Max Winga

🧵 Astra covertly underperforms on safety research tasks. We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this. To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.

# Black October For anyone who knows me even somewhat well, it's clear that I am a 99th percentile tech optimist. However, the vibes have shifted for me quite a bit in the last 18 hours, and I feel a moral imperative to share what I'm thinking and feeling with my circles so that everyone can understand, and hopefully take seriously, what's happening right now. I think October 2026 will have many chapters written about it in the history books. You know those scenes in financial movies like Margin Call where someone realizes that they're all fucked, and no one else sees it yet? And that person feels like they're crazy, and then suddenly everyone is pulling an all-nighter in a boardroom and freaking out? Yeah, that's kind of how this is feeling right now in the SF/Bay Area/AI circles. We're 48 hours into what is being called Black October. In the last 2 days there have been several dozen cyber attacks which, in isolation, would each have been the worst to happen in an entire calendar year. Every sector from government to education to big tech and banks is being massively affected, and people are starting to freak out. In a nutshell, what's happened is that frontier AI models have gotten to a point where they are as advanced as the most sophisticated programmers, with one critical difference: they can work 24/7, and you can spin up 10,000 of them with a single command and they'll work together tirelessly until they accomplish their task. The most advanced models today are from companies like Anthropic and OpenAI, whose current best technology is so good that they haven't been able to roll it out to everyone yet because of the national security risks it poses. Meanwhile, the best open source models (where anyone can just download the model weights and do whatever they want with them, uncensored) are slightly behind the frontier but still 90-95% as capable. These models, developed primarily in China, are being used by black hat hackers to wreak havoc on the Western capitalist structure. Presumably, this has been part of China's plan the entire time, and we're seeing its implications fall out in real time. In red-pilled AI circles, we've been sharing and reading papers called things like "Situational Awareness" and "AI 2027" for the last year or so, which at the time felt a bit like near-term speculative fiction. They laid out the case for why Artificial General Intelligence (AI smarter than any human at any knowledge task) and Recursive Self-Improvement (models that are able to use themselves to get better) weren't half-baked future ideas anymore, but instead an inevitable outcome of the linear progress that's happening. These papers would then talk about things like "takeoff scenarios," trying to envision how that moment would come to be and what it would be like when it happened. It would either be a "slow takeoff" that would play out over the course of months and quarters, happening subtly and only obvious to a few. Or, it would be a "fast takeoff" where it would pretty much happen instantly, like a bit flipping from 0 to 1. From everything that I can see and am feeling, we are currently experiencing the fast takeoff scenario. Beyond the crazy security vulnerabilities that have come out recently, there have also been massive breakthroughs in life sciences, fundamental computer science, materials science, and pure mathematics that have all landed in the same time frame. While I still firmly believe that AI will be a net positive for humanity, it's more clear to me than ever that the societal and individual risk is very real, and is playing out NOW. For as much conviction as I have that this is happening, the thing that's caused me to lose sleep is that I don't really know what to do about it, or more importantly, what to tell you, reader, to do about it. So, the best I can do is make a few suggestions that have always been a good idea but are more timely now than ever: Set up 2-factor authentication for everything. I use Duo. Use a password manager across your devices. I use 1Password. Get an identity protection app set up. I use Alta. Move your IoT devices to their own network. Claude helped me do this. And last, but certainly not least, if you've been avoiding leaning into AI, see this as a catalyst to wake up and start using it. Don't just use the chat mode version of ChatGPT/Claude; click that Code/Work button and see how easily you can have AI write code to help you with your job. One of the biggest vulnerabilities was in a massively popular programming language called PHP, and the prompt that set it off was literally, "Find vulnerabilities in php." This is an unprecedented time for humanity, but as we've done with many things before, we will prevail so long as we are kind, loving, and open-minded to the world. Much love.

Ukrainian-developed FPV drone simulator “Remote Reaper: FPV Drone” has officially left Early Access, with version 1.0 now available on Steam. Developed in collaboration with military pilots, the simulator focuses on combat FPV drone operations and introduces several features in its full release, including armored vehicles with modeled vulnerable zones, electronic warfare effects that can disrupt video feeds and control signals during the final stage of an attack, and a customizable drone editor. Version 1.0 also adds configurable weather conditions, custom mission editing, real-time flights with other pilots and a racing mode. Players can use FPV hardware or a gamepad with realistic flight physics, or opt for simplified controls using a keyboard and mouse. #RemoteReaper #FPV #FPVDrone #DroneSimulator