@78
I don't understand the question? Mac circuitry is only special because they were in the perfect position to do it though. X86 probably isn't on the way out, but I doubt it'll be the first choice for new stuff much longer. Amazon has cheaper datacenter offerings based off Arm already that you can rent from them.
@77
This probably isn't the place for the argument but I'm bored so I'll have it anyway. It doesn't matter if you can push the AI just far enough that it can start solving these problems. We haven't, but AI is something where every single goddamn time I say "x won't happen until" x has happened sooner than I'd expect, and while the headlines are all ChatGPT lots of people are busy applying this stuff to solving real science, e.g. searching for medicine. And while perhaps model minimization needs to come along, and maybe power is an issue, both are problems that can be solved on a 5-10 year time horizon. Plus, really let's be honest, we've gone from using f64 to 4-bit quantization in just the last 5 years to mention only one thing I am aware of, that's over a 10x improvement.
Still you and a lot of other people always talk about how we need to democratize it and it's just like, no we don't, I am personally extremely glad that it requires expertise, the absolute last thing I want is for anyone to be able to make one, I would literally and genuinely prefer to just give everyone nukes first. No joke, I view nukes as safer. Whoever gets an AI above the threshold of say junior programmer first wins. Not something specific, they win everything it is possible to win. Or, more likely, we botch alignment and get the horrifying endings.
Get that far, though, and we are at most months from the singularity, and we are terrifyingly close to that right now if you consider that for the most part it turns out that you build giant buildings and stuff them full of computers and all you have to do for a better AI is, apparently, stuff more computers in. Like if I made a novel out of this and published it a decade ago I'd have been laughed out of even the highly speculative science fiction genre because you're kidding, right? They just kept stacking chips in a pile? That's your plot device? Everyone is so quick to say gimmick and also so quick to forget what they'd have thought 5 or 10 years ago if you describe the modern world to them back then. Clearly AI is more than this, but seriously "just use enough electricity and your problems are solved" level solutions? They weren't supposed to even be on the freaking table. If you look at how all of this works on the one hand you can always point at a million obstacles, that's true of any field, but on the other hand AI is happening so easily--relatively speaking--that it's almost like having enough compute in one place just naturally wants to let intelligence happen or something.
If I were you--or anyone else--decide what line AI has to cross before you take it seriously. Write it down right now so that you all will stop changing your minds on what that is. Then check if it has already, a lot of the time it turns out it did, and if it hasn't check every 6 months. I didn't; my opinion was changed by a bunch of "nah can't happen" and then it did predictions going against me. If I had to say mine, probably GPT4 (not Copilot, just the normal one) passing our coding interviews, the music offerings becoming passable or, going way way back, when that friend of mine got Ai Dungeon to engage in incredibly, unbelievably niche kink roleplay.
I should really blog what I call the idea oracle, and I could continue on at length, but put simply if you give me an AI which can come up with an important scientific insight 1 time out of 100000, I could likely leverage that plus some relatively simple hierarchical human structures shaped like mechanical turk to filter out most of them before passing the possibly good ones on up to the next level of qualification and so on. Like if you had a box that is that good it's better than humanity is right there, even if you had to check them all with college-level or higher people it's still better than what we have, checking 100000 ideas for validity is something we could almost automate. The other part of this that always gets me so hard is that no one realizes how low the threshold is for it to push us over the line like that. Imagine 1/10000? 1/1000? And you can point these at "soft" fields like chemistry where "synthesize this molecule cheaply" for example has tens of thousands of ways to do it so that it's not like physics where only exactly one solution exists to mine for, plus also we have reasonably good simulators for biology and chemistry and a lot of other "engineering"-type science that can check the output before even getting to the expensive and slow people...
My BlogTwitter: @ajhicks1992