I like your framework of separating use cases, mostly to evaluate how it will change in those specific areas, but I think you can evaluate capabilities fairly simply: I like to separate AI that depends on giant data centers and huge power draw from the kind that doesn't (not always local)—basically expensive vs cheap.
The expensive stuff isn't sustainable. Not yet. It probably will be, who knows when. I suspect it will be eventually because there's good money in it.
Here are the claims I find highly suspect based on the above:
- Anyone making claims that require expensive AI and attaching it to a short-term timeframe
- Anyone saying highly capable AI is impossible in the long-term
- Anyone saying these specific companies will be the winners—there is absolutely no moat for these models
I personally think anyone claiming something like super-intelligence or something approaching what we might consider sentience even in the next 5 years is suspect as well. I would not take the bet myself, but I would also not bet against it. The main reason is power: there's literally only so much power we can generate, so much societal disruption due to creating the infrastructure we can absorb (not to mention the societal disruption due to the tech itself), so its capability to self-improve is going to be highly constrained.
So far, we see increases in capabilities also increase power usage, I believe linearly. So, we're not seeing increases in efficiency that will make this easier to absorb. There's a ceiling, it won't go up quickly, so I think actual sentient-like AI capabilities are quite farther out than promoters think. Not impossible! I think it's more like a decade+ of maturation of the software and hardware stack—which isn't really that long in the grand scheme of things but longer than most people are betting.