Robots Might Be 1000x Harder Than Superintelligence

When assessing how difficult a task will be for AI, we mostly use humans as a proxy. For example, most humans could learn construction work, while fewer humans could do really advanced math – so advanced math must be a harder task than construction. This is a fallacy.

Our brains have perfected manipulating objects and moving around the world for millions of years. Modern mathematics is at best one thousand years old – a timespan so short our brains have not changed to be better at it. In machine learning terms: Math is still out-of-distribution for our monkey brains!

The underlying task, solving navigation and object manipulation, might well be 1000x harder than solving mathematics. Humans are just naturally 1000x better at it so the complexity is hidden from us.

Most likely, there are many more tasks that humans are naturally insanely good at but will take much longer for machines to conquer. ARC-AGI might be one of these.

Currently, the best way to sort tasks is to build a machine and see if it succeeds. This strategy is hardly optimal. Being able to sort tasks perfectly would let us vastly accelerate progress – so it's a worthwhile study.

Maybe a good razor is that if our ancestors were doing it millions of years ago, it could be hard for AI. If the task is only thousands of years old, it's most likely pretty easy.

October 20, 2024 – Max Rumpf