Jonathan Stray

Hi! Ready your paper, very interesting. Thought about it, and I'm not sure I find the computational complexity proof convincing because it bites only for learning all possible distributions D -- human-like D is plausibly easier. I found this, makes argument more rigorously. ar…

Barriers to Complexity-Theoretic Proofs that "AGI" Using Machine Learning is Impossible arxiv.org

There's a long list of content-based signals we might want to rank on. We are limited by the classifiers we actually have. BUT we are working on LLM scoring with arbitrary prompts! The challenge is performance, but we have a cunning plan docs.google.com/document/d/1...

LLM-based scoring for GreenEarth docs.google.com

While there are many criticisms of LLMs I agree with, and correspondingly many lines of attack... this one I feel is a dead end. For two reasons. 1) there is deep theory showing that replicating the output of a system implies building a causal model of the system. E.g. arxiv.o…

Robust agents learn causal world models arxiv.org

Interesting! But there's no human testing here. "Quality" was evaluated only using automated metrics (BLUE, COMET) and while classifiers could tell human vs LLM apart, that doesn't tell us which one is preferred by humans. Meanwhile, www.nature.com/articles/s41...

AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably - Scientific Reports nature.com

Well, there are examples in the paper and the dataset is open source! github.com/humanCompati...

github.com

Our definition turns “neutral” into something empirically testable, generalizes to any conflict, and is grounded in political theory. And it really does find better answers that everyone can agree on. Preprint arxiv.org/abs/2605.28911 Dataset github.com/HumanCompati... /FIN

Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses arxiv.org

This may be the first widely deployed AI conflict mediation system. It's used by Chinese citizens in Hangzhou to resolve disputes with businesses. We're going to be seeing a lot more of this. www.sohu.com/a/932941980_...

AI调解员：国产大模型赋能消费纠纷化解，效率提升25% | 法律AI应用 sohu.com

You may remember we did a 10,000 person experiment testing AI-powered healthier feed algorithms. Well, now we’ve built it as a product for BlueSky/ATProto. We are currently recruiting pre-release users to test it out and give us feedback. Want to try it? survey.qualtrics.com/j…

Qualtrics Survey | Qualtrics Experience Management survey.qualtrics.com

Humanity's ability to know, reason, judge, and act well is the foundation of science, democracy, crisis response, & management of AI itself. AI poses serious risks to that foundation. New paper on epistemic risks by 30 experts calls for attention and proposes solutions. Link in thread.

Fusion power is gonna go a lot like AI. Empty promises for decades, then suddenly here faster than anyone can adjust to.

Me: I’m working on AI and human conflict BigLab employee: Cool, like multi-agent games? Me: No, like human conflict BigLab: Cool, so like simulating people’s responses? Me: no, like actual human field experiments 🤷‍♂️

What could it mean for an AI to be "politically neutral”? And can we measure it? New paper + dataset. We propose a definition that applies to any type of conflict on any topic: a neutral response should maximize approval on both sides of an issue, while keeping that approval balanced. 1/🧵

The AI models of today are the worst they will ever be. And yet, pretty much every "AI will never..." claim has now been shattered. I don't understand people who still bet against AI. How much more evidence do you need that these machines are going to be smarter than us in every way?

Seeing a flurry of evals and startups promising to test the mental health effects of AI. Literally all of them test what the model says in various conditions... none of them measure actual outcomes on actual people. A big gap, fixable with privacy-preserving experiments.

Is this a good logo for the GreenEarth feed? It's a healthier, user-controllable, open-source, LLM-powered, transparent feed we're building -- now in alpha testing. Try it? Tell us what you think! bsky.app/profile/did:...

I want to make sure AI doesn't incite human conflict. It's sometimes hard to explain what I do, but that's the core of it -- it won't happen automatically. And we're making progress! Both theoretically, and in field experiments that test how AI alters human relationships.

AI safety typically assumes one well-meaning user. I'm working on the case where two of them are at war.

Hello I am in Montreal for the week! Anyone interesting in AI safety I should meet here?