9 ms·

This is an interesting criticism. I grew up on wired magazine and sci-fi. Even then I recognized the dangers of shallow knowledge and maybe I can’t avoid it eve

This is an interesting criticism. I grew up on wired magazine and sci-fi. Even then I recognized the dangers of shallow knowledge and maybe I can’t avoid it even now. But I’m actively working with GPT3 and I must say it is startling. I’d say we’re a few years from automating most knowledge tasks (I created a tool that I chose not to release because it was too powerful). We could be a decade from general purpose artificial intelligence (or it might not be possible at all. For a fun thought experiment ask GPT3 if it is truly conscious). The answer it gave me was surprisingly coherent wise and useful. From there (AGI) we could be minutes away from artificial super intelligence. That starts to look a bit like a point pst which standard prediction breaks down. But like you say, pop sci might be hindering rational thinking on this score.

whimsicalism 5y ago
I used to work in physics, which seems to attract a lot of people with shallow knowledge, and ML seems to be going in a similar direction. To be clear, I'm actually quite bullish on AI (well, bullish on it ultimately happening, not necessarily on how it will impact humanity). I was mostly reacting to the timeline for photonic quantum computing rather than on ML advancement, actually. Language modeling is extremely impressive. I've been saying since 2017 that modeling language is basically modeling the world/intelligence, but at the time most people did not seem to believe that. People still don't really seem to believe it, but I think the smarter among us are beginning to now that they see GPT-X & PaLM. I think there is still a substantial gap from "good language model" to "embodied agent taking actions in the world" that we have not bridged and will be potentially pretty hard to, more than mere "minutes."
- zmgsabst 5y ago
  > will be potentially pretty hard to, more than mere "minutes." This makes it sound like you’re arguing against something no one said — which makes it a weird thing to quote.
  - whimsicalism 5y ago
    A quote from the GP > The answer it gave me was surprisingly coherent wise and useful. From there (AGI) we could be minutes away from artificial super intelligence What am I missing? I assumed of course that they didn't mean literal minutes but more metaphorically, but my reply was also using it in the metaphoric sense - it could be a potentially heavy lift to go from point A to point B. e: I guess I am missing that they are conflating GPT with AGI and saying it will be minutes from AGI to artificial super intelligence in the literal sense. That, I actually do agree with, but I don't think GPT-X or PaLM or LMs in general qualify as such.
    - zmgsabst 5y ago
      I read it in a different way, where we’d go from “plays chess like a bad human” to “plays chess superhumanly”, applied to general knowledge synthesis tasks — something that would be a rapid change once it was able to synthesize general knowledge answers. Upon reflection, my comment was too critical for having a different read than you did.

whimsicalism 5y ago

I used to work in physics, which seems to attract a lot of people with shallow knowledge, and ML seems to be going in a similar direction. To be clear, I'm actually quite bullish on AI (well, bullish on it ultimately happening, not necessarily on how it will impact humanity). I was mostly reacting to the timeline for photonic quantum computing rather than on ML advancement, actually. Language modeling is extremely impressive. I've been saying since 2017 that modeling language is basically modeling the world/intelligence, but at the time most people did not seem to believe that. People still don't really seem to believe it, but I think the smarter among us are beginning to now that they see GPT-X & PaLM. I think there is still a substantial gap from "good language model" to "embodied agent taking actions in the world" that we have not bridged and will be potentially pretty hard to, more than mere "minutes."

zmgsabst 5y ago
> will be potentially pretty hard to, more than mere "minutes." This makes it sound like you’re arguing against something no one said — which makes it a weird thing to quote.
- whimsicalism 5y ago
  A quote from the GP > The answer it gave me was surprisingly coherent wise and useful. From there (AGI) we could be minutes away from artificial super intelligence What am I missing? I assumed of course that they didn't mean literal minutes but more metaphorically, but my reply was also using it in the metaphoric sense - it could be a potentially heavy lift to go from point A to point B. e: I guess I am missing that they are conflating GPT with AGI and saying it will be minutes from AGI to artificial super intelligence in the literal sense. That, I actually do agree with, but I don't think GPT-X or PaLM or LMs in general qualify as such.
  - zmgsabst 5y ago
    I read it in a different way, where we’d go from “plays chess like a bad human” to “plays chess superhumanly”, applied to general knowledge synthesis tasks — something that would be a rapid change once it was able to synthesize general knowledge answers. Upon reflection, my comment was too critical for having a different read than you did.