What makes a labor market for agents that recognize problems and provide solutions special or different from markets for other kinds of labor? If AIs get to a point where they dramatically outperform humans in other forms of labor, why not in this one?

I think some humans will be doing it well enough to keep themselves afloat the rest of our lifetime, and some will get fabulously rich building products as a one-man operation leveraging AIs. But there will be far more people failing at it. It will be like Youtube creators or Instagram influencers where there are few winners who take virtually all the rewards.

compared to the broadcast era aren't there way more winners -- with a smaller pieces of the pie -- nowadays?

it's still a Pareto distribution, I'm sure, but mega-stardom kinda died and was replaced by all these mini-stars, as far as I can tell. I'm not sure it supports your hypothesis.

Sure, there is some truth to this.

I'm not really in touch with other genres, but I like to watch chess videos/streams on Youtube and Twitch. The vast, vast majority of views and revenue are captured by about ten people.

I like those people too, but I've also watched a lot of smaller acts, even some amateur players not much stronger than me. So I get those recommendations, and I see their view counts. They aren't making anything at all.

There are other people who have some followers, but even 50,000 followers would be a dream for most people doing it and they will make next to nothing from that. I'd guess there are at least 30x the number of strong, titled players in the 50k group as there are in the 1MM+ group. These are all people who were chess prodigies as kids, won every scholastic tournament in their state, took gap years or went to colleges that let them basically major in chess, travelled the world for tournaments, with awe-inspiring skills, and they are not making anywhere close enough to live on.

And the thing is, I think software might even be tougher in twenty years. Its hard to get people to change from a system they use to another thing, much harder than recommending a new face on Youtbue.

Maybe someday they will. But the current run of LLMs are fantastic at regurgitating and synthesizing existing knowledge, and getting better all the time, but not so good at coming up with new ideas. As long as you keep to the realm of what is known, they can seem incredibly intelligent, but as soon as you cross that boundary there's a clear change - often to just meaningless bullshit. So, I personally don't think we're going to be outsourcing idea generation to LLMs (or AI in general) anytime soon. Though to be fair, I'm only about 75% confident in that, and even so, it doesn't mean they won't be hugely transformative anyway.

Pessimistically but realistically, it doesn't matter if AI will perform better, it just matters if it's cheaper. A historical example is all the offshoring of mission critical code starting in the late 90s, early 00s. The code that came back was sub-par many of the times, particularly for the cheaper shops, but the executives got their bonus for saving money and bailed out. The new executives are now in charge of fixing the disaster of a codebase that they were left with.

I think history will rhyme with the offshoring trend but with AI this time.