10 ms·

I think the "sparks of AIG" presentations are compelling. It does feel like there are a just a few remaining big problems away from AGI level capabilities where

by mdale 3y ago

I think the "sparks of AIG" presentations are compelling. It does feel like there are a just a few remaining big problems away from AGI level capabilities where it was more futuristic unclear projections in the past. If the issues as synthetic data conversations like alpha go but for LLMs; setting much longer token contexts (or dramatically cheaper training on custom data sets) And scaling via software and hardware a few more orders of magnitude. I think then we have some big waves that will wash over any observations we could make at this point in time.

cmdli 3y ago
Honest question, not meaning to disagree with you: how do you view previous predictions of the success of AI, such as Marvin Minsky predicting in the 60s that AI “would substantially be solved in a generation”? What reasons would there be that the experts are correct this time?
- famouswaffles 3y ago
  Current Capability would be the biggest one. We're at the point where any testable definitions of GI that the sota LLM fails (GPT-4) is also failed by a good chunk of humans. You couldn't say that a few years ago nevermind 60. What we have now (so no hypotheticals) coupled with the fact that scaling hasn't yet shown any performance walls makes a pretty good shout that things will probably be different this time.
  - skepticATX 3y ago
    That's not really true though. LLMs are abysmal at planning, for example. Something that comes quite naturally to humans.
    - runsWphotons 3y ago
      They are probably better than 10% of people
      - danielmarkbruce 3y ago
        You meant 70%, right?
    - famouswaffles 3y ago
      They're really only abysmal if you attempt to one-shot it and probe with tasks that would require a human a scratchpad to accomplish. Humans can't one-shot non trivial planning tasks either. It's the one problem i have with all the papers that try to evaluate planning for LLMs. Step away from that approach and they're ok. https://innermonologue.github.io/ https://innermonologue.github.io/ https://tidybot.cs.princeton.edu/ https://tidybot.cs.princeton.edu/
    - pulvinar 3y ago
      I'm curious as to your source, or particular examples, since they (or at least GPT-4) seem to me to be rather decent at planning. E.g., for writing code.
      - climatologist 3y ago
        Ask it to write a backtracking sudoku solver with coroutines and/or fibers and let me know how it performs in your language of choice. We are nowhere near generally intelligent software systems.
        jtmoulia 3y ago
        I was curious where GPT-4 would come up short on the problem and I was surprised -- it seemed to solve it pretty well whether or not using coroutines. (I dropped both solns into a python interpreter and both appeared to solve the problem.) There could def be bugs I missed tho. https://chat.openai.com/share/ef77507e-cb75-4112-97f1-a16cfc03cd98 https://chat.openai.com/share/ef77507e-cb75-4112-97f1-a16cfc...
        climatologist 3y ago
        That's a good attempt but the coroutine solution is incorrect. See if you can figure out why and how to improve it. You can also ask it to propagate constraints and see what happens.
        el_nahual 3y ago
        This isn't the dunk you think it is since GP is a human (I presume), and he thought the solution worked.
        jtmoulia 3y ago
        Sigh classic LLM -- without you, the expert, I can't quickly tell from the code / output how the answer the LLM produced is wrong. I also asked it to solve sudoku by "propagating constraints" and the answer seemed to work for me :/ Again, I'd guess the soln produced is wrong because I trust you more than the LLM but I don't have the mental horsepower to figure it out without resorting to tests & debugging.
29athrowaway 3y ago
Yann LeCun says human intelligence is not "general intelligence".
- tru1ock 3y ago
  We are not GI, and GI is too wasteful to be useful in a scarce environment.
- esafak 3y ago
  Is monkey intelligence or dolphin intelligence qualitatively different? What is general intelligence, then?
flangola7 3y ago
It utterly befuddles me that so many people still can't (or refuse to) sense what's coming. Be it bad or good the magnitude of what is now unfolding is beyond everything that homo sapiens have ever witnessed. Unless we think that MI advancement will miraculously stall and never improve beyond current capabilities, I cannot conceive of future, even a near future, that we recognize as real. >I think then we have some big waves that will wash over any observations we could make at this point in time. This is spot on. Exactly as how we didn't know what the internet would be, the wiser amongst us realized it was a new and strange era. "An alien lifeform. Unimaginable, both amazing and terrifying." - Bowie. MI will far eclipse the internet.
- apsurd 3y ago
  I don't know, that's quite the romantic take. I can easily admit I don't know what's coming, but _homo sapiens have ever witnessed_... _cannot conceive of near future that we recognize as real_. Come on, lol Imagine yourself in the movie Apocalypto and I'm using a Hollywood film on purpose here - the entire Universe of being as you see it all the sudden explodes its Universe with new Gods/demons different and more powerful _everything_ coming to kill you. And then there's the boring discovery of "we're just one planet in one solar system in one Galaxy... LLMs gonna be orders of magnitude beyond that (in)comprehension? I mean maybe it will be more powerful and maybe it will come to kill us, but that's not an inconceivable new story line. Funny that both of us can accuse one another of hubris.
  - flangola7 3y ago
    Is this GPT-2 generated? I've tired reading it three times and I still can't make heads or tails of it.
    - apsurd 3y ago
      I see what you did there gpt-"2". at least you don't dislike enough to downvote. not sure what's not understandable. I think parent is overly hyperbolic. i get that chat GPT is unprecedented, but come on "impossible to conceive of near-term reality" ??? edit: oh! you're the parent. yeah i think you're wildly hyperbolic. There's more drastic historical precedents but of course the future is by definition going to be inevitably more inconceivable over a long enough arc, so I don't think of this as a debate. I just think you're being overly dramatic for effect.

by mdale 3y ago

cmdli 3y ago

Honest question, not meaning to disagree with you: how do you view previous predictions of the success of AI, such as Marvin Minsky predicting in the 60s that AI “would substantially be solved in a generation”? What reasons would there be that the experts are correct this time?

famouswaffles 3y ago
Current Capability would be the biggest one. We're at the point where any testable definitions of GI that the sota LLM fails (GPT-4) is also failed by a good chunk of humans. You couldn't say that a few years ago nevermind 60. What we have now (so no hypotheticals) coupled with the fact that scaling hasn't yet shown any performance walls makes a pretty good shout that things will probably be different this time.
- skepticATX 3y ago
  That's not really true though. LLMs are abysmal at planning, for example. Something that comes quite naturally to humans.
  - runsWphotons 3y ago
    They are probably better than 10% of people
    - danielmarkbruce 3y ago
      You meant 70%, right?
  - famouswaffles 3y ago
    They're really only abysmal if you attempt to one-shot it and probe with tasks that would require a human a scratchpad to accomplish. Humans can't one-shot non trivial planning tasks either. It's the one problem i have with all the papers that try to evaluate planning for LLMs. Step away from that approach and they're ok. https://innermonologue.github.io/ https://innermonologue.github.io/ https://tidybot.cs.princeton.edu/ https://tidybot.cs.princeton.edu/
  - pulvinar 3y ago
    I'm curious as to your source, or particular examples, since they (or at least GPT-4) seem to me to be rather decent at planning. E.g., for writing code.
    - climatologist 3y ago
      Ask it to write a backtracking sudoku solver with coroutines and/or fibers and let me know how it performs in your language of choice. We are nowhere near generally intelligent software systems.
      - jtmoulia 3y ago
        I was curious where GPT-4 would come up short on the problem and I was surprised -- it seemed to solve it pretty well whether or not using coroutines. (I dropped both solns into a python interpreter and both appeared to solve the problem.) There could def be bugs I missed tho. https://chat.openai.com/share/ef77507e-cb75-4112-97f1-a16cfc03cd98 https://chat.openai.com/share/ef77507e-cb75-4112-97f1-a16cfc...
        climatologist 3y ago
        That's a good attempt but the coroutine solution is incorrect. See if you can figure out why and how to improve it. You can also ask it to propagate constraints and see what happens.
        el_nahual 3y ago
        This isn't the dunk you think it is since GP is a human (I presume), and he thought the solution worked.
        jtmoulia 3y ago
        Sigh classic LLM -- without you, the expert, I can't quickly tell from the code / output how the answer the LLM produced is wrong. I also asked it to solve sudoku by "propagating constraints" and the answer seemed to work for me :/ Again, I'd guess the soln produced is wrong because I trust you more than the LLM but I don't have the mental horsepower to figure it out without resorting to tests & debugging.

29athrowaway 3y ago

Yann LeCun says human intelligence is not "general intelligence".

tru1ock 3y ago
We are not GI, and GI is too wasteful to be useful in a scarce environment.
esafak 3y ago
Is monkey intelligence or dolphin intelligence qualitatively different? What is general intelligence, then?

flangola7 3y ago

It utterly befuddles me that so many people still can't (or refuse to) sense what's coming. Be it bad or good the magnitude of what is now unfolding is beyond everything that homo sapiens have ever witnessed. Unless we think that MI advancement will miraculously stall and never improve beyond current capabilities, I cannot conceive of future, even a near future, that we recognize as real. >I think then we have some big waves that will wash over any observations we could make at this point in time. This is spot on. Exactly as how we didn't know what the internet would be, the wiser amongst us realized it was a new and strange era. "An alien lifeform. Unimaginable, both amazing and terrifying." - Bowie. MI will far eclipse the internet.

apsurd 3y ago
I don't know, that's quite the romantic take. I can easily admit I don't know what's coming, but _homo sapiens have ever witnessed_... _cannot conceive of near future that we recognize as real_. Come on, lol Imagine yourself in the movie Apocalypto and I'm using a Hollywood film on purpose here - the entire Universe of being as you see it all the sudden explodes its Universe with new Gods/demons different and more powerful _everything_ coming to kill you. And then there's the boring discovery of "we're just one planet in one solar system in one Galaxy... LLMs gonna be orders of magnitude beyond that (in)comprehension? I mean maybe it will be more powerful and maybe it will come to kill us, but that's not an inconceivable new story line. Funny that both of us can accuse one another of hubris.
- flangola7 3y ago
  Is this GPT-2 generated? I've tired reading it three times and I still can't make heads or tails of it.
  - apsurd 3y ago
    I see what you did there gpt-"2". at least you don't dislike enough to downvote. not sure what's not understandable. I think parent is overly hyperbolic. i get that chat GPT is unprecedented, but come on "impossible to conceive of near-term reality" ??? edit: oh! you're the parent. yeah i think you're wildly hyperbolic. There's more drastic historical precedents but of course the future is by definition going to be inevitably more inconceivable over a long enough arc, so I don't think of this as a debate. I just think you're being overly dramatic for effect.