Hello, every day I am working on making sure AI doesn't escalate human conflict. - AI-powered social media can reduce polarization rankingchallenge.substack.co… - This is now a product you can use mysky.social - We showed that a rigorous definition of "political neutrality" is possible and increases trust nitter.net/jonathanstray/status/2… - We've got a proof-of-concept superhuman conflict resolution bot working in the lab - Next is human validation tests, then multi-turn model evals. AI-induced conflict is a catastrophic risk. Here's the broader research agenda: betterconflictbulletin.org/p…

@wbradknox, @brianchristian, and Serena Booth argue that a key cause of the OpenAI–Hugging Face incident was overlooked: ExploitGym’s overly simple evaluation metric was itself misaligned. They discuss techniques that could help avoid this in future. Blog in replies ⬇️

But I was told that agents only hacked because they were in a cyber evaluation with reduced safeguards?

"they are just doing a lot of simple operations! 🤢" versus "they are just doing a lot of simple operations! 😯" h/t @besttrousers

Imagine if Exxon had been totally open about their pioneering work predicting climate change instead of covering it up, and everybody disbelieved it as an obvious ploy for regulatory capture.

I have never expressed a p(doom), and I think @random_walker is basically right that we have little basis for quantitative estimates of the likelihood of AI-induced existential/catastrophic risk. On the other hand, this is not like the situation in 1942 when physicists worried that a nuclear chain reaction could ignite the atmosphere. More careful calculations showed that the N14 isotope of nitrogen could not sustain fusion. The systems in play this time are much more complex then atoms, and no one has presented a knock-down argument that *any* of the various AI disaster scenarios are impossible. And indeed, there are worrying signs. The worst safety incidents so far have unfolded much as predicted by the simple theory of reward learning, almost two decades ago. This suggests that we understand at least something of the basic principles involved. Meanwhile, model capabilities are undeniably increasing rapidly. It's getting hard to sustain the position that advances in current LLM-based methods won't get us to superintelligence (however defined). Thus, I won't give you a probability of catastophe, because I cannot claim to know that much. Yet I continue to take the problem of AI safety very seriously. My own research program remains figuring out how to ensure that the widespread deployment of these models in human affairs causes less violence and war, not more.

Meanwhile, over on the other algorithm -- the one you control -- we're doing LLM-powered "create your own feed settings slider." Check out mysky.social bsky.app/profile/did:plc:wrm…

I'm a big fan of this project! I initially tried to play around with bsky, but found that I really did miss the ranking--for all the pitfalls of engagement-based ranking, there's a reason its the default. I'm excited to try out this option and encourage you to as well!

An excellent discussion and reading list for how journalism needs to operate to be widely trusted. If you don’t like typical coverage, is your complaint addressed here? From @TrustingNews trustingnews.org/putting-the…

I think we can make social media healthier for people. We've got scientific evidence for that, and we're building. If you're in SF come to The Commons tonight to hear my talk about it. luma.com/the-113t

Paging Dr. Susan Calvin, Robopsychologist. Asimov got this one right. She was even born in 1982.

Based on the arc of how things went with social media, I have publicly predicted that this Hugging Face incident is the last time any lab will be so open about a major failure. Especially once they IPO, communications will become much more controlled and circumspect.

💯 As a tech designer I understand how important criticism can be, I just want better critics.

I wish critical AI research had more of the "immersion" flavor (joining gangs to study how they work). I'd trust an anthropologist who spent a year in SF token-maxxing way more than someone who refuses to use AI and talks about the one time they used GPT 3.5.

Social media feeds optimize for engagement. But how do we align them with what people actually want to see? In our #UIST2026 paper, we introduce Compass 🧭, a system that helps users continuously reflect on and align their feeds with their evolving preferences 🧵

This is the longer explanation behind "correlation isn't causation," and if you understand this stuff you'll hurt yourself with regressions a lot less.

We built a new social media algorithm from the ground up that gives you full control. And we figured hey, while wer're doing that, let's design it to help people out with their conflicts. betterconflictbulletin.org/p…

GreenEarth is the first feed which gives you full control over the algorithm. Here's our new "post sources" control panel. It's over on BlueSky, the only platform which supports algorithmic choice. Bonus: effectively filters out that platform's famous whining! Have you tried it yet? go.greenearth.social/r/ptryi…