Peeking my head out from social media hiatus to say that making AI safer by teaching it to disobey us is very, very silly. gift link: ter.li/ydoxnqeq @technologyreview.com
Arthur Holland Michel
Arthur Holland Michel
@writearthur.bsky.social
Friendly reminders about spooky technologies. Bylines in The Economist, The Atlantic, Wired, MIT Technology Review, etc. Author of 'Eyes in the Sky' Research on AI and its discontents. https://www.arthurhollandmichel.com/
387 followers774 following432 posts
I'm struggling to see the value-add for the Vatican of doing this. Having discussions with tech leaders is one thing. Bringing them onstage is another. Maybe the Vatican thinks it shows that it has buy-in from a key mover. But many observers will read this as corporate capture.
Pope Leo: “I want to thank Mr. Olah,” co-founder of Anthropic AI accept his invitation to walk together, to listen and to speak, and together to find a path for humanity in this time of artificial intelligence.”
Arthur Holland Michel reposted
In other AI news, prompt injections are here to stay based on this new paper from Abdelnabi and Bagdasarian. "an adversary can always construct a context under which a blocked flow appears legitimate, or a defender who tightens norms will block genuinely legitimate flows." arxiv.org/abs/2605.17634
It's good to see more people challenging the metaphor of the AI race. I wrote a report about this for @chathamhouse.org in 2023, and I think a lot of the findings still hold water today. www.chathamhouse.org/2023/04/reca...
The "AI race" with China is Washington's excuse to deregulate and skip safety. But AI knowledge spreads fast, racing accelerates rivals, and there's no finish line. From @arbel.bsky.social and @mtokson.bsky.social www.lawfaremedia.org/article/the-...
The fact that the AI hallucinated a quote from @merbroussard.bsky.social’s “Artificial Unintelligence” is truly perfection.
"The author of a nonfiction book about the effects of artificial intelligence on truth acknowledged on Monday that he had included numerous made-up or misattributed quotes concocted by A.I." 🫠🫠🫠
Look I know peer reviewers are super busy and they often don't get paid, but if you're not checking the citations then what you are doing is not peer review.
Two new facts stand out: 1. 85% of hallucinated citations in preprints are also in the subsequent journal version (thanks, peer review!) 2. Fake cites more likely to use the names of (male) scholars who are already highly cited, creating a fake-citation Matthew effect. arxiv.org/abs/2605.07723
Arthur Holland Michel reposted
Dozens of empty Waymos invaded an Atlanta neighborhood and circled a cul-de-sac for hours with no passengers www.wsbtv.com/news/local/a...
Modern warfare is weird. Baltic States are probably going to have to start installing interceptors to shoot down Ukrainian drones that get hijacked by Russia and diverted to targets inside their territory. I was on @deutschewelle.dw.com last night to discuss. www.youtube.com/watch?v=g7R7...
In case you needed any further proof of just how thoroughly drones have transformed warfare. Nobody is safe from aerial attack anymore, no matter where they are.
Russian Nuclear Submarines, Thousands of Miles From Ukraine, Get Anti-Drone Defenses New satelite imagery from @vantortech.bsky.social www.navalnews.com/naval-news/2...
For @economist.com I went deep on a big scary question: Could AI enable terrorists to develop bioweapons? The answer: yes, they could. But probably not quite yet.
How easily could a malicious person with no scientific expertise and an axe to grind create and spread a nasty pathogen? The bar is constantly being lowered
Arthur Holland Michel reposted
UK staff of Google's AI research lab hope to block the use of the company's artificial intelligence models in military settings. www.wired.com/story/google...
Arthur Holland Michel reposted
Arthur Holland Michel reposted
Except for the reference to ChatGPT, every single word in this story—from insisting that a breakthrough is imminent, to the glossy portrayal of a world where robots make coffee and stock supermarket shelves, and the buried lede on ethical questions—could have been written a decade ago.
Scientists say they've made a key breakthrough that would allow robots to figure out complex tasks on their own — but experts say it raises questions about how much risk comes with letting robots be in charge of their own learning. n.pr/4e2OlgF
This company started out making small unarmed robotic sailboats for environmental monitoring. Now it has unveiled the 170-foot, 250 ton "Spectre," which can be equipped with a range of missiles for anti-sub warfare and other attack roles. www.saildrone.com/news/introdu...
London's live facial recognition dragnet lives to see another day—and to see everyone's faces, all the time, all over the city. www.biometricupdate.com/202604/londo...
The U.S. spends around two million dollars to shoot down each $35,000 Iranian drone. Iran spends a few hundred thousand dollars shoot down each $30 million U.S. drone. dronexl.co/2026/04/20/2...
Everyone who was paying attention saw it coming. In some circles the drone cost asymmetry discussion has been raging for a decade.
Arthur Holland Michel reposted
Arthur Holland Michel reposted
People are having a great time dunking on this by saying things like "my car also goes faster than a human" but I doubt they'll be so smug when they're getting chased down an alley by one of these things for attending a protest.
The winner from Honor, a Chinese smartphone maker, completed the race in 50 minutes and 26 seconds, running faster than the human world record in a show of China's technological leaps. n.pr/4sP50Ij
This makes no sense.
Perhaps the darkest argument I’ve heard in favour of writing with AI is that there’s no point writing without it; we’ll all be consuming so much AI-generated content unwittingly that it will re-programme our writing styles and thought patterns. Humans who write without AI will sound like AI anyway.
Drone operators controlling drones in the battlefield from outside of Ukraine. Interesting questions here about individuals taking on active combatant status despite being far beyond an active war zone.
‼️ Incredible record: controlling interceptor STING at a distance of 2,000 km using HORNET VISION Ctrl While abroad, the pilot controlled interceptor STING operating in northern Ukraine. 1/2
I wrote about some previous studies on this last year, and what struck me is that while it is obvious that people who use AI to think for them will do less thinking, it is alarming just how drastic the effect is and how quickly these dependencies take hold. www.economist.com/science-and-...
NEW: A new pre-print study offers the first causal evidence that outsourcing reasoning tasks to AI can rapidly erode users' independent performance AND their will to persist despite difficulty. "People’s persistence drops... they’re also not willing to try without AI.” futurism.com/artificial-i...