4 ms·

And all of it just driven on vague statements of "AGI next year" (a year ago) and now "cancer cure on 10 gigawatts". Almost every podcast nowadays is just "Whe

by mlnj 1y ago

And all of it just driven on vague statements of "AGI next year" (a year ago) and now "cancer cure on 10 gigawatts". Almost every podcast nowadays is just "Where is all this money coming from?", "This is just infinite money and you are just out of the loop." Feels so weird to hear.

impossiblefork 1y ago
I have actually become less skeptical about claims like this over time. I didn't expect LMs to be able to attack IMO problems this soon. There's also this video stuff recently discussed here, where it was explained that video models can be prompted to segment, solve blind deconvolution problems, solve maces, etc. I'm not sure it's going to work out economically-- I think a crash is unavoidable because P/E ratios for firms like Tesla of 200 or whatever it is are silly, but I think that technology is going to keep evolving throughout and I don't see why some kind of mild not-quite-AGI-but-beating-humans-on-what-we-pay-humans-to-do-for-their-jobs type thing is not achievable. If these video models get better and more available we might actually be able to use them to spot patterns in the natural world and in humans.
- marginalia_nu 1y ago
  Yeah it's worth remembering that the dot com bubble didn't come crashing down because the ideas weren't viable, we're very much living in a world dominated by the very things people were hyped on back then. The dot com bubble came crashing down because we were still not there yet, at the turn of the millenium. We lacked the logistics to make large scale ecommerce viable, and we also lacked the smartphones to make it practical. I generally feel we are in a very similar position right now. LLMs shows real promise, but also real limitations. I'm very skeptical it will bring significant utility in the next few years. Maybe in a decade or more, when other pieces of the puzzle are in place.
  - tasuki 1y ago
    > I'm very skeptical it will bring significant utility in the next few years. LLMs are already bringing significant utility to me. I don't think they're a good investment though: they're very expensive to train and somewhat expensive to run. And I'm extremely fickle and will jump ship to the next one without a second thought.
    - marginalia_nu 1y ago
      Sure, but is this a degree of utility that stands in proportion to the trillion dollar valuation of the market sector?
- conartist6 1y ago
  That's the mirage of value. It's a mirage 'cause: a) it isn't creating that value but taking it from somewhere else (past IME contestants and test preparers, StackOverflow, the job market, etc) b) users fight to bid up the value of a single set of fixed capabilities, like rats fighting over the same scrap of leftovers. The market floods with generic copies of whatever stuff it can do. That is to say it's utterly, humorlessly laughable to think that we're trending towards these things doing a whole economy worth of work.
  - _DeadFred_ 1y ago
    Not sure why you are getting downvoted. AI only worked when you had an internet full of people taking their time to post/help others out. When AI takes those peoples jobs, they stop posting. I don't post anything AI useful on public sites (I answer peoples questions via DM if I want to help). Imagine the constraint on every breaking change going forward being 'can AI be trained to handle these changes against the pre-AI training dataset'. Because AI is an inertia machine. It takes all the momentum from everything it touches. Next step after it takes the devs jobs is to freeze out all breaking changes that could make the AI training set useless (and we don't have a new one yet, because it's a new breaking change, and also because, you know, no one is posting knowledge for free AI scrapping anymore, yay, huge step forward for humankind!!!). Inertia machine sucking the momentum out of everything. They tried working with a movie studio for a video tool. They needed more movies than the output of the studio. How do you train in new things if you just happen to need 5 studios worth of 'new thing' but output is all from 'old thing' and AI generated? Inertia machine.

by mlnj 1y ago

impossiblefork 1y ago

I have actually become less skeptical about claims like this over time. I didn't expect LMs to be able to attack IMO problems this soon. There's also this video stuff recently discussed here, where it was explained that video models can be prompted to segment, solve blind deconvolution problems, solve maces, etc. I'm not sure it's going to work out economically-- I think a crash is unavoidable because P/E ratios for firms like Tesla of 200 or whatever it is are silly, but I think that technology is going to keep evolving throughout and I don't see why some kind of mild not-quite-AGI-but-beating-humans-on-what-we-pay-humans-to-do-for-their-jobs type thing is not achievable. If these video models get better and more available we might actually be able to use them to spot patterns in the natural world and in humans.

marginalia_nu 1y ago
Yeah it's worth remembering that the dot com bubble didn't come crashing down because the ideas weren't viable, we're very much living in a world dominated by the very things people were hyped on back then. The dot com bubble came crashing down because we were still not there yet, at the turn of the millenium. We lacked the logistics to make large scale ecommerce viable, and we also lacked the smartphones to make it practical. I generally feel we are in a very similar position right now. LLMs shows real promise, but also real limitations. I'm very skeptical it will bring significant utility in the next few years. Maybe in a decade or more, when other pieces of the puzzle are in place.
- tasuki 1y ago
  > I'm very skeptical it will bring significant utility in the next few years. LLMs are already bringing significant utility to me. I don't think they're a good investment though: they're very expensive to train and somewhat expensive to run. And I'm extremely fickle and will jump ship to the next one without a second thought.
  - marginalia_nu 1y ago
    Sure, but is this a degree of utility that stands in proportion to the trillion dollar valuation of the market sector?
conartist6 1y ago
That's the mirage of value. It's a mirage 'cause: a) it isn't creating that value but taking it from somewhere else (past IME contestants and test preparers, StackOverflow, the job market, etc) b) users fight to bid up the value of a single set of fixed capabilities, like rats fighting over the same scrap of leftovers. The market floods with generic copies of whatever stuff it can do. That is to say it's utterly, humorlessly laughable to think that we're trending towards these things doing a whole economy worth of work.
- _DeadFred_ 1y ago
  Not sure why you are getting downvoted. AI only worked when you had an internet full of people taking their time to post/help others out. When AI takes those peoples jobs, they stop posting. I don't post anything AI useful on public sites (I answer peoples questions via DM if I want to help). Imagine the constraint on every breaking change going forward being 'can AI be trained to handle these changes against the pre-AI training dataset'. Because AI is an inertia machine. It takes all the momentum from everything it touches. Next step after it takes the devs jobs is to freeze out all breaking changes that could make the AI training set useless (and we don't have a new one yet, because it's a new breaking change, and also because, you know, no one is posting knowledge for free AI scrapping anymore, yay, huge step forward for humankind!!!). Inertia machine sucking the momentum out of everything. They tried working with a movie studio for a video tool. They needed more movies than the output of the studio. How do you train in new things if you just happen to need 5 studios worth of 'new thing' but output is all from 'old thing' and AI generated? Inertia machine.