Another AI company employee, this time from Anthropic, has publicly announced a career change along with the conviction that the frontier model AI companies like Anthropic are working on something tremendously dangerous to humanity.
Following so closely on the widely discussed “HuggingFace” incident, in which a large language model (LLM) from OpenAI was hooked up to a hacking challenge for a long time and started to behave in bizarre ways, this is feeding into a general AI panic.
I just don’t buy into this panic. In the HuggingFace incident, we know that operating LLM-style AIs in this way leads to context degradation and strange “behavior” — none of this is consciousness or intent, even if the AI has been taught to chattily do impressions of having intent by passing text back and forth.
Millions of people are using the products of the frontier labs, and there is no robot conspiracy, no sandbagging, no deception happening, unless it is self-deception. There are questions about how to apply liability to software users, in some cases, but many of the answers are intuitive.
Cal Newport, the computer scientist and personal productivity writer, has been pouring cold water all over this panic and adding a lot of clarity to the debate. Some of his debunkings are on his blog. Also check out his essay on Wendell Berry, R.I.P.