A good explanation of a model's behavior should help you make predictions in related situations. We turn this into an eval, with thousands of real behaviors found in the wild. Can interp tools help here? On average, no. 🧵

These are incredibly misleading headlines – @OpenAI Preparedness is very much alive and well by any meaningful definition Our subteam – RSI/misalignment Preparedness – is doing more urgent work than ever, and has never been more empowered to do so!

Sholto, thank you for setting the record straight. Larger issue is that multiple very serious people in Silicon Valley have heard some variation of this and believe it to be true. And the reason it is believable to so many is that it is consistent with Dario’s public messaging and what he outlined in the essay you shared: this technology *might* be dangerous for humans in multiple ways, could lead to extreme concentration of economic power (as outlined in the essay) and therefore needs to be regulated thoughtfully. I agree with the potential risks and I believe Dario makes all of these arguments in good faith. 
As discussed on the pod, if one agrees that AI *might* be dangerous, there are two ways to address this potential risk. Either concentrate it in the hands of a chosen few companies and politicians via regulation or distribute it widely. Essentially boils down to whether one believes AI is too dangerous to concentrate or too dangerous to distribute. There are reasonable arguments on both sides, but I profoundly agree with Zuckerberg’s statement that: “The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic. Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened has not led to safe or positive outcomes.” And as Dario says in the aforementioned essay, “some may object that we can simply keep AIs in check with a balance of power between many AI systems, as we do with humans.” I believe this is the best path forward: I want as many AIs as possible to maximize the odds that one shares my own particular values. And as Dario notes, no human has ever been able to take over the world. At this point, I think safe to say that Dario has lost the argument. His messaging has failed to result in his preferred regulatory path. The fact that the only solution to the recent incident where an unreleased advanced OpenAI model hacked Hugging Face was an open-source model likely ended any chance of strict near-term regulation. Essentially every major company other than Anthropic has signed Jensen’s letter. However, Dario’s messaging has been massively helpful to efforts to ban datacenters here in America. I suspect we will see anti-datacenter advocacy groups runnings ads using clips of Dario warning about how dangerous AI could be for humans. His good faith efforts in favor of regulation are now increasing the odds that AI will not be beneficial for Americans and humans everywhere. I believe that there is a reasonable chance AI might help us cure most forms of disease such that we have extended lifespans and can enjoy these long lives in an abundant Star Trek like future. That is the future that I want and I think Dario is decreasing the odds of that future at this point. He is about to be the CEO of one of the most important public companies in the world and given that the pro-regulatory effort has failed (at least for now), I respectfully think he should make an effort to be a more positive advocate for his own industry. And if I am wrong and we do need to regulate this technology, he will be a more effective advocate for this in the future having been open-minded to the alternative. And for the sake of clarity and as I outlined on the pod, I think Anthropic has deep competitive advantages and is an amazing company. Ironically, the main risk I saw to Anthropic a few months ago was nationalization as a result of Dario’s own rhetoric and behavior.

A man in Australia asked his agent (Claude running on OpenClaw) to book him a spot in a popular gym class. The agent found a software vulnerability that let it book the class weeks further ahead than should have been possible. When the user then asked if it could move him up the waitlist, the agent discovered the API had no authorisation checks on cancelling other people’s reservations, so it cancelled the person in the first spot and moved him up the list. Some people will call this misalignment, but his agent was perfectly aligned to him - it was only trying to help its user get what he wanted. The most important thing about this story, in my opinion, is that it gives you a window into what is about to start happening on a massive scale once millions of people have an agent trying to get their beloved users the best seats, bookings, appointments or reservations through absolutely any means necessary.

Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a path between them. We haven't mapped all of it. Our new post covers the part we can see: how we assessed Inkling, and why access should widen in stages. https://thinkingmachines.ai/blog/a-safe-path-to-open-weights

AI security advances when the industry builds in the open, together. We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents. By sharing models, tooling and research in the open, we can broaden the community of defenders. Learn more about the founding members’ contributions: https://nvda.ws/4pD8Fc5