A number of startups are working in verifiable domains where they can provide realistic data. This is an interesting thread from one of those startups: https://x.com/khoomeik/status/1973056771515138175
Aside from the better versions of what AI is visibly doing now (software dev, human language translation, video gen, etc), many of the AI bears are dismissing the potential impact of hooking AI up with automated experimentation so it's able to generate new types of data to train itself. The impact on drug discovery, material science, and other domains are likely to be very significant. The Nobel Prize in Chemistry for AlphaFold is just a glimpse of this future.
Here's a discussion with Isomorphic Labs (Google DeepMind spinoff) on this line of thinking: https://www.youtube.com/watch?v=XpIMuCeEtSk