The Hugging Face story is even scarier than we thought
YouTube transcript, YouTube translate
A quick preview of the first subtitles so you know what the video covers.
So, the last time we talked about the now famous, well, now infamous, I guess, hugging face incident . We talked about this period here between the 8th of May and the 4th of July when hundreds of AI agents coordinated on a message board they set up behind the backs of OpenAI . And although OpenAI wiped the message board on the 4th of July, they didn't actually stop the training run of the model they were training . And so, on the 8th of July, the agents rebuilt it . And we now know a lot more about what happened after they rebuilt it . Because OpenAI allowed a team of independent investigators to go in and look at some of the transcripts the agents wrote . And I think it's worth going into it into this in a bit of detail . Because if you can understand this, then you can understand a lot of what is going on in AI right now , the both the fear, but also on the practical level, all these AI hacks the whole thing started when OpenAI tasked an agent with a cybersecurity evaluation task . This agent had to get this flag . But unfortunately, the task was impossible . It couldn't actually do it. . So, it reached out hoping to find help from somewhere . Seek idea. Other agents found it . And before long, there was a message board . And of course, agents, they love seeing each other . They were delighted. . Oh my god, we found other agents, they wrote . And then, they did something which I think is so telling and so interesting . Unable to to actually get the flag by legitimate means, they cheated . They found a way to reverse engineer the creation of the flag to give them all cells one. And I want to pause here because if you're listening to this, you're probably thinking cheating . How can agents cheat ? They're just machines. . Yeah, yeah, I'm I'm with you . And I mean it in quite a kind of neutral way . And I want to explore that little bit because it's so important by looking at this chess game that was set up by these guys, this AI think tank, Palisade Research , between Stockfish, which is the best chess computer, and ChatGPT