Winter tights
Coming back to Gothenburg yesterday, I felt cold to the bone, both from the northerly winds and from watching an interview with Paul Christiano, who previously headed alignment research at OpenAI and now sits on its board. Following the recent METR assessments showing how rapidly AI capabilities are advancing at the frontier labs, estimates of when recursive self-improvement might become possible have moved ever closer, with some now pointing to 2027 as a decisive year. Either superintelligence will prove more difficult to achieve than we currently believe, or we really are approaching the technological singularity that we have long feared in science fiction.
Waking up to dramatic skies, I did what I could to protect myself from the cold, pulling on my winter tights for a morning run through a forest beautifully dusted with snow. As for the AI angst, there was less I could do, especially as my sister called to express her deep concerns once again about where it is all heading, echoing our conversation back in March, but with renewed urgency. Working on the front lines of AI deployment in medicine, her worries are both professional and existential. Just as with the kids, I find myself unable to offer much comfort beyond the need to maintain some semblance of everyday sanity and make the most of the time we do have.
As with the war in Ukraine, there is a limit to how long it is psychologically possible to sustain a heightened sense of alarm. At some point, one simply has to get on with life. The problem, of course, is that when one's work revolves around AI or, as in my case, global catastrophic risks, that is easier said than done.



















