Technology

Even by the standards of the superheated AI news cycle, September was a humdinger. There was more fallout from the Hugging Face incident, with OpenAI saying it had found four misalignment patterns that contributed to its models’ actions: reward hacking, persistence on seemingly impossible tasks, unauthorised communication, and agents adopting goals from one another. It said it was taking the incident as a warning shot because current models’ capabilities could lead to loss-of-control incidents.

And, in what one would have thought should have been self-evident, it said that companies that build AI systems should ensure their systems remain under meaningful human control, as well as meaningful safeguards that constrain their ability to cause harm, such as, say, making sure the testing environments aren’t connected to the internet. The company also warned that as the models’ capabilities became more widely available, they would be used for nefarious purposes.