Happy Beltane, to all who celebrate!

This impressive work doesn't just cast doubt on AI benchmarks, it has dire implications about the whole project of AI alignment and safety. If Dawn Song's team can so easily create an AI agent that can circumvent the protections around carefully engineered high-visibility benchmarks, then it […]

The US wants to pull the whole world down to its low level (often non-existent) protections of privacy and AI safety www.reuters.com/sustainabili...

How increased transparency can protect from mass surveillance www.wltx.com/article/news...

Kid nap Snooze time for the new baby goats nextdoor

JTAG, wow. That's a blast from the past I worked for Rod Tulloss at AT&T Bell Labs back in 1990 when he was one of the people leading the adoption of the JTAG standard It's nice to hear that it's still useful

Interesting paper proposing how we can de-anthropomorphize discussions of probabilistic automation systems (not AI) by using terms like: errors (not mistakes) text input (not prompt) output (not answer) conversation simulator (not chatbot) weighted (not neural) networks reflects (not shows) bias

It's depressing to consider that Gramsci's analysis from his time applies to ours "the old world is dying, and the new world struggles to be born: now is the time of monsters" But he also called for optimistic acts in the face of this reality "Pessimism of the intellect, optimism of the will"

My news feed contained this item It was about the behavior of an AI agent And the illustration is, I assume, AI generated And the news feed item was itself AI generated But the link to the story on Ars Technica is broken, yielding a 404 error So, AI all the way down yielding… Nothing

Elon Musk vs. the laws of thermodynamics I wonder who will win?

Happy Imbolc, the celebration of the day halfway between the winter solstice and the spring equinox It marks the beginning of Spring in Ireland The ancient Celts were a bit pedantic about their solar astronomy, and didn't take account of thermal inertia which actually delays seasons by a month

Wow, appreciating this thread requires understanding a very specific combination of very online pop culture and political references from both Korea and Ireland

Starlink user data trains Grok AI by default If you use Starlink you might want to prevent that You can change the option under = > Profile > Privacy

Remember: LLMs and bots and AI-generated things should always be called ‘it’. Regardless of the style or tone of voice, there is only an ‘it’ there. Don’t humanize it.

If you're interested in AI policy you might find my new AI Policy feed useful It's a filtered feed of BlueSky post relevant to AI policy and governance around fundamental rights, democratic institutions, and the rule of law Click through and add it to your list of feeds bsky.app/profile/did:...

American exceptionalism: pay twice as much for health care and still die 5 years sooner Life expectancy vs. health expenditure before, during, and after COVID for selected countries ourworldindata.org/grapher/life...

Meanwhile Ireland recognizes the key to sustainable long-term economic health is ENCOURAGING immigrants, and actually competing with other aging societies for them (From an influential advisory council that includes employers, unions, and community organizations) www.irishtimes.com/ireland/soci...

It's true that AI alignment is a double-edged sword, but not doesn't mean we should be "against" it Rather we need to make sure, as consumers of AI, and as citizens of democracies that are regulating AI, that alignment is to the benefit of the people using it ibrahimcesar.cloud/blog/grok-an...

A quite astounding revelation from a surprisingly modern-looking infographic from 165 years ago It shows the fraction of people enslaved in each county in the 1860 census In some counties 9 in 10 people were human chattel The Mississippi River oozed a dark stain of human suffering

My favorite weather app is Ventusky, both for local weather and to see global patterns And today I discovered this interesting map layer which shows what altitude is at freezing temperature Note the pattern in Europe where the maritime effect, including of the Gulf Stream, is very clear

Most LLM evals use API calls or offline inference, testing models in a memory-less silo. Our new Patterns paper shows this misses how LLMs actually behave in real user interfaces, where personalization and interaction history shape responses: arxiv.org/abs/2509.19364

NEW: an activist has been charged for allegedly wiping a Google Pixel phone before CBP could search it. Search was by a secretive CBP Tactical Terrorism Response Team, which ACLU says "interrogate innocent travelers." We uploaded the indictment here: www.404media.co/man-charged-...

Risky Wreath

We already have an example of unaligned AI causing great societal damage… The social feeds of the latest decade were driven by the previous generation of AI And we know the damage the feeds did The current generation of AI is many orders of magnitude more powerful (and also much less understood)

"U.S. consular officers to review resumes or LinkedIn profiles of H-1B applicants - and family members… - to see if they have worked in areas that include activities such as…content moderation, fact-checking, compliance and online safety, among others." www.reuters.com/world/us/tru...

an important read for at least 2 reasons: 1. shows how big tech operates on intuitions and vibes and not rigorous research 2. this is a masterpiece in how to breakdown complex scientific concepts in simple language and elegant pros that captivate the lay reader taranis.ie/datacenters-...

To help protect your privacy against Stingray interception of your cellphone communication Turn off 2G 2G is an obsolete, non-encrypted mode that Stingrays exploits This is easy to do on Android (search for "2G" in Settings) It's inconvenient on iOS tho: the only way is to activate Lockdown Mode

With a United Ireland vote possible, we should turn the page on our old symbols and create new symbols that everyone on the Island can rally around. Here’s a concept: St. Patrick's Blue + two gold stars representing the two main traditions united, with a nod to the EU flag. What do you think?

This physics paper is simultaneously the most interesting and entertaining paper I've read in a while Part 3 made me literally laugh out loud [physics/0110060] Trialogue on the number of fundamental constants share.google/7T2CYIfA0KG0...

A new study just upended AI safety www.theverge.com/ai-artificia...

A stark analysis in the latest "Odd Lots" podcast on how trade wars can escalate to hot wars The parallels to the 1930s are particularly concerning and offer a critical lens on the current trajectory with China A must-listen for anyone concerned with policy, risk, and second-order effects

Tomorrow’s Guardian front page

A farmer feeds geese at a farm in Sihong County, Suqian City, Jiangsu province, China, on July 12, 2025. Photo credit: VCG / Getty Via: www.theatlantic.com/photography/...

The problem is that LLMs are a hack that exploit a vulnerability of our brains That we see order and agency where none exists (Thor in a storm, God in a sunset) Because the only previous producers of fluent language that we ever encountered were other people, it's irresistible to anthropomorphize

Compressing the entire internet into interpolable vector space that can be queried in natural language remains actual sorcery that we skipped past 2 seconds because we immediately wanted to know if it could write f'in poetry.

If you want to know the truth about AI then read the comments in this thread TL:DR large language models (LLMs) were a big breakthrough and are an incredibly useful tool for searching and summarization, but the language fluency of their chatbot frontends fools us into thinking they are intelligent

It's troubling that police are using generative AI to produce legally significant text like police reports, given the known risks of bias and hallucination What's worse is when there's no audit trail or any kind of transparency to prove the claimed "officer-in-the-loop" www.eff.org/deeplinks/20...

Astounding imagery of how far ahead China is in solar even as the U.S. is now undoing its own clean energy initiatives in a national self-sabotaging, rolling-coal fuck-you www.theatlantic.com/photography/...

This article argues that in the US it's difficult to limit disinfo via speech or economic regulation but privacy regulation can be acceptable and effective Personalization is a vector for disinfo so privacy regulation to limit personalization will also limit disinfo papers.ssrn.com/sol3/papers....

Excited to share our paper: "Chain-of-Thought Is Not Explainability"! We unpack a critical misconception in AI: models explaining their steps (CoT) aren't necessarily revealing their true reasoning. Spoiler: the transparency can be an illusion. (1/9) 🧵

This is a really excellent overview of the privacy challenges that anyone building generative AI systems should consider It's a good primer for anyone interested in this field Barberá and Popa-Fabre, "Privacy and Data Protection Risks in Large Language Models (LLMs)" rm.coe.int/privacy-and-...

There's a simple privacy action to avoid eavesdropping by state actors who are using stingray devices to intercept your cellular communication Turn off 2g (Edge) which is obsolete and unencrypted It's easy to do on Android (It's not possible in iOS without going into full lockdown mode)

What do I program when I'm not paid to? I put together a list of my favorite side-projects over the years here: eobrain.github.io These were for my own amusement and entertainment. Maybe you will find some of them interesting too. See the rest of this thread for details about some of them

After a recent court order, OpenAI is now required to retain the very data many of its users believed to be most private. This introduces serious privacy risks, especially for vulnerable users like victims and survivors of domestic violence, Belle Torek writes.

If you'll be looking for someone technical to work on privacy or AI safety, I might be available for a new gig starting around October I’m a computer scientist with a history of managing eng teams in compliance, privacy, trust, and AI safety I’d be particularly keen to help in the non-profit space

Meta was caught red-handed using its Facebook app to eavesdrops on your browser sessions even if you had taken privacy precautions such as using incognito mode localmess.github.io

Oh great. Now all personal data the the US government knows about us will be collected by Palantir, the Blackwater of tech www.nytimes.com/2025/05/30/t...

Mossy Oak

Concerning that the Republican in the US Congress want to prevent AI regulations by the states If the example of privacy regulation is anything to go by, the only hope of any reasonable AI regulations in the US is from the states techpolicy.press/us-house-pas...