After the bubble collapses, I believe there is going to be a rule of thumb for whatever tiny niche use cases LLMs might have: “Never let an LLM have any decision-making power.” At most, LLMs will serve as a heuristic function for an algorithm that actually works.

Unlike the railroads of the First Gilded Age, I don’t think GenAI will have many long term viable use cases. The problem is that it has two characteristics that do not go well together: unreliability and expense. Generally, it’s not worth spending lots of money on a task where you don’t need reliability.

The sheer expense of GenAI has been subsidized by the massive amounts of money thrown at it by tech CEOs and venture capital. People do not realize how much hundreds of billions of dollars is. On a more concrete scale, people only see the fun little chat box when they open ChatGPT, and they do not see the millions of dollars worth of hardware needed to even run a single instance of ChatGPT. The unreliability of GenAI is much harder to hide completely, but it has been masked by some of the most aggressive marketing in history towards an audience that has already drunk the tech hype Kool-Aid. Who else would look at a tool that deletes their entire hard drive and still ever consider using it again?

The unreliability is not really solvable (after hundreds of billions of dollars of trying), but the expense can be reduced at the cost of making the model even less reliable. I expect the true “use cases” to be mainly spam, and perhaps students cheating on homework.