Starting Monday, I’ll be joining Google as a research scientist. Being at a frontier research corporation is a very interesting position to be in.
I suspect my beliefs will change during this experience, so I want to keep track of how they change. Here are my initial thoughts.
I suspect writing down my beliefs will also cause them to become more calcified or place me in different camps. I want to make explicit that these will most likely change, and the point of this exercise is to see them change. However, this is where they stand based on my AI research work up to this point.
Intelligence explosion is an ill-defined concept, so I don’t know what an intelligence explosion happening or not happening would mean. I think algorithmic progress, as measured by some pre-2026 metric like pretraining loss, could be 1000x, but I don’t know how much this actually means. Productivity metrics like number of papers can go up 100x while gains in end metrics like human lifespan could be very minimal.
I don’t believe in most other large-scale AI claims like 15% growth, AI controlling Earth, human immortality, a Kardashev Type 1 civilization, etc. by 2030.
Almost all verifiable research tasks that I was doing last year will be automated.
We will see big advances in AI research like we see in AI for math (this is both an optimistic and a pessimistic take).
I’m more optimistic about AI safety/security. It seems like we are not doing much at the moment and there is a lot of low-hanging fruit. However, AI hacking looks like it will be a big deal. I think we’ll have some fair warning shots, in line with other technologies like bio and chemical sciences, so not much out of the ordinary compared to these.
At the moment I’m underhyped by AI. It seems to have radical intelligence but has not transformed any of the end metrics that I really care about (anti-sleep medication?, cure for cancer, greater job satisfaction, etc.). In fact, I wish it went faster, at least in the sense that we could see more positive benefits for people.
I think having AI development in the hands of AI companies is still a good idea, but we do need more regulation.