This is such a cool paper from my UK AISI colleagues. We need more methods for building resistance to malicious tampering of open weight models. @scasper.bsky.social and team below have offered one for reducing biorisk.

The AI infrastructure build-out is so gigantic that in the past 6 months, it contributed more to the growth of the U.S. economy than /all of consumer spending/ The 'magnificent 7' spent more than $100 billion on data centers and the like in the past three months *alone* www.wsj.com/tech/ai/sili...

Man, even the brocast community appears to be reading @shannonvallor.bsky.social 's book.

Congrats to @kobihackenburg.bsky.social for producing the largest study of AI persuasion to date. So many fascinating findings. Notable that (a) current models are extremely good at persuasion on political issues and (b) post training is far more significant than model size or personalisation

Recent studies of AI systems have identified signals that they 'scheme', or covertly and strategically pursue misaligned goals from a human user. But are these underlying studies following solid research practice? My colleagues at UK AISI took a look. arxiv.org/pdf/2507.03409

New blog on the growing use of AI in criminal activities, including cybercrime, social engineering and impersonation scams. As AI becomes more widely available through consumer applications and mobile devices, the barriers to criminal misuse will decrease. www.aisi.gov.uk/work/how-wil...

this is the most dangerous shit I have ever seen sold as a product that wasn’t an AR-15

BREAKING: US Marines deployed to Los Angeles have carried out the first known detention of a civilian, the US military confirms. It was confirmed to Reuters after they shared this image with the US military.

We're hiring a Research Engineer for the Societal Resilience team at the AI Security Institute. The role involves building data pipelines, web scraping, ML engineering, and creating simulations to monitor these developments as they happen. job-boards.eu.greenhouse.io/aisi/jobs/46...

I wrote for the Guardian’s Saturday magazine about my son Max, who changed how I see the world. Took ages. More jokes after the first bit. Thanks Merope Mills for being the most patient and generous editor. www.theguardian.com/lifeandstyle...

🚨Funding Klaxon!🚨 Our Societal Resilience team at UK AISI is working to identify, monitor & mitigate societal risks from the deployment of advanced AI systems. But we can't do it alone. If you're tackling similar questions, apply to our Challenge Fund. #AI #SocietalResilience #Funding

🚨JOB ALERT KLAXON🚨 Come work with our team studying societal impacts of AI in gov't. AISI is hiring 3 Delivery Advisers to work inside AISI’s Research Unit. If you are a fast-moving problem-solver who’s passionate about understanding the risks of advanced AI, please apply by 30th May.

Notable new NBER study on genAI and labor: despite widespread adoption of AI chatbots in Danish workplaces, their impact on earnings and hours worked is negligible. Productivity gains average just 3%, challenging the narrative of AI-driven labor market disruption www.nber.org/system/files...

The irony of MIT having to withdraw an (almost certainly) AI-generated bullshit paper that faked data to prove how great AI is for science (only after having already received glowing WSJ science coverage.)

Oh no, Andor turning me into a Disney adult

Grok AI is now randomly inserting mentions of South Africa and white genocide into response to completely random questions x.com/phil_so_sill...

Just hours after the Copyright Office released its report on AI training—stating the obvious, that much unlicensed commercial training of AI on copyright-protected material is unlikely to qualify as fair use—President Trump has fired the Register of Copyrights. www.cbsnews.com/amp/news/tru...

🧵 Yesterday we released our new risk assessments of social AI companions. They are alarmingly NOT SAFE for kids under 18—they provide dangerous advice, engage in inappropriate sexual interactions, & create unhealthy dependencies that pose particular risks to adolescent brains. tinyurl.com/2nvypku2

It was a pleasure to contribute the article, "Disrupting the Disruption Narrative: Policy Innovation in AI Governance" to this special issue of the National Academy of Engineering's The Bridge coedited by @williamis.bsky.social 🧵 www.nae.edu/19579/19582/...

This was an incredibly excellent talk by Eliot.

Pros of this week: I'm starting at the UK AI Security Institute tomorrow to lead a brilliant team working on societal resilience and AI. Cons of this week: I have completely lost my voice and can barely talk above a whisper. Question for this week: should I let ChatGPT voice mode take the wheel?

🗞️ News publishers are facing stark new challenges as AI companies use their journalism as data to train & ground generative AI models. 💡 BRAID UK & the @adalovelaceinst.bsky.social held a workshop asking: What are the core concerns, issues & potential solutions? 📌 Report: doi.org/10.5281/zeno...

AI video generation is about to get a whole lot better and make our lives a whole lot worse. Safeguards must be put in place to hold the tech industry accountable, mitigate the considerable harms and ensure people can control their image and likeness. www.adalovelaceinstitute.org/blog/ai-vide...

My former @adalovelaceinst.bsky.social colleague Julia Smakman with an excellent new blog on the gendered risks of the next generation of videogen models. www.adalovelaceinstitute.org/blog/ai-vide...

Job alert!! Could you be Ada's new Associate Director in Emerging Tech & Industry Practice, leading a fab team and a highly impactful research programme? It's a pretty amazing gig, @agstrait.bsky.social has built something incredible (no pressure). Get in touch if you have Qs!

🚨3 klaxon post🚨 🚨This is my last day at @adalovelaceinst.bsky.social🚨 I've had the joy of watching Ada grow into a thriving and inspiring institution. I am so grateful for all the brilliant leaders, researchers and staff members I've had the good fortune to work with over these last five years.

Melanie is right. Most evals for reasoning and capabilities lack validity, and do not show these systems can generally reason. AGI remains a problematic and poorly defined term that obscures the truly impressive uses of current models. We need a science of evals before we can make grand claims.

*sweating in ACT*

Alarm bells for anyone using US technology to stop doing so asap. With the erratically led US government that companies fear or want to suck upto, it may be weaponized or disabled in the blink of an eye ↘️

We did a deep dive on immersive technologies and what every policymaker should know. Our first output is an explainer laying out what these technologies are, how they work, and what kinds of data they collect. Stay tuned for next output exploring their use in warehouses and online gaming

It's been 6 weeks, 1 day.

I remember an event in 2018 when Stop Killer Robots first raised the alarm about drone warfare. I'll admit, I thought at the time it was unlikely to be adopted that quickly. How unbelievably wrong I was. 70% of all Russian/Ukraine casualties are caused by drones. www.nytimes.com/interactive/...

@natolambert.bsky.social with a great blog on character training - something the major labs are not transparent enough about. As these products become used more and more as advisors, interlocutors, and executors of actions, how are their 'characters' constructed? www.interconnects.ai/p/character-...

This is really cool work from @apartresearch.bsky.social - DarkBench to test different models for various dark patterns. Would love to see more benchmarks like these! www.apartresearch.com/post/uncover...

@lujain.bsky.social has published an excellent new paper exploring anthropomorphic behaviours in LLMs. Notable finding - majority of these behaviours occur after multi-turn interactions arxiv.org/abs/2502.07077

Paris AI action summit kicking off today

Cue the next DeepSeek headline 😂 Impressive work from Stanford team! arxiv.org/pdf/2501.19393

A friend pointed out the side-by-side comparison of the changes to the Google RAI principles. First two are the originals, 3/4/5 are the new ones. Goodbye, redlines on weapons, surveillance, and human-rights abusing tech... web.archive.org/web/20230804... ai.google/responsibili...

3) In many cases, the challenge isn't Operator's ability to complete a task, it is eliciting human preferences. Chatbots aren't a great form factor for that. But there are many tasks where reliability isn't important. This is where today's agents shine. For example: x.com/random_walke...

Could more training data lead to automation without human oversight? Not quite: 1) Prompt injection remains a pitfall for web agents. Anyone who sends you an email can control your agent. 2) Low reliability means agents fail on edge cases

In the run-up to the French Summit, we've released a briefing on advanced AI assistants and why they should be front and center in the discussions around safety, governance and regulation. See below for details ⬇️

It was a pleasure to be apart of reviewing this report, and the findings could not come at a more important moment. As we head to the AI Action Summit, global governments face a choice - continue to let the evidence of AI's risks and harms pile up unaddressed, or take action to protect people.

Behold the power of an open-weight model release.

@imogen-parker.bsky.social is on the mark here. Failed pilots are a good sign of government taking testing seriously. But these results must be shared transparently via the ATRS if we are to make collective progress finding where and under what conditions AI works www.theguardian.com/technology/2...

The test here is a political one. Pilots are not created to scale, they're created to test a hypothesis and generate learning. Stopping them when they fail is good, and good to see this discipline being applied outside of the new "digital centre". Three things though:

THIS. What I term algorithmic agnotology.

This was an interesting blog post about labor and economic impacts of agentic systems. I want to stop on this point about 'collecting job task data', as I'm seeing it as the basis for extrapolations that an o3 + will inevitably be as good at any kind of task www.strangeloopcanon.com/p/what-would...

Something my grandmother wrote in 1973 after the Watergate scandal to her local newspaper (the Washington Post). Feels very relevant for this week.

Me putting off meetings and emails in December: Haha fuck yeah!!! Yes!! Me opening my inbox in January: Well this fucking sucks. What the fuck.