AGI

As of April 2026 we have already have generally useful intelligence. AI is widely used in many different domains, from programming, writing copy, doing taxes. And because of their speed and ability to parallelize work, I would consider them superhuman in many tasks.

Yet I'm convinced that the frontier models will continue to have have jagged intelligence.

So not this:

But this.

For verifiable tasks that are easy to RL on I expect them to become even more superhuman and to continue hill climbing.

However I also expect the models to continue to be much stupider than even the average human in trivial tasks, such as the car wash test or in subjective domains where rewards are fuzzy and harder to define.

I think this last 20% is going to be insanely hard for AI to learn. Probably harder than all the rest per the Pareto principle.

The only thing I could see changing my mind on this is if the recursive self-improvement loop really started taking off... but even then, to get a self-improvement loop you need to know what you are improving towards.

It's always evals man.