The most useful way to think about AI capabilities is to think of target domains in terms of 3 orthogonal axis:
- Ease of building a verifier (coding is easy, lab equipment manipulation is hard)
- Causal complexity in terms of number of confounds (math problems is low as answers don’t depend on external factors, startup success is high complexity as it depends on many random factors)
- Economic attractiveness (high for coding, low for many domains)
What we’ve seen the first to be automated are domains where building verifiers is easy, causal complexity is less and economic attractiveness is high.
No wonder coding is the first one to see big jump.
But inflated valuations of AI companies need to be justified, so I fully expect that these companies will keep attacking the next best domain they can until they exhaust the economic attractiveness constraint.
Join 200k followers
Follow @paraschopra
Get new essays on your email: