4 ms·

Unfortunately, No AGI Yet

3 points by georgehill 3y ago

ggm 3y ago
Also no FTL, no perpetual motion machines, no time machines. And no cold fusion, and it looks like no room temperature superconductors. The problem isn't "no <x>" the problem is the extreme hype and FOMO in people who want to believe "<x> is around the corner" and build systems, belief, money, ponzi schemes, you-name-it predicated on <x> but, as you note, AGI isn't happening. Taking LLMs and generative systems as evidence of emergent behaviour is taking fools gold in the rock for signs of the valuable stuff. Ask yourself: How many times have you downvoted people like me who say "it's not happening" because of your emotional ties to the belief it is happening.
- mitthrowaway2 3y ago
  I'm not aware of FTL, perpetual motion, time machines, or cold fusion having 8 billion extant biological instances populating the planet and acting as compelling proofs of the thing being not only physically possible, but even possible just by putting molecules in a sunlit jar and shaking them for a billion years. But sure, maybe it's just like perpetual motion, time travel, and other things that violate the known laws of physics. No way to know, right?
  - ggm 3y ago
    That's true. We know how to make GI organically. It's the inorganic kind which is proving elusive. Your faith that training over giant corpus of text is the path, is frankly touching. Bless.
    - mitthrowaway2 3y ago
      Actually, I was expressing faith that shaking molecules in a jar is the path to creating general intelligence. Training over text might work too, I don't know; what I can say is that the more AI tempts investors with the prize of replacing all human mental labour, and the more billions of dollars and brilliant researchers are thrown at picking that prize, the more likely it is that someone will eventually crack the challenge. The apparent effectiveness of transformer-based LLMs is merely weak evidence that, whatever the eventual solution does end up being, it might only require a relatively small number of core insights that work unexpectedly well when applied at large scale. What I don't have is blind confidence that it's an impossible problem that nobody will figure out no matter how hard they try. (Whereas I do have that confidence with respect to perpetual motion remaining unsolved).

3 points by georgehill 3y ago

ggm 3y ago

Also no FTL, no perpetual motion machines, no time machines. And no cold fusion, and it looks like no room temperature superconductors. The problem isn't "no <x>" the problem is the extreme hype and FOMO in people who want to believe "<x> is around the corner" and build systems, belief, money, ponzi schemes, you-name-it predicated on <x> but, as you note, AGI isn't happening. Taking LLMs and generative systems as evidence of emergent behaviour is taking fools gold in the rock for signs of the valuable stuff. Ask yourself: How many times have you downvoted people like me who say "it's not happening" because of your emotional ties to the belief it is happening.

mitthrowaway2 3y ago
I'm not aware of FTL, perpetual motion, time machines, or cold fusion having 8 billion extant biological instances populating the planet and acting as compelling proofs of the thing being not only physically possible, but even possible just by putting molecules in a sunlit jar and shaking them for a billion years. But sure, maybe it's just like perpetual motion, time travel, and other things that violate the known laws of physics. No way to know, right?
- ggm 3y ago
  That's true. We know how to make GI organically. It's the inorganic kind which is proving elusive. Your faith that training over giant corpus of text is the path, is frankly touching. Bless.
  - mitthrowaway2 3y ago
    Actually, I was expressing faith that shaking molecules in a jar is the path to creating general intelligence. Training over text might work too, I don't know; what I can say is that the more AI tempts investors with the prize of replacing all human mental labour, and the more billions of dollars and brilliant researchers are thrown at picking that prize, the more likely it is that someone will eventually crack the challenge. The apparent effectiveness of transformer-based LLMs is merely weak evidence that, whatever the eventual solution does end up being, it might only require a relatively small number of core insights that work unexpectedly well when applied at large scale. What I don't have is blind confidence that it's an impossible problem that nobody will figure out no matter how hard they try. (Whereas I do have that confidence with respect to perpetual motion remaining unsolved).