Joseph Miller

Had a lot of fun editing this. So many good clips to choose from.

Please invite your friends!

PauseAI the Movie 🍿

Just completed my application to PauseCon DC 🤞

Some AI CEOs in fact do not want to risk the fate of the world on an insane race to superintelligence.

I gave a talk at my college in Oxford about why I'm trying to PauseAI.

TikTok is the Mirror of Erised. The feed of a perfectly happy person would just be a black screen that reflects their own image.

Today, we’re launching SPEAK. Every year, thousands of people in the UK are arrested for what they say online. We’re going to change that. 🎥

Overheard in Oxford: "I didn’t ask him to start buying me charcuterie, but he consistently does... So I have to buy bread."

Although on first attempt the model seems a bit worse when answering questions about a video.

Kinda crazy that Gemini 3 has native video input. Idk why they're not advertising that. Seems like a very useful feature potentially.

Note to self: If you're putting off a task, add it to your todo list, even if it's extremely tiny and trivial (like replying to a message), so you get some satisfaction from completing it.

@LRudL_ I have similar feeling about "research taste".

@LRudL_ Is big model smell a real thing? It's feels a like analyzing abstract modern art imo. I'm not sure if there are real things being tracked or people are just trying to show how sophisticated they are.

What're top benchmarks for "big model smell", NOT the ability to grind through easily-verifiable, closed-form maths/code/reasoning puzzles? I want a model with creativity, writing chops, good at adapting to a worldview. Intellectual gravitas, not empty performative shape rotation

@Algon_33 @MichaelTrazzi See AI Lab Watch: ailabwatch.org

@MichaelTrazzi TY. Do you know of any other labs breaking safety pledges? IIRC, anthropic changed some of their RSPs but that's (maybe?) not the same thing as breaking a pledge.

After 12 days of hunger strike, DeepMind gave their first public statement regarding the strikes to The Verge, saying: "Safety, security and responsible governance are and have always been top priorities as we build a future where people benefit from our technology while being protected from risk." This totally fails to address the breathtaking recklessness of the race to the bottom that Google DeepMind is in. In reality, safety cannot be a priority, since they are in a relentless competition with other AI companies to build superintelligence first. DeepMind couldn’t even wait a couple of weeks to take basic precautions before releasing Gemini 2.5 Pro, violating the Frontier AI Safety Commitments they made in Seoul last year. That is the point of the Google DeepMind hunger strike. We want AI CEOs to acknowledge the race they are in and the danger this poses to humanity. They would then be able to coordinate to slow down frontier AI development before it’s too late. Denys Sheremet (on his 11th day without eating) and I are still waiting for a public statement from Google DeepMind CEO @demishassabis, as requested in our letter forwarded directly to him one week ago.

@NinaPanickssery If you were aligned with IGF you would want to spread your particular genome right? And on the current margin it would be a very good idea for people to have more children to increase their IGF, even if everyone doing this would decrease the population in the long run.

@MichaelTrazzi You should be proud. Not many people would have the guts to do what you did. Stay safe.

I've just received a medical report from two doctors who have examined my blood tests, and upon reading their findings, I have decided to stop my hunger strike. They strongly advised me to end my fast given developing acidosis and “very low blood glucose” (even for somebody on a 7 day fast) that "runs the risk of hypoglycemia which can be life threatening". I had decided to see a doctor today after experiencing two near-fainting episodes over the past 48 hours. Given the significant risk of seizure, brain damage, or death if I continue, I am stopping my fast today and beginning refeeding. I want to thank everyone who has supported me in this journey, both in person and online. My hunger strike ends now but the movement continues, with Guido and Denys still on strike in front of Anthropic and DeepMind. The emergency remains real. We need to build coordination between AI companies now before it's too late. When we reach the critical point where AI systems become powerful enough to pose catastrophic risks, we must have the mechanisms in place to pause frontier development and focus on safety. My letter to Demis Hassabis has been delivered directly to him, and I'm still waiting for his response to this first step: publicly commit that DeepMind would halt frontier AI development if other major labs agreed to do the same.

@NinaPanickssery I think this definition still leaves room for a lot of crucially different interpretations. For example, does moral realism mean that you should be able to physically construct a good-o-meter? I don't think moral realists would have a consensus answer.

@JosephMiller_ Do you dislike this definition?

It grates on me when I open yet another blogpost or essay that starts discussing “human values” without even a nod to whether or not such a construct is justified.

@NinaPanickssery I think moral realism does not have a widely agreed definition / shared conception. Nor does morality, which is why I think it's a confused concept. I'm sure there are some versions of moral realism which are a coherent position. How are you defining it?

@JosephMiller_ Moral realism is a coherent position, and I think holders of it should be more, and not less, direct about that, instead of smuggling it in as an implicit assumption.

@NinaPanickssery Yeah I was being a bit facetious. I agree those words do convey some meaning. I just favor anti-realism and I find morality to be a confused concept, and think people would be less confused if they talked about preferences instead of morality.

@JosephMiller_ It’s fine to discuss “human values” if you start off by saying that what you mean is the objective moral good, or what most people value empirically. Of course ideally you say those things directly. If you think these categories overlap, you should also explicitly say so.