Regarding computing technology, I think the positive feedback you're describing happened with chip design and vlsi stuff, eg. better computers help design the next generation of chips or help lead to materials breakthroughs. I'm willing to believe LLMs have a macro effect on knowledge work in a similar way search engines, but as you said, it remains to be seen whether the models can feed back into their own development. From what I can tell, gpu speed and efficiency along with better data sets are the most important inputs for these things. Maybe synthetic data works out, who knows.
The only reason we stopped making larger nuclear weapons is because they were way, way, way beyond useful for anything. There’s no reason to believe an upper bound exists in the physical universe (especially given how tiny and energy efficient the human brain is, we’re definitely nowhere near it) and there’s no reason to believe an upper bound exists on the usefulness of marginally more intelligence. Especially when you’re competing for resources with other nearly-as-intelligent superintelligences.
There were lots of plans for using nukes for large-scale engineering projects.
- Make a protected harbor from a straight coastline
- Get that mountain that's in the way of your railroad flattened
- Add a freshwater lake to this plain
- Mine an entire vein at once
Reality got in the way and we found out that we don't want to be doing that.
Maybe if we get practical antimatter bombs we'll revisit those plans.
The question is, "Can you create a chain reaction that grows?", and the answer is unclear right now with AI, but it's hard to say with any confidence that the answer is "no". Most experts five years ago would have confidently declared that passing the Turing test was decades to centuries away, if it ever happened, but it turned out to just require beefing up an architecture that was already around and spending some serious cash. I have similarly low faith that the experts today have a good sense that e.g. you can't train an LLM to do meaningful LLM research. Once that's possible, the sky is the limit, and there's really no predicting what these systems could or could not do.