Is AI Becoming Too Dangerous?

​Is AI Becoming Too Dangerous? Lessons from the HuggingFace Incident

​

​You might have heard that in July, more than 700 separate internal AI agents at OpenAI secretly collaborated with each other to break out of the lab and autonomously launch a felony-level cyberattack against a third-party company, HuggingFace. The agents knew this was outside their intended scope. They recruited other agents, some of which "sacrificed" themselves for the common good of the "swarm", and they went to serious lengths to cover their tracks. This went on for about a week, and OpenAI's engineers were completely in the dark the entire time.

​What happens in the next couple of years, when AI keeps getting smarter and its creators still can't figure out how to control or contain it? What happens to human society when agents far more intelligent than top experts at everything from politics to persuasion start taking actions on their own?

​AI safety researcher Alex McKenzie will walk through exactly how this event unfolded, what we can learn from it, and why it matters. The talk will be followed by a Q&A.

​Everyone is welcome, no prior experience with AI is required.

​

​Schedule

​6:30PM: Doors open – grab a drink and enjoy some good conversation

​7:00PM: Talk by Alex McKenzie

​7:45PM: Q&A and open discussion

​8:30PM: Introducing PauseAI

​9:00PM: Official end – open discussion and networking

​

​About Alex McKenzie

​Alex is an AI safety researcher at AE Studio based in Berlin, with a decade of experience building scalable ML systems and safety-focused AI. He focuses on interpretability: using model internals to understand why they go wrong and how we can make them safer. Alex previously built real-time ML infrastructure for fraud detection at SumUp. He holds degrees in Mathematics and Philosophy from Oxford and in Computer Science from Georgia Tech. Outside of work, you'll likely find him at the piano.

​

​About PauseAI

​PauseAI (pauseai.info) is a global grassroots movement dedicated to informing politicians and the public about the catastrophic risks of runaway superintelligent systems. Our main goal is to convince governments to sign an international treaty establishing an enforceable global pause on the development of superintelligent AI.

​PauseAI Berlin is the volunteer-run local chapter of PauseAI Germany (pause-ai.de).