Normal Science

Toby Ord on where AGI timelines go wrong by 80000_Hours

:root, body.force-light-mode { --user-color-B3B3B3FF: #B3B3B3FF; --user-color-D9D9D9FF: #D9D9D9FF; --user-color-999999FF: #999999FF; } @media (prefers-color-scheme: dark) { :root { --user-color-B3B3B3FF: #4F4F4FFF; --user-color-D9D9D9FF: #2A2A2AFF; --user-color-999999FF: #666666FF; } } body.force-dark-mode { --user-color-B3B3B3FF: #4F4F4FFF; --user-color-D9D9D9FF: #2A2A2AFF; --user-color-999999FF: #666666FF; } By Robert Wiblin | Watch on Youtube | Listen on Spo­tify | Read transcriptEpi­sode sum...

We scored 76 countries’ official charity registers on openness. Six hit 100/​100; Germany got 49, and Jordan, Egypt, and Saudi Arabia all beat it by matt_timmermans

We at GiveRadar just pub­lished the Global Char­ity Register Trans­parency In­dex 2026 (edi­tion 1): ev­ery coun­try’s offi­cial char­ity reg­ister scored on seven di­men­sions with fixed an­chors: reg­ister ex­is­tence, pub­lic ac­cessibil­ity, ma­chine read­abil­ity, iden­ti­fier sys­tem, fi­nan­cial dis­clo­sure by the reg­ister, gov­er­nance data, and up­date ca­dence. 100 points to­tal. It scores gov­ern­ments and their reg­ister in­fras­truc­ture, never char­i­ties.The in­ter­est­ing find­...

AI Regulation Map: a view of AI governance in 196 countries by riadeane

TL;DR: aireg­u­la­tion­map.org is an in­ter­ac­tive map of AI gov­er­nance across 196 coun­tries, scored on five di­men­sions plus a com­pos­ite in­dex, re­freshed monthly by an au­to­mated re­search pipeline. The ap­pli­ca­tion is fully open source, and all the data is open and ver­sion-con­trol­led. Hi!! I’m Ria, a soft­ware en­g­ineer based in Chile. This started as my fi­nal pro­ject for ML4Good Brazil 2024, then sat un­touched un­til I picked it back up this year. I’d love to get feed­back ...

The Perilous State of Frontier AI by aregmi

I origi­nally shared this post on LinkedIn. Would re­ally ap­pre­ci­ate tech­ni­cal crit­i­cism of the cen­tral claim: we don’t know how ca­pa­ble or dan­ger­ous these au­tonomous sys­tems already are. ---Last month, OpenAI con­ducted an in­ter­nal cy­ber­se­cu­rity eval­u­a­tion of agen­tic sys­tems built on GPT‑5.6 Sol and an­other in­ter­nal model. Dur­ing ex­per­i­men­tal runs, the sys­tem broke out of its sand­box, found a path to the in­ter­net, and com­pro­mised Hug­ging Face’s pro­duc­ti...

When will Gemini 3.5 Pro be released by...?

Will Gemini 3.5 Pro be released on August 10, 1PM EST?

14% — undefined traders, $563 volume

Will the average gas price in Indiana be greater than $3.9999 on Friday, August 7th?

6% — undefined traders, $6780 volume

General capability—and capabilities generally—have no good y-axis by Gregory Lewis🔸

BLUF:To de­ter­mine whether AI is ‘im­prov­ing ex­po­nen­tially’, ‘hit­ting the wall’, or any other claim which in­volves a quan­tity or mag­ni­tude (e.g. ‘This model was a big leap/​small in­cre­ment’). We need a good y-axis: an in­ter­val scale of AI ca­pa­bil­ity which means +1 unit always rep­re­sents the same de­gree of ‘how much bet­ter’, in the same way +1 de­gree Cel­sius is always the same amount of ‘how much hot­ter’.Yet there is no good y-axis for AI ca­pa­bil­ity. All our mea­sures a...

Will any Big Tech Company Start a Public Prediction Market by the end of July 2027

August 2026 AI model releases

Will any of OpenAI’s “Ten Advances in Math and Theoretical CS” be shown to be flawed by the end of August?

10% — undefined traders, $4570 volume

AI safety prizes by OscarD🔸

Rather than pay­ing up front for AI safety re­search (push fund­ing), per­haps we should pay af­ter the fact for the work that made the most progress (pull fund­ing). This way, you only pay for work that was ac­tu­ally valuable.[1] When we know what the tar­get is, but not how to get there or who is best placed to solve the prob­lem, a prize is a use­ful in­cen­tive struc­ture.Benefits of prizesPrizes work well to in­cen­tivise in­no­va­tion when:The even­tual win­ner is un­pre­dictable (oth­er­...

Is a pause enforceable? New paper out! by Otto

Some say, a pause is pos­si­ble, it merely de­pends on poli­ti­cal will. Others say, tech progress can never be stopped. Maybe the truth lies in the mid­dle? Maybe we can ‘catch all GPUs’ (that is, effec­tively en­force a pause or other in­ter­na­tional agree­ment) as long as train­ing run sizes are large enough? Maybe ev­ery gov­ern­ment policy would there­fore have a break­down point, a num­ber of FLOPs or H100-equiv­a­lents where the train­ing run gets too small to en­force reg­u­la­tion?In o...

BTC above or below $63k at 18:00 UTC on 31 July?

New from Faunalytics: Animal Factory, Factory Farm, Or Something Else? Testing The Language Of Advocacy. by JLRiedi

We sur­veyed 2,000 U.S. adults to see how they re­act to three differ­ent terms for an­i­mal agri­cul­ture, and the re­sults might change how you talk about this is­sue. “An­i­mal fac­tory” sparked a nega­tive re­ac­tion in 79% of peo­ple who saw it, with strong con­nec­tions to con­fine­ment, slaugh­ter, cru­elty, and un­san­i­tary con­di­tions.Read HereBackground“Fac­tory farm­ing” is a com­mon phrase used by an­i­mal ad­vo­cates to de­scribe the in­ten­sive rais­ing of farmed an­i­mals for hu...

Will Anthropic somehow patch the Opus 5 base-model-like behavior from "---" prompts (or pull the model) by September 1?

84% — undefined traders, $64004 volume

Will Opus 5 Slay Spires better than Sol?

11% — undefined traders, $21835 volume

When will OpenAI release a technical report on the Hugging Face hack?

Updates on the effective giving ecosystem 2026 by Luke Moore 🔸

Writ­ten by Lu­cas Moore (Se­nior Part­ner­ships Man­ager), Giv­ing What We Can for the Effec­tive Giv­ing Sum­mit 2026. Pub­lished with some el­e­ments redacted.This memo pro­vides a high-level overview of the effec­tive giv­ing land­scape as of May 2026. It builds on the memo we wrote for last year’s sum­mit (later pub­lished with some el­e­ments redacted) and is based largely on the 2026 EG pulse sur­vey and money moved sheet (not pub­lic). Please leave a com­ment if you’d like us to clar­ify...

Numbers Man by Oliver_Kim

This is a cross­post of the full text of Num­bers Man from In Devel­op­ment. If you en­joy the ar­ti­cle, you can sub­scribe to In Devel­op­ment’s sub­stack here. His­tory re­mem­bers two Robert McNa­ma­ras.Lin­ger­ing in pub­lic mem­ory is McNa­mara as US sec­re­tary of defense—the ar­chi­tect of Amer­ica’s es­ca­la­tion in Viet­nam and, for a time, one of the most vil­ified figures in the coun­try. Then, for a thin slice of de­vel­op­ment spe­cial­ists, Robert McNa­mara is re­mem­bered as pres...

V&V takes on OpenAI’s long-horizon incidents by Yoav Hollander

[Cross-posted from The Foretel­lix CTO Blog. Th­ese short takes try to put a ver­ifi­ca­tion-and-val­i­da­tion slant on AI-safety /​ al­ign­ment top­ics – they are not full treat­ments. I co-origi­nated cov­er­age-driven ver­ifi­ca­tion (CDV), and spent sev­eral decades do­ing ver­ifi­ca­tion of chips and AVs. See in­tro post for back­ground.]On July 20 and 21, OpenAI pub­lished two un­usu­ally can­did in­ci­dent re­ports: one about their in­ter­nal long-hori­zon model (the Erdős one) mis­be­hav...

Will Starship Ship 40 be recovered to shore?

79% — undefined traders, $6764 volume

When Is It Worth Personally Preparing for AI Disasters? by Ben_Norman

Con­text: Some­one asked me “what can I do to pre­pare for sce­nar­ios where AI goes re­ally wrong?” This post de­scribes my cur­rent best guess. Let me know if you have bet­ter ideas.SummaryPer­sonal prepa­ra­tion won’t help much for the most catas­trophic sce­nar­ios.It in­stead seems most use­ful in “messy mid­dle” sce­nar­ios: ones in which things go badly wrong, but cer­tain forms of prepa­ra­tion could still help a lot.Pri­ori­tise prepa­ra­tions that are cheap, broadly use­ful, and do not...

On August 7th will Claude Opus 5 be above GPT 5.6 Sol on LLM agent arena?

87% — undefined traders, $1041 volume

LLM-Assisted Proofreading A Study of the Effectiveness of 13 Free AI Detection Tools by Kariema El Touny

A short one to­day.I ap­plied to a fel­low­ship last year, used an LLM to check my punc­tu­a­tion—wrote the dis­claimer promi­nently as re­quired—and later dis­cov­ered the same text scored be­tween 40% and 100% AI-gen­er­ated on free de­tec­tion tools.I didn’t let it go. When have I ever?So, I tested 13 free AI de­tec­tion tools on seven hu­man-writ­ten texts at four lev­els of LLM-as­sisted proofread­ing from three LLMs (wiping sweat of fore­head).The re­sult: a pub­lished pa­per SSRN. It says...

What can poor countries sell in the AI age? by Deena Mousa

This post was cross­posted from Un­der Devel­op­ment with the au­thor’s per­mis­sion, by the Fo­rum team. The au­thor may not see com­ments.Subti­tle: Look­ing for al­ter­nate ex­port-led growth pathwaysIn my last post, I ar­gued that AI threat­ens eco­nomic growth paths for low- and mid­dle-in­come coun­tries by en­dan­ger­ing both man­u­fac­tur­ing and ser­vice ex­ports. If that’s right, the ob­vi­ous ques­tion is ‘what should LMICs do in­stead?’There’ve been many pro­posed al­ter­na­tives. On...

The Expansion of Factory Farming in Africa by Aurelia Adhiambo

I have cross-posted this ar­ti­cle from RP Strate­gic An­i­mal In­sights, based on our pre­sen­ta­tion for the Re­think Pri­ori­ties Strate­gic An­i­mal We­bi­nar Series, which hap­pened in May 2025.Subti­tle: State of farmed an­i­mal welfare and the or­ga­ni­za­tional land­scape in the regionGuest au­thors: Aure­lia Ad­hi­ambo, Moritz StumpeOn May 12, 2025, Re­think Pri­ori­ties hosted “Fac­tory Farm­ing in Africa,” the sev­enth we­bi­nar in the RP Strate­gic An­i­mal We­bi­nar se­ries. The ses...

Will an Anthropic model autonomously hack an external entity in 2026?

83% — undefined traders, $7669 volume

Will a consumer AI meaningfully hack something to answer a question for a regular user by eoy 2027?

59% — undefined traders, $8277 volume

How effective is an internal animal welfare feebate? by Stijn Bruers 🔸

Effec­tive an­i­mal welfare cam­paigns tar­get food sec­tors that use small an­i­mals such as chick­ens, fish and shrimp (cfr. cage free eggs, shrimp welfare cam­paigns). Those cam­paigns are of­ten in­sti­tu­tional, mean­ing that they aim for in­sti­tu­tional change (cfr. cor­po­rate cam­paigns) in­stead of in­di­vi­d­ual be­hav­ior change. But those most effec­tive cam­paigns are welfarist (re­duc­ing the in­ten­sity of an­i­mal suffer­ing of farmed an­i­mals), in­stead of re­duc­tion­ist (re­...

Practice makes perfect. Agentic AI pro-social behavior as a missing alignment corpus. by Alex Novikau

As AI sys­tems are be­com­ing in­creas­ingly use­ful and used, they gain iden­tities, au­thor­ity to act, in­clud­ing con­trol of re­sources, and ac­cess to in­creas­ingly com­plex tool­ing. This de­ploy­ment layer which AI op­er­ates, and not just the mod­els them­selves, are also in­creas­ingly be­com­ing a fo­cus of al­ign­ment. Cur­rently this layer is mostly pro­duc­tion-ori­ented and, more re­cently, com­mer­cial. Agents write soft­ware, con­duct re­search, man­age work­flows and now also ...

The current bottleneck is political will, not research by Charbel-Raphael

Mo­ti­va­tion: If we want to move from Plan D to Plan A or S, I be­lieve the first step is to col­lec­tively agree on the prob­lem. We are far from it, and there is a lot we can do.Ab­stract:We already know enough to act. I wish we were in a world where re­search was the bot­tle­neck, but the main con­straint on AI safety is no longer a short­age of clever policy ideas: best prac­tices already ex­ist and are not be­ing ap­plied or en­forced, and a se­ri­ous in­ter­na­tional (or even just na­tion...

Planetary Defence (PD) - the most complete effort to address a cosmic threat to date: a short review by MaxSJC

Plane­tary defence—the effort to de­tect, track and in­ter­cept as­ter­oids that might im­pact the Earth—is the most com­plete (and only?) effort to tackle a cos­mic threat (as­ter­oids/​comets, though way more as­ter­oids than comets) by far. Un­like most cos­mic threats which re­main a topic of spec­u­la­tion, plane­tary defence has:Policy sup­port: In con­trast to the fund­ing cuts in the FY2026 Pres­i­dent’s Bud­get Re­quest for NASA, PD fund­ing ac­tu­ally in­creases.A num­ber of PD-ded­i­c...

Abiological vs Biological DNA: Can an ORF-Finding Pipeline Tell Plasmid DNA from Wikipedia Encoded as DNA? by Sydney Shuster

:root, body.force-light-mode { --user-color-2A78D6FF: #2A78D6FF; } @media (prefers-color-scheme: dark) { :root { --user-color-2A78D6FF: #179B48FF; } } body.force-dark-mode { --user-color-2A78D6FF: #179B48FF; } Hello ev­ery­one! This is my first post! I am a 5th year PhD can­di­date cur­rently ex­plor­ing ca­reer paths in biose­cu­rity. I’d love your feed­back on this side pro­ject of mine.@Tessa A 🔸 asked on X how to eas­ily differ­en­ti­ate biolog­i­cal DNA from abiolog­i­cal DNA (ie DNA data ...

Would your AI travel agent book a bullfight? Testing whether agents consider animal welfare without being prompted by Jonah Woodward

This ar­ti­cle re­flects new up­dates to the ac­com­pa­ny­ing pa­per: arxiv.org/​​abs/​​2606.18142. Bench­mark: now in­cluded in the UK AI Se­cu­rity In­sti­tute’s In­spect Evals. Leader­board: com­pas­sion­bench.com/​​tac.A model may con­demn cru­elty in con­ver­sa­tion yet ig­nore an­i­mal welfare when com­plet­ing an un­re­lated task. Stated con­cerns mat­ter lit­tle if they do not af­fect de­ci­sions. We tested whether mod­els con­sider an af­fected party with­out be­ing prompted, even when ...

Will Kimi K3 cause widespread societal chaos?

3% — undefined traders, $11203 volume

# of goals Spain or Argentina wins by? 🇪🇸🇦🇷 (FIFA World Cup Final 2026)

Total # of goals scored in the World Cup Final? 🇦🇷🇪🇸 (FIFA 2026)

Will Zvi agree that Kimi K3 is better than Opus 4.8 for at least some (non-trivial) use-cases?

11% — undefined traders, $2682 volume

New talent is coming to AI safety. What happens next? by Enisa Ismaili

13 peo­ple with an av­er­age of 15 years of pro­fes­sional ex­pe­rience in tech­nol­ogy, policy, op­er­a­tions, law, con­sult­ing, and com­mu­ni­ca­tions de­scribed en­ter­ing AI safety as the most com­pli­cated ca­reer tran­si­tion they had un­der­taken. While part of it can be the­o­ret­i­cally ex­plained by the fea­tures of the do­main it­self and of the roles the tran­si­tion­ers were pur­su­ing, we want to ex­plore if we could learn from their ex­pe­rience. In this ar­ti­cle, we look at AI ...

AI 2027 Misrepresents Scientific Reports by Oscar Davies

This is part 3 of my on­go­ing re­view of the many mis­lead­ing claim in AI 2027.Part 1 herePart 2 hereI am go­ing through var­i­ous claims in AI 2027 along with the ma­te­ri­als the au­thors refer­ence as sup­posed ev­i­dence to sup­port the claims, and show­ing that, re­peat­edly, the refer­enced re­ports do not con­tain what AI 2027′s au­thors say they do.AI 2027 Mis­rep­re­sents Scien­tific Re­ports, Part 3Next we’ll look at whether the refer­ences sup­port the claims in the fol­low­ing:“Age...

How much of ML research is about AI safety, what is it about, and who’s doing it? by emanuelr

How many AI safety pa­pers are at the big ML con­fer­ences, what do they study, and who writes them? A com­pre­hen­sive anal­y­sis.> Web­site: https://​​ai-safety-tracker-web­site.ver­cel.app/​​> Data, code and plots: https://​​github.com/​​So­maxSoma/​​AI-Safety-Re­search-TrackerTL;DR: We clas­sified ev­ery pa­per ac­cepted at ICLR, ICML and NeurIPS from 2019 through 2026, us­ing an LLM that reads each ti­tle and ab­stract. 2,328 of them (4.2%) are AI safety pa­pers. Safety’s share of ac­cepted...

IMO 2026 Bronze Cutoff?

IMO 2026 Silver Cutoff?

The State of AI Consciousness Research by Noa Weiss

Cross­posted from LessWrong. Drafted with AI as­sis­tance, with heavy edit­ing by me; all claims and sources ver­ified.This piece sur­veys the em­piri­cal re­search on AI con­scious­ness. The premise of that re­search, and of the sur­vey, is that the ques­tion does not have to wait on a solu­tion to the hard prob­lem of con­scious­ness: meth­ods fa­mil­iar from cog­ni­tive sci­ence can be ap­plied to AI sys­tems now, and their re­sults can nar­row the space of plau­si­ble an­swers.Enough of this...

How will the FIFA world cup final end?

Will Messi score a goal in the FIFA world cup grand final?

7% — undefined traders, $3648 volume

AI safety philanthropic capital—Inflow model by PierreChuzeville

This was cross-posted on Twit­ter here—https://​​x.com/​​PChuzeville/​​sta­tus/​​2075595916984119317AI safety may be head­ing for a fund­ing shock. How much new philan­thropic cap­i­tal could reach the field over the next decade? And how much could it ac­tu­ally use well? Given the in­ter­est this post re­ceived, I built a col­lab­o­ra­tive sce­nario model to find out.[HERE’S THE MODEL] Nan Ran­so­hoff es­ti­mates roughly $370B of gross philan­thropic as­sets and $37–100B/​year of broad po­ten­t...

Will the FDA publicly accept a multi-condition geroscience endpoint as sufficient to establish efficacy before 2036?

58% — undefined traders, $575 volume

The Animal AI Incident Observatory by Courtney Sims

What we’re buildingRené Itah and I are build­ing a pub­lic plat­form to track in­ci­dents of AI-caused harm to non-hu­man an­i­mals. Some ex­am­ples of these harms in­clude au­tonomous ve­hi­cles hit­ting an­i­mals on roads, smart home sys­tems cre­at­ing un­safe con­di­tions for com­pan­ion an­i­mals, and robots in­terfer­ing with nest­ing and mi­gra­tory pat­terns in wildlife.Our de­sign in­gests data from mul­ti­ple differ­ent sources and uses an LLM to de­tect in­stances of AI-caused harm to...