Had fun researching this piece, which cites Keynes & Marx but also folks of much more recent vintage, including the economists Phillip Trammell, @briancalbrecht.bsky.social, @aleximas.bsky.social; plus @dwarkesh.bsky.social, @sebk.bsky.social, Boaz Barak, Geoffrey Hinton.

An Apparition. Slide from my collection, 1975.

super interesting take blog.cosmos-institute.org/p/coasean-ba...

ICYM: International bodies for which the US Administration has apparently stopped paying assessed dues include (thread):

What will AI look like by 2030 if current trends hold? Our new report zooms in on two things: (1) whether scaling continues (compute, data, power, capital), and (2) the capabilities this enables—especially for scientific R&D.

AI can help protect democracies or empower those who threaten freedom: corporations, oligarchs, or governments. The key question is: who does AI serve? We should all care about the answer, writes Richard Reisman (@rreisman.bsky.social).

🧵 Several people have raised questions about what I meant by "deservedly so" when referring to the decline of institutional trust. It's a fair question, so I want to explain that, because it sits at the heart of my work on how democracies collapse when their epistemic foundations rot.

I know Yglesias isn't everyone's favorite, but it's amazing how angry many people -often with PhDs-get at a suggestion that there are tradeoffs in politics. Remarkably, many who purport to believe the republic is hanging by a thread are in the "no compromise with the voters!" camp.

I'm quoted in this fact-check for the claim that the UK has "the 12th largest female health gap in the world", which was widely shared by MPs & media in the UK. The answer is no, and the male wellness company who calculated this statistic reviewed & deleted their page while this piece was written.

Announcing our latest arxiv paper: Societal and technological progress as sewing an ever-growing, ever-changing, patchy, and polychrome quilt arxiv.org/abs/2505.05197 We argue for a view of AI safety centered on preventing disagreement from spiraling into conflict.

Join us for our final symposium panel, Panel 7: Transitions to Transformative AI, at 2:15pm ET. Panelists: @jbullockruns.bsky.social, @ruchowdh.bsky.social, @sebk.bsky.social, and Sam Manning. Moderator: @sethlazar.org. #AIDemocraticFreedoms

Not you. take CURSE OF RA 𓀀 𓀁 𓀂 𓀃 𓀄 𓀅 𓀆 𓀇 𓀈 𓀉 𓀊 𓀋 𓀌 𓀍 𓀎 𓀏 𓀐 𓀑 𓀒 𓀓 𓀔 𓀕 𓀖 𓀗 𓀘 𓀙 𓀚 𓀛 𓀜 𓀝 𓀞 𓀟 𓀠 𓀡 𓀢 𓀣 𓀤 𓀥 𓀦 𓀧 𓀨 𓀩 𓀪 𓀫 𓀬 𓀭 𓀮 𓀯 𓀰 𓀱 𓀲 𓀳 𓀴 𓀵 𓀶 𓀷 𓀸 𓀹 𓀺 𓀻 𓀼 𓀽 𓀾 𓀿 𓁀 𓁁 𓁂 𓁃 𓁄 𓁅 𓁆 𓁇 𓁈 𓁉 𓁊 𓁋 𓁌 𓁍 𓁎 𓁏 𓁐 𓁑 𓀄 𓀅 𓀆

This Thursday 4/10 & Friday 4/11, we're hosting our symposium "AI and Democratic Freedoms." @jbullockruns.bsky.social, @ruchowdh.bsky.social, @sebk.bsky.social & moderator @sethlazar.org will help wrap up the event with a final panel. #AIDemocraticFreedoms RSVP: www.eventbrite.com/e/artificial...

EVENT: Join us for Artificial Intelligence and Democratic Freedoms on April 10-11 at @columbiauniversity.bsky.social & online. Hosted with Senior AI Advisor @sethlazar.org. Co-sponsored by the Knight Institute & @columbiaseas.bsky.social. Panel info in 🧵. RSVP: knightcolumbia.org/events/artif...

important 🚨

Many amoral 'useful geniuses' getting one-shotted by power-seeking Bad Guys. Hannah Arendt's 'banality of evil' is more relevant than ever, and it's sad how fragile values are and how quickly we can sleepwalk into dystopia. Here's to hoping that better angels of our nature win!

Sloan Foundation/UKRI postdoc on metascience and AI. For researchers in the social sciences & humanities who are interested in building a career in understanding the implications of AI for the science & research ecosystem. sloan.org/programs/dig...

Cool new reporting from @aylinwoodward.bsky.social on the ancient history of lead pollution - Roman silver smelting produced enough lead to lower population IQ by 3 points, potentially contributing to its decline. Hopefully we can eradicate the problem of lead exposure this century!

Yep I'm very close to leaving X now, not worth staying despite the better technical AI discourse there.

Some great news: the FDA just approved a new chikungunya vaccine for people aged 12 and older.

Metascience and AI news (part 1): we're launching a new metascience and AI fellowship programme, in partnership with @SloanFoundation.bsky.social to study and develop the uses of AI in research! Apply by 10/4; info webinar 28/2 www.ukri.org/opportunity/...

New Google DeepMind safety paper! LLM agents are coming – how do we stop them finding complex plans to hack the reward? Our method, MONA, prevents many such hacks, *even if* humans are unable to detect them! Inspired by myopic optimization but better performance – details in🧵

Is there a good starter pack for technical AI safety work?

Identity doesn't protect people from adopting fascist politics. Marginalisation isn't a corrective for political/moral position. Everyone is human, and unchecked power seeking is a common human failing.

Easy prediction for 2025 is that the gains in AI model capability will continue to grow much faster than (a) the vast majority of people’s understanding of what AI can do & (b) organizations’ ability to absorb the pace of change. Social change is much slower than technological change.

I like the vibes and people of Bsky, but the AI research and content is much better at X unfortunately. Therefore I must continue digging for gems through the trash there, despite the frequent vile vibes. 😔

since i'm always boosting bluesky, i'll say that one major weakness is that the qualitiy/quantity of AI discussion following the o3 announcement is still way better on twitter the only intellectual domain I care about where that's still very clearly true.

very unfair penguins can do this

Behavioral economics: people cannot add or subtract Macroeconomics: *maybe* people cannot deal with the fact that the entire cross sectional distribution is a state variable

I'm going to instruct AGI agents to go through everyone's skeets and create a human card for each user, showing their track record of wrong predictions (both explicit and implied). At the EoY, everyone will get their own Spotify Wrapped-style report of their top biases, errors, and inconsistencies.

new benchmark just dropped: if you sound like this, consider 'mitigating' your skill issues

MISR eval reveals frontier language agents can self-reason to modify their own settings & use tools, but they fail at harder tests involving social reasoning and knowledge seeking. Opaque reasoning remains difficult. arxiv.org/abs/2412.03904

Finally read this great anthropological study the hidden spiritual and cultural references within the latent space of models, by @bvlsingler.bsky.social & @mpshanahan.bsky.social. Puzzled by terms like 'hyperstition' or 'noosphere'? This is a great place to start!https://arxiv.org/abs/2411.13223

Texas has become the blueprint for the militarization of the US-Mexico border. This politically-driven show of force has sent the Rio Grande death toll soaring. We reveal for first time the true human cost of this

Did you know that 99% of email today is spam? Your inbox isn’t 99% spam because AI is used to filter it. The same 99% will happen here too, but if AI researchers continue to get perma-banned for making available the datasets needed to filter it, it’s going to make this platform unusable.

me post labour automation

Innovation job market papers round up! Over the last week, I found 50+ papers from PhDs going on the job market and I’ll post a few abstracts to this thread every day until they’re all up.

This banger from @sebk.bsky.social applies to OpenAI too x.com/sebkrier/sta...

The coming years will test whether Fukuyaman democracy can defend its values while avoiding the authoritarian drift Schmitt predicted. The solution lies in recognizing that liberal democracy isn't merely procedural neutrality, but a set of core values requiring active protection.

CoT can actually degrade outputs in cases where overthinking hurts humans too, and models share the same limitations. E.g. CoT worsened face recognition (language can't capture subtle visual details) but helped with complex decisions where memory was a critical prerequisite. arxiv.org/abs/2410.21333

Great post. While not strictly RL agents ofc, models like o1, trained with a greater emphasis on RL, can exhibit agent-like behaviors, optimizing for efficiency at the cost of interpretability of reasoning traces and exploiting flaws in goals. alexirpan.com/2024/12/04/l...

Compared to the 1920s, the average US dairy cow now yields around 6x more milk(!) In a great new article for Works in Progress, @pablohutch.bsky.social explains how big data, artificial insemination, cryopreservation, and more, created the modern dairy cow. worksinprogress.co/issue/how-bi...

🚀🚀PaliGemma 2 is our updated and improved PaliGemma release using the Gemma 2 models and providing new pre-trained checkpoints for the full cross product of {224px,448px,896px} resolutions and {3B,10B,28B} model sizes. 1/7

🪔 imma keep it real with you chief: genie 2 is far cooler than your language model that scores 76.29 on the questions-i-don't-care-about eval. shape rotate all you want, i'll be exploring a generated 3D interstellar ocean with a level 100 octopus-inspired humanoid.