[visarga]

I believe AI will be amazing at driving job creation - it can be applied to almost all fields and it never works well on its own, it always suffers from limited autonomy.

Say your model makes 99% correct predictions, in 20-30 time steps that error rate drops to 73%, that is just not ok for automation. But real accuracy rates per time step are much lower.

Compounding errors even when they are small lead a system astray. I think there is no form of AI that achieves autonomy in any field. We're at minutes of autonomy, or seconds. Nobody can go on vacation and leave AI do their job.

On the other hand AI will open up new opportunities and markets. Demand will scale up to meet the new productivity level. We can always desire more, it's not a fixed sum game.

More fundamentally, LLMs are interpolating between and combining known skills. They don't do radically new discovery. Why? They train on human text instead of human text + world feedback. In order to surpass human experts AI needs labs, experiments or the ability to create its own experience.

I can only name AlphaZero, AlphaTensor and AlphaFold as superhuman AIs, and they have been trained on massive experimental feedback, not just text.