4 ms·

Three years? That’s a stunningly optimistic (pessimistic?) timeline…

by germinalphrase 5y ago

FeepingCreature 5y ago
What can I say, PaLM is scary.
- NoGravitas 5y ago
  The hand-picked examples are pretty scary, but I'm also reasonably sure that AGI is not going to come (directly) from LLMs no matter how big they get. Reasoning: https://medium.com/@emilymenonbender/on-nyt-magazine-on-ai-resist-the-urge-to-be-impressed-3d92fd9a0edd https://medium.com/@emilymenonbender/on-nyt-magazine-on-ai-r...
  - FeepingCreature 5y ago
    Seems too focused on the NYT article? I don't think this because the NYT said so, I think this because PaLM's chains-of-thought approach seems to match how my own brain works. I think LMs have huge intelligence overhang atm that is hidden by the way they are forced to respond. If we figure out how to train LMs with propagation through hidden chains of thought, we'll see a huge jump in capability, and also this is the obvious next step that the industry will aim for.
zmgsabst 5y ago
I’m betting 2025-2030 as well. We have quantum AI math developed and photonic accelerators are currently in development. The moment those work reliably, machine learning will hit an exponential curve — particularly since those photonic circuits are already partially designed by AI. The feedback loop of quantum-accelerated AI refining its own hardware is one breakthrough away.
- whimsicalism 5y ago
  Stop reading so much popsci, it is bad for thought.
  - zmgsabst 5y ago
    I’m sorry that you’re ignorant. Microsoft QuArC has published a number of papers on QAI, and I both work on photonics and discuss with other companies who do the accelerators being developed. Accelerating AI with photonics is an active area of development. I think you’re projecting that your knowledge of this is only based on pop-sci.
    - whimsicalism 5y ago
      There are a ton of unsolved problems in both ML and quantum computing with photonics. We are so far away from being able to do this with photonic quantum computers that I am really skeptical that you actively work in this field. FWIW, I know quite a lot of people involved in AMO physics (which involves related techniques) and work in ML myself. 2025-2030 is way too soon for these things to come to fruition.
      - zmgsabst 5y ago
        I guess you know less talented people then. We’re not more than a single breakthrough from being able to make photonic accelerators — and there’s already worked out algorithms to accelerate ML once those circuits are available. I’m confident in 2025-2030.
        whimsicalism 5y ago
        This is one of the ~top two universities in the country for AMO research, so I don't think so. Such false bravado, not going to keep replying.
        zmgsabst 5y ago
        > Such false bravado, not going to keep replying. Yawn — you’re the one who came in with ad hominems and trying to tell other people what they work on, lol I’m sorry your friends are ignorant and you don’t know what I do.
- 10amxn10 5y ago

Three years? That’s a stunningly optimistic (pessimistic?) timeline…

FeepingCreature 5y ago

What can I say, PaLM is scary.

NoGravitas 5y ago
The hand-picked examples are pretty scary, but I'm also reasonably sure that AGI is not going to come (directly) from LLMs no matter how big they get. Reasoning: https://medium.com/@emilymenonbender/on-nyt-magazine-on-ai-resist-the-urge-to-be-impressed-3d92fd9a0edd https://medium.com/@emilymenonbender/on-nyt-magazine-on-ai-r...
- FeepingCreature 5y ago
  Seems too focused on the NYT article? I don't think this because the NYT said so, I think this because PaLM's chains-of-thought approach seems to match how my own brain works. I think LMs have huge intelligence overhang atm that is hidden by the way they are forced to respond. If we figure out how to train LMs with propagation through hidden chains of thought, we'll see a huge jump in capability, and also this is the obvious next step that the industry will aim for.

zmgsabst 5y ago

I’m betting 2025-2030 as well. We have quantum AI math developed and photonic accelerators are currently in development. The moment those work reliably, machine learning will hit an exponential curve — particularly since those photonic circuits are already partially designed by AI. The feedback loop of quantum-accelerated AI refining its own hardware is one breakthrough away.

whimsicalism 5y ago
Stop reading so much popsci, it is bad for thought.
- zmgsabst 5y ago
  I’m sorry that you’re ignorant. Microsoft QuArC has published a number of papers on QAI, and I both work on photonics and discuss with other companies who do the accelerators being developed. Accelerating AI with photonics is an active area of development. I think you’re projecting that your knowledge of this is only based on pop-sci.
  - whimsicalism 5y ago
    There are a ton of unsolved problems in both ML and quantum computing with photonics. We are so far away from being able to do this with photonic quantum computers that I am really skeptical that you actively work in this field. FWIW, I know quite a lot of people involved in AMO physics (which involves related techniques) and work in ML myself. 2025-2030 is way too soon for these things to come to fruition.
    - zmgsabst 5y ago
      I guess you know less talented people then. We’re not more than a single breakthrough from being able to make photonic accelerators — and there’s already worked out algorithms to accelerate ML once those circuits are available. I’m confident in 2025-2030.
      - whimsicalism 5y ago
        This is one of the ~top two universities in the country for AMO research, so I don't think so. Such false bravado, not going to keep replying.
        zmgsabst 5y ago
        > Such false bravado, not going to keep replying. Yawn — you’re the one who came in with ad hominems and trying to tell other people what they work on, lol I’m sorry your friends are ignorant and you don’t know what I do.
10amxn10 5y ago