If you're talking about human-level intelligence, I think you're on to something. Old-fashioned AI was probably closer to the right track than we are now.
> did attempts to generalize some of that research fail because the approach was fundamentally flawed, or was it because such efforts themselves weren't as good as initial projects?
In many cases there haven't _been_ attempts to follow up on that early work. It's more like the zeitgeist just shifted to other things. For example, SHRDLU was just a couple years before the first AI winter, and when spring came (~1980) people had largely moved on. (Which isn't to say there weren't also flaws with the approach.)
Not the GP, but some examples are: (1) how do you get a computer to have a human-like train of thought? (2) how do you get a computer to acquire new concepts (e.g. "debt", "global warming", "weed") and then reason about them correctly, without any reprogramming? (3) automated acquisition of common sense through experience (e.g. "if you pour water on the floor you will get a puddle") (4) deep natural language understanding (i.e. how do you make a chatbot that really understands, and isn't just a thin illusion of understanding).
None of the problems you mention is actually solved though. They're all things that work sort of, some of the time, with caveats about how you define "work." They work well enough to be useful, but not well enough to argue we're converging on human-level intelligence.
Yes and no. For safety of narrow AI systems, yeah, there's a lot of scope for research, and that's what your first link gets at.
But for AGI (which is what Tegmark talks about), there's no good way to get a handle on safety yet (other than working towards figuring out AGI).
As for MIRI's agenda, I don't buy that it will help with AGI safety at all. There are a variety of reasons for that, some of which are discussed in the piece I linked above.
Yes, this is the problem with AI risk---there's a community pushing hard to gather resources to the cause, but little or no scientific work to be done. This is a rather pathological situation---among other things, the AI risk community makes their own cause look silly, and they promote an unduly negative vision of AGI. I've written more about this here: http://www.basicai.org/blog/ai-risk-2017-08-08.html.
On a positive note, as a piece of science fiction, this was an enjoyable read!
An interesting read, but the conclusion that "the singularity is nowhere near" was reached by assuming that only neural modeling could get us there, and that assumption wasn't defended well. (In fact it looks rather dubious, given all the quasi-intelligent things computers have achieved without copying neural dynamics.)
The biggest deficiency in AI is that we still don't have artificial systems which simulate human thought with any fidelity. Sooner or later that's bound to become a focus of attention.
The presentation briefly mentioned simulating the brain, but I think what's more likely to succeed is mimicking the mind at a high level of abstraction (i.e. a level we can study with introspective or even linguistic methods rather than neuroscience). There's some precedent for this with projects like Soar and ACT-R (and even some recent interest from mathematicians [1]).
IMHO this kind of methodology could be pushed much further.
Eventually I think AI safety will be solved through some mixture of design choices, supervision/monitoring, and human-administered "incentives" for good behavior (not unlike the reward signals in reinforcement learning).
But to flesh that out in detail requires a specific AGI design, something we're far from achieving. The current inability to get specific is probably why AI risk doesn't get more attention (though it does get a lot).
> did attempts to generalize some of that research fail because the approach was fundamentally flawed, or was it because such efforts themselves weren't as good as initial projects?
In many cases there haven't _been_ attempts to follow up on that early work. It's more like the zeitgeist just shifted to other things. For example, SHRDLU was just a couple years before the first AI winter, and when spring came (~1980) people had largely moved on. (Which isn't to say there weren't also flaws with the approach.)