses known physics — no worm­holes or faster-than-light travel — to iden­tify the only three weapons that could work at an in­te...

Will the US attack Cuba before the 2026 midterm elections?

32% — undefined traders, $2143 volume

Will Anthropic release a non-Fable model before Fable is rereleased?

17% — undefined traders, $5675 volume

Will a current moderator lose their role before end of month?

28% — undefined traders, $304 volume

Introducing A Beginner’s Guide to Digital Minds by Lucius Caviola

We’ve launched A Begin­ner’s Guide to Digi­tal Minds (digi­tal­minds.guide), a web­site for peo­ple who want an in­tro­duc­tion to AI con­scious­ness, AI welfare, and the broader im­pli­ca­tions of the pos­si­bil­ity that AI sys­tems could mat­ter morally.The guide pro­vides an overview of the ques­tions mo­ti­vat­ing digi­tal minds re­search, in­clud­ing: What are digi­tal minds? Could AI sys­tems be con­scious? How might they work? Could they mat­ter morally? And given the un­cer­tainty, how s...

Tactical and Operational Exploratory Modeling for AI Governance by Dawn Drescher

Us­ing com­pu­ta­tional meth­ods to im­prove our pre­pared­ness via more ro­bust and adap­tive strate­gies in AI gov­er­nance. A pro­ject pro­posal for a think tank, con­sul­tancy, or soft­ware.OverviewOver the years, I’ve come across or come up with a num­ber of pro­ject ideas in AI safety and gov­er­nance that I find promis­ing. My top list has less than ten, but in to­tal there are hun­dreds. Either way, too many for me to re­al­ize them all. In­stead I want to pro­mote these ideas in the hop...

What Is It Like to Be a Bass? Red Herrings, Fish Pain and the Study of Animal Sentience by Vasco Grilo🔸

This is a linkpost for What Is It Like to Be a Bass? Red Her­rings, Fish Pain and the Study of An­i­mal Sen­tience by Ge­or­gia Ma­son and Michelle Lav­ery, which was origi­nally pub­lished in Fron­tiers in Ve­teri­nary Science on 27 April 2022. The ab­stract and last sec­tion are be­low. Re­lat­edly, I liked the ar­ti­cle In search for con­scious­ness in an­i­mals: Us­ing work­ing mem­ory and vol­un­tary at­ten­tion as be­hav­ioral in­di­ca­tors by An­dreas Nieder, which was pub­lished in Neu­r...

AWASH: Intervention Selection—Improving farmed fish welfare in Ghana by Naveeth Basheer

This is the ex­ec­u­tive sum­mary of a re­port, pub­lished on 15 May 2026 by AWASH. You can find the full re­port here: linkEx­ec­u­tive summaryAWASH has com­pleted two phases of re­search on fish welfare in Ghana­ian aqua­cul­ture: a desk-based liter­a­ture re­view ti­tled Aqua­cul­ture and Fish Welfare in Ghana, and an on-site scop­ing study that in­cluded vis­its to eight large com­mer­cial cage farms on Lake Volta, along with stake­holder in­ter­views with farm­ers, hatcheries, vet­eri­nar­i...

Epistemic Exclusion of the Hausa Language in AI Safety by Hadiya Usman

I. I was sit­ting with a pam­phlet in front of me, an AI safety aware­ness flyer we were prepar­ing for a pro­ject called Safer Schools, Safer Com­mu­ni­ties, de­signed to be printed and dis­played as posters in schools and com­mu­nity spaces across north­ern Nige­ria. My job was to trans­late it into Hausa so that the stu­dents, teach­ers, and com­mu­nity mem­bers who walked past it ev­ery day could ac­tu­ally read it and un­der­stand it.I got stuck al­most im­me­di­ately.Not be­cause the Hausa...

Fable reenabled for Europeans before July 1?

22% — undefined traders, $140967 volume

Will Claude Fable 5 be re-enabled for Non-Americans before July 1st, 2026?

40% — undefined traders, $4160 volume

When will Anthropic's API customers get access to Fable 5 again?

are we locked into P(doom)? by keivn

start­ing with the conclusionear­lier this year I was a part of the lat­est co­hort at AI safety camp (AISC). the pro­ject I had joined — In­ves­ti­gat­ing as­sump­tions of the doom de­bate — had a goal to build a prob­a­bil­is­tic as­sump­tion tree for AI doom ar­gu­ments.sev­eral mem­bers made ver­sions of the tree, but the one that we col­lec­tively worked on, started from a root node “Danger­ous AI oc­curs.” one branch — “model ca­pac­ity”broke down into a dozen subn­odes, from in­fer­ence a...

Will the biosecurity interventions succeed in a resource-limited setting like Nigeria in the event of an engineered pandemic? by Nnaemeka Emmanuel Nnadi

TL;DRMany biose­cu­rity in­ter­ven­tions as­sume that if we cre­ate effec­tive defences to man­made pan­demics, so­ciety will be able to em­ploy them when nec­es­sary. How­ever, this as­sump­tion is mostly pred­i­cated on fac­tors preva­lent in high-in­come coun­tries, in­clud­ing strong in­sti­tu­tions, pub­lic trust, de­pend­able health­care sys­tems, ro­bust surveillance net­works, and ap­pro­pri­ate fi­nance. In this pa­per, I sug­gested that mil­i­tary mod­els de­signed and as­sumed to func...

How bad would it be if GPS satellites were shot down? by Jackson Wagner

Los­ing GPS isn’t an X-risk, but would cre­ate a huge dis­aster on the scale of Covid-19 or big­ger.Hi! From 2020 − 2023 I was one of the early em­ploy­ees at Xona Space Sys­tems, a com­pany work­ing on es­sen­tially a next-gen­er­a­tion ver­sion of GPS. I ended up learn­ing a lot about how GPS works and what it’s used for, and (due to my per­sonal in­ter­est in effec­tive al­tru­ism) ended up do­ing some re­search into what would hap­pen if to­day’s GPS sys­tems sud­denly failed. This post is t...

Safe for What World? Why the AI safety field may be asking the wrong question by Benoît L.

I used an LLM to help re­draft some ar­gu­ments for an EA spe­cific au­di­ence and it likely con­tains ca.10% AI-gen­er­ated text, but I’ve ed­ited/​rewrit­ten it ex­ten­sively and en­dorse it.Epistemic sta­tus: The em­piri­cal claims rest on in­sti­tu­tional sources cited at the end. The cen­tral ar­gu­ment, that AI safety re­search as a field does not ask whether its de­ploy­ment tra­jec­tory is com­pat­i­ble with the phys­i­cal sys­tems it draws on, is visi­ble in what the field’s re­search a...

[New Paper] Prioritizing Risks from AI: A Delphi Study of 272 Experts by Peter Slattery 🔸

TL;DR: We ran a Delphi study with 272 in­ter­na­tional AI ex­perts to pri­ori­tize 24 AI risk do­mains from the MIT AI Risk Do­main Tax­on­omy. In a busi­ness-as-usual sce­nario, ex­perts judged a more than 10% chance of catas­trophic out­comes (i.e., ‘more than 1 mil­lion hu­man deaths or more than a USD 100B in fi­nan­cial loss or civ­i­liza­tional-scale in­tan­gible im­pacts’) from 18 of our 24 AI risk do­mains over the next five years. They also iden­ti­fied a re­spon­si­bil­ity gap: AI user...

Will the WTI Crude Oil Spot Price be above $94 on June 22, 2026?(See Description)

1% — undefined traders, $10454 volume

Are top existential risk estimates 50,000 times too high? An optimizer’s curse model and analysis by titotal

.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabin...

Timelines to what? A proposal by tlevin

This post was cross­posted from Mul­ti­plier with the au­thor’s per­mis­sion, by the Fo­rum team. The au­thor may not see com­ments.Subti­tle: Why you might want to DIAL in to short timelinesPeo­ple in the AI field love to talk about “timelines.” So­phis­ti­cated peo­ple in the AI field love to ask: “timelines to what?” Au­toma­tion of AI R&D? (By which met­ric?) AI that far sur­passes hu­mans at nearly all cog­ni­tive tasks, or “TEDAI”? Ar­tifi­cial gen­eral in­tel­li­gence? Some­thing more am­...

Will silver hit $60/oz by close on Friday?

6% — undefined traders, $1089 volume

Will Claude Fable 5 be a accessible in a Claude max 20x subscription on June 23 2026?

38% — undefined traders, $5152 volume

Incumbents Might Be Replaced During AI Takeoff, Including in Animal Agriculture by Hazo

I used an LLM to help draft this post and it likely con­tains >10% AI-gen­er­ated text, but I’ve ed­ited/​rewrit­ten it ex­ten­sively and en­dorse it.TL;DRIt’s un­clear how much the in­tel­li­gence ex­plo­sion will di­rectly af­fect agri­cul­ture, be­cause it’s one of the least cog­ni­tive-la­bor-in­ten­sive in­dus­tries. But the in­dus­trial ex­plo­sion that may fol­low could make even an­i­mal agri­cul­ture sus­cep­ti­ble to dis­rup­tive in­no­va­tion (pos­si­bly the only time in his­tory that...

CATL receives license renewal for Yichun Lithium mine by end of Feb 2027?

57% — undefined traders, $50089 volume

Coverage-driven alignment—What ‘Teaching Claude Why’ can borrow from AV verification by Yoav Hollander

:root, body.force-light-mode { --user-color-FFFFFFFF: #FFFFFFFF; } @media (prefers-color-scheme: dark) { :root { --user-color-FFFFFFFF: #000000FF; } } body.force-dark-mode { --user-color-FFFFFFFF: #000000FF; } Cross-posted from The Foretel­lix CTO Blog. Also on LessWrong.Epistemic sta­tus: A method­ol­ogy im­port ar­gu­ment, not a re­sult. I’m fairly con­fi­dent that cov­er­age-driven ver­ifi­ca­tion (CDV), bor­rowed from au­tonomous-ve­hi­cle de­vel­op­ment, has trans­fer­able struc­ture for al...

Human Presence as External Variable in AI Self-Expression: A Pilot Study by L.Raeva

Edi­to­rial as­sis­tance pro­vided by Claude (An­thropic). The con­cep­tual frame­work, hy­poth­e­sis, study de­sign, ob­ser­va­tions, and anal­y­sis are the au­thor’s ownA grow­ing field of re­search ex­am­ines the moral sta­tus and welfare im­pli­ca­tions of ad­vanced AI sys­tems. Long, Sebo and col­leagues (2024) ar­gue that AI welfare re­quires rigor­ous as­sess­ment amidst sub­stan­tial un­cer­tainty sur­round­ing con­scious­ness and agency. But­lin, Long and col­leagues (2023) provide a co...

Claude Fable released by June 17?

99% — undefined traders, $11153 volume

Will there be more than 4 mass shootings in America this weekend?

91% — undefined traders, $5930 volume

Will Claude Mythos be released by the end of June 2026?

99% — undefined traders, $22537 volume

At the end of the next 30 days, starting from today (June 7 2026), how many blocks will be broken on mysterybox.art ?

Assert, don’t describe. Linguistic Features that shift LLM reasoning about animal welfare by Jasmine Brazilek

tl;drThe way things are said (“lin­guis­tic fea­tures”) in fine-tun­ing data af­fect AI views on an­i­mal welfare. Some fea­tures de­grade com­pas­sion to­wards an­i­mals, some have neg­ligible im­pact, and oth­ers bolster it. We recom­mend you be­come fa­mil­iar with these fea­tures If your writ­ing may end up in LLM fine-tun­ing train­ing data. In short: as­sert a po­si­tion with moral vo­cab­u­lary rather than de­scribe a scene neu­trally, and avoid hedg­ing and overly con­crete sen­sory de­s...

Estimating the costs of pandemic-proofing the EU by Jeanne Marie Jacqueline (JMJ/​Evana)

BackgroundCovid-19 was the trial by fire of con­tem­po­rary times in terms of cop­ing with pan­demic risks. It showed how un­pre­pared the world was to such an out­break but also was a cau­tion­ary tale for more dan­ger­ous dis­eases that could arise in the fu­ture. Nonethe­less, many co­op­er­a­tive ac­tions and pro­grammes came out of this crisis, vac­cines never de­vel­oped so fast, and cam­paigns meth­ods evolves or were tested to in­crease effi­ciency. It was a les­son-learned for the EU, w...

Will Navier-Stokes be considered solved on June 3th 2027?

16% — undefined traders, $9891 volume

What will Elons Musk Net Worth reach before July 2026?

Sixteen schemes for AI safety by Austin

Th­ese days, I of­ten run across whip­per­snap­pers ex­cited to do some­thing for AI safety — but aren’t quite sure what. One of the fun things about the Fu­ture Fund era were the big lists of pro­ject ideas; as we en­ter a new era of crazy money slosh­ing around, it might be time to bring back the lists!Note that these ideas range from “very con­fi­dent this is good” to “com­pletely hare­brained”; I’m not tel­ling you which are which.If you’re ex­cited for ideas like these, con­sider join­ing S...

Feedback wanted on a curated AI safety & governance jobs board by AI Safety Careers

Dis­clo­sure: I used AI as­sis­tance to draft parts of this post, then re­viewed and ed­ited it be­fore post­ing.I’ve been work­ing on a small pro­ject called AI Safety Ca­reers: a cu­rated job board for roles in AI safety, AI gov­er­nance, policy, eval­u­a­tions, red team­ing, re­spon­si­ble AI, and re­lated fron­tier AI work.The rea­son I started build­ing it is pretty sim­ple: when I was look­ing through roles in this space, I found that rele­vant jobs were scat­tered across a lot of differ­e...

Will Spacex close above a $2.5 Trillion Market Cap on it's first day of trading (IPO)?

16% — undefined traders, $15304 volume

Critique of current AI safety bug bounty programs by clickyquack

The po­ten­tial value of AI safety bug bounty programsGen­er­ally, AI labs should (and most do) put their mod­els un­der ex­ten­sive safety test­ing be­fore de­ploy­ing them to pre­vent mi­suse, schem­ing, and other dan­ger­ous be­hav­iors. This may in­clude in­ter­nal tests, red-team­ing efforts by third-par­ties, etc. How­ever, edge case safety vuln­er­a­bil­ities will likely slip through, and these can still cause dam­age. If any of the risks of AI sys­tems from labs that im­ple­ment strong s...

What will Anthropics Public Ticker be?

US Gas Price $4.250 or below in June 2026?

99% — undefined traders, $13190 volume

We Need Breadth-First AI Safety Plans by MichaelDickens

Cross-posted from my web­site. Depth-first plans lay out a path from here to al­igned su­per­in­tel­li­gent AI. We need those kinds of plans. But depth-first plans de­pend on many as­sump­tions: “We will make AI safe by do­ing step 1, then step 2, then step 3.” Step 1 only works un­der con­di­tion A, step 2 re­quires con­di­tion B, step 3 re­quires con­di­tion C. If A or B or C is false, the whole plan fails (and there’s a good chance we all die). Con­sider Google’s safety plan from April 2025. ...

Who controls compute supply chains? Looking into Japan as an underexplored case by Mahi

This post is part of a com­par­a­tive pro­ject on East Asia com­pute gov­er­nance, ini­ti­ated as part of the inau­gu­ral round of AI Gov Sprints. Many thanks to David Sanchez Gar­cia, who led this pro­ject and whose feed­back greatly benefited this piece, and to all who gave feed­back on ear­lier drafts. This post was lightly ed­ited for flow us­ing Claude; all ideas and ar­gu­ments are my own. All mis­takes, too, are my own. A lot of com­pute gov­er­nance dis­course fo­cuses on the US-China dy...

Will Bitcoin be below 71K at end of week?

99% — undefined traders, $4207 volume

Who will have the highest sharpe ratio of the website by June 20th 8 PM CEST [ADD ANSWERS] (2026)

What body fat % will my next dexa scan show?

OPQA (OpenAI-proof QA) hits 20% before 2027

59% — undefined traders, $3988 volume

To what extent will Developmental Cognitive Interpretability be successful [Read Updated Description]