AGI

I believe that as of April 2026 we have achieved my definition of AGI, which is generally useful intelligence. This means we have AI that is useful and widespread for work across many domains. I think that is true in coding, in science, writing copy, doing taxes. It's quite incredible what the models can do and I would say due to their speed and parallelization abilities, they are superhuman.

"Maybe, but maybe it looks like this"

For verifiable tasks that are easy to RL on I expect them to become even more superhuman and hill climb all of these skills very well. However there are simultaneously many obvious places in which they fail and they are much stupider than humans. Many such examples: car wash test.

I think that the last 20% is going to be insanely hard for AI to learn. Probably harder than all the rest per Pareto. What is this 20%? Almost certainly subjective, non-verifiable tasks.

The only thing I could see changing my mind on this is if the recursive self-improvement loop really started taking off.