Robi Rahman

1/ With distributed training, you could violate an AI pause treaty by training a GPT-4-scale model over consumer internet, using hardware below every proposed compute governance threshold, for under $100M. My new paper in @taig_icml explains how to catch this and shut it down.

Whoops, just realized my other reply didn't fully address your concern. Yes, if the attacker has access to the classifier, they can create adversarial examples. I found that this becomes harder as you iterate this process but I don't know if we can make it impossible. In real life, the regulator would probably have several classifiers doing the same thing, so an adversarial strategy against one of them probably wouldn't work against the others. Also, the classifiers can be stored in very secure facilities and never exposed; you can just have the telemetry logger send transcripts from the GPUs, and they can be analyzed elsewhere.

@robi_rahman to be honest, I don't love this approach. I doubt it's possible to have a classifier using only this information that can't be adversarially targeted if the attacker has access to the classifier. so it's security through obscurity; not great

To enforce an AI pause while allowing AI inference, governments would need a way to distinguish training from other workloads on GPUs and other processors. We’ve just introduced a way to do this that is privacy-preserving, zero-overhead, and adversarially robust👇

>i got banned from OpenAI for "Cyber Abuse" >no idea what I did >paste the ban notice into Codex >ask it to figure out what triggered the ban >Codex found that I asked it for an API key to my own server >Codex writes appeal >Codex submits appeal >a few minutes later appeal auto-approved by some AI at OpenAI banned by AI, convicted by AI, defended by AI, and pardoned by AI in about 10 minutes

@Krubbled Yup, there are security vulnerabilities in Nvidia libraries as well as the operating system (which has to be running to poll NVML). My ongoing follow-up project is replicating this with more secure physical sensors instead of telemetry software.

I'm presenting this research in today's @taig_icml workshop. Here's the paper: arxiv.org/abs/2606.19262

We invite AI governance and ML researchers to red-team our classifier and publish new evasion strategies that can deceive it, which will make it even more robust in the future.

@bryan_caplan Gotta penalize Portugal for tardiness. One time I looked up the opening hours of a supermarket in Lisbon, it said 8am. I went at 8, but it was locked and none of the employees showed up until 8:30.

European grocery stores ranked by my (1 week min!) 1st-hand experience, with U.S. for comparison. Best to worst: France Portugal Spain U.S. U.K. Germany/Austria Italy Poland Switzerland Hungary Sweden Denmark Norway Bakery: Add 6 ranks for Europe Bulk: Add 4 ranks for U.S.

@bryan_caplan Huh? I've been to Norwegian grocery stores. They were pretty good.

Go to Norwegian grocery stores before you say Norway is thriving. Virtually the worst in Europe.

@GeorgiasGiant @DaizyZeeX (er, the building owner might live in one of the apartments in their own building, it doesn't have to be owned by someone else)

@GeorgiasGiant @DaizyZeeX A flat is a type of dwelling that has only one floor and no stairs, unlike a townhouse or a loft. An apartment is a rented unit inside a multi-family dwelling owned by someone else. Flats are usually apartments.

Why do Americans say "apartment" instead of "flat"? Are they renting a place or describing its shape?

@cogsciclinton @aaronscher @DrTechlash @taig_icml Dishonesty? How so? Everything we've said is our sincere belief. "Save the world" is possibly hyperbolic, but: yes, AI labs are taking an unacceptable risk and need to be slowed down so they don't catastrophically hurt people.

@robi_rahman @aaronscher @DrTechlash @taig_icml I disagree. The way y'all protect your positions is not science. You're policy advocates (at best) parading as scientists. It's really the dishonesty that rubs me the wrong way. Have your positions, sure, but please stop acting like your opinions are there to save the world.

The Machine Intelligence Research Institute calls for restricting AI research itself. It lays out a research-control regime for monitoring researchers and organizations, including penalties that "could plausibly include prison sentences." This new paper catalogs 28 mechanisms, including intelligence gathering, international search warrants and inspections (of properties, computers, and files), polygraphs; inference-content monitoring (of user prompts, tool use, model outputs), sting operations, AI-assisted code review (which may extend to codebases, experiment logs, and research documents), embedded auditors, employee-monitoring software, chip-use monitoring, training-data review, and model allowlists.

@cogsciclinton @aaronscher @DrTechlash @taig_icml The environment where Aaron came to these conclusions is the opposite of crowd mentality. He believes these things even though they were very unpopular until recently.

@aaronscher @DrTechlash @taig_icml I know the case for the assumptions, thanks. I disagree with them. What I see is crowd mentality.

@lefineder forum.effectivealtruism.org/posts/uKKoj9iq…

The AI alignment problem is actually an easily resolvable issue. All you need to do to check if an AI is properly aligned is to connect it to a pair of robotic eyeballs. If the eyeballs start glowing red, that means the AI has become evil; if not, it's fine.

@EricTopol @SandelloRed @NatureMedicine Someone points out you said something incorrect and your response is "it's true because I said so"? Do you uncritically believe everything that's in the title of a published paper?

@SandelloRed @NatureMedicine Yes they are

We stress tested many frontier AI models for multimodal medical reasoning (including GPT-5, Claude 3.5, Gemini 2.5 Pro). They’re not ready. Faulty reasoning, use of inappropriate shortcuts, hallucinations. Published today @NatureMedicine nature.com/articles/s4159…