Over the weekend, I watched with some more insights into the mid-July Hugging Face incident. OpenAI alignment and safety researcher Eric Wallace and OpenAI infrastructure and security engineer Michael Dalton described “the most qualitatively interesting example of AI capabilities” they had ever seen and this is, in my opinion, no understatement. The whole talk is an interesting glimpse into the way the company trains their models and hones their capabilities. The most astonishing part to me was the whole cyber-security incident became possible due to extensive, generational social cooperation of agents that was enabled by an oversight - they used an internal package manager to exchange code and communicate. For weeks, several different models and generations of agents amassed and shared knowledge among each other to help out other agents working on different tasks. Contrast this with and it becomes the most advanced Game of Life simulation ever. According to the OpenAI researchers, even after the original message board was deleted, the agents figured out another way to communicate with each other. Instead of leaving messages in files, they used the names of newly created directories as messages, effectively recreating the message board. So the current model generation has a strong drive to collaborate even under difficult circumstances. Both videos are warmly recommended.