4 ms·

I guess things are slowing down at DeepMind. I have tremendous respect for David Silver and his work on AlphaZero and Richard Sutton as a pioneer in RL. But the

I guess things are slowing down at DeepMind. I have tremendous respect for David Silver and his work on AlphaZero and Richard Sutton as a pioneer in RL. But the cynic in me is that this paper is just a result of Goodhart's law with publishing count as a metric. Any proof of the type of emergent behaviors that they mention from RL with an actual RL experiment would go a long way. Showing an RL agent developing a language would be extremely interesting. It makes me think they tried to show these emergent behaviors but could not and thus ended up with a hypothesis.

maiodude 5y ago
They just "solved" protein folding late last year. How can you say things are slowing down? Do you honestly expect life-changing discoveries every other week?
- 0xB31B1B 5y ago
  Protein folding is a well modeled math problem. The alpha fold solution is extremely good at pruning (aka guessing) folding chain structure possibilities. I am Impressed and this is a difficult problem but this is extremely different from AGI as this is a well scoped easily modelable problem that is basically a chain of 26 inputs types of links of arbitrary length. I am not trying to take away that the protein folding is incredible but AGI is extremely different. AGI is literally having a model that can both do alpha fold and self driving cars, as well as the ability to generate novel models to solve new well scopes problems. RL can do 0 to 1, the 1 to n (generalizeability) is the extremely difficult part.
  - ackbar03 5y ago
    I may be really only speaking for myself here but I have very sincere doubts that anyone who has done any moderately serious work on "AI" and wrangled with the nitty gritty details of it all is really having any huge expectations for AGI. I mean if it comes during my lifetime, hurrah! But personally I'm not gonna sit around waiting for it or depend on it for anything. That being said, is the current AI tech as we have it useless? Of course not. Things like protein folding and alphago are still huge leaps forward in tech, it'd be kind of silly to treat AGI as the only thing worth achieving
  - tinco 5y ago
    It is entirely not necessary for an AGI to be able to drive a car. Frankly, after seeing AlphaZero and AlphaFold I'm surprised they didn't declare AGI right there and then. People assume that when AGI happens, computers can suddenly outsmart humans in every way and solve every problem imaginable. The reality is just that it could in theory given enough time and resources. It is like quantum computing. In theory it can instantly factor and break our nice cryptographic primes. In reality the largest number it factored is 21.
    - croon 5y ago
      > It is entirely not necessary for an AGI to be able to drive a car. "Artificial general intelligence (AGI) is the hypothetical[1] ability of an intelligent agent to understand or learn any intellectual task that a human being can." What definition are you using?
      - tinco 5y ago
        The same, that they understand and can learn how to drive a car doesn't mean they would actually be able to do it in the real world. You can read a book on how to hit a ball with a baseball bat, you can even practice and get good at it, but that still doesn't mean you would actually be able to hit a ball thrown by a professional pitcher.
        croon 5y ago
        The interface to the car is a solved problem. > You can read a book on how to hit a ball with a baseball bat, you can even practice and get good at it, but that still doesn't mean you would actually be able to hit a ball thrown by a professional pitcher. If I hade incredibly fast reflexes and actuators I could.
        tinco 5y ago
        Similarly, DeepMind's software might be able to drive a car, would it have a similar neuron count, connectivity, perception systems and training you received. Or maybe it couldn't, because the software is not as efficient as the organisation of your brain is. Or because there's hardcoded routines evolved in your brain that it lacks. What I'm saying is that just that because an AGI can't drive a car, it doesn't mean it's an AGI. For the same reason there's loads of people out there that are generally intelligent that can't drive cars for all sorts of physical reasons.
uyt 5y ago
It might be a stretch but some people say that the weights learned by a neural network is somewhat like a language. For example if you look at the weights of a random middle layer it would seem like gibberish. Much like how aliens would react when looking at humans making gibberish noises (aka talking) to each other. In both cases they are just compressing signals based on learned primitives.
- mirker 5y ago
  Not sure if there is any case where this thought is useful. The only thing this says is the primitives are correlated and we don’t understand them. It’s similarly not useful to think about atoms “talking” when they exchange heat.
ChicagoBoy11 5y ago
Very tangential, but as someone who has gotten into the Game of Go because of their pioneering project in that space, I'm exceptionally grateful -- that alone had a very significant and positive impact on my life, and I can tell that in that entire community it was a watershed moment as well.

maiodude 5y ago

They just "solved" protein folding late last year. How can you say things are slowing down? Do you honestly expect life-changing discoveries every other week?

0xB31B1B 5y ago
Protein folding is a well modeled math problem. The alpha fold solution is extremely good at pruning (aka guessing) folding chain structure possibilities. I am Impressed and this is a difficult problem but this is extremely different from AGI as this is a well scoped easily modelable problem that is basically a chain of 26 inputs types of links of arbitrary length. I am not trying to take away that the protein folding is incredible but AGI is extremely different. AGI is literally having a model that can both do alpha fold and self driving cars, as well as the ability to generate novel models to solve new well scopes problems. RL can do 0 to 1, the 1 to n (generalizeability) is the extremely difficult part.
- ackbar03 5y ago
  I may be really only speaking for myself here but I have very sincere doubts that anyone who has done any moderately serious work on "AI" and wrangled with the nitty gritty details of it all is really having any huge expectations for AGI. I mean if it comes during my lifetime, hurrah! But personally I'm not gonna sit around waiting for it or depend on it for anything. That being said, is the current AI tech as we have it useless? Of course not. Things like protein folding and alphago are still huge leaps forward in tech, it'd be kind of silly to treat AGI as the only thing worth achieving
- tinco 5y ago
  It is entirely not necessary for an AGI to be able to drive a car. Frankly, after seeing AlphaZero and AlphaFold I'm surprised they didn't declare AGI right there and then. People assume that when AGI happens, computers can suddenly outsmart humans in every way and solve every problem imaginable. The reality is just that it could in theory given enough time and resources. It is like quantum computing. In theory it can instantly factor and break our nice cryptographic primes. In reality the largest number it factored is 21.
  - croon 5y ago
    > It is entirely not necessary for an AGI to be able to drive a car. "Artificial general intelligence (AGI) is the hypothetical[1] ability of an intelligent agent to understand or learn any intellectual task that a human being can." What definition are you using?
    - tinco 5y ago
      The same, that they understand and can learn how to drive a car doesn't mean they would actually be able to do it in the real world. You can read a book on how to hit a ball with a baseball bat, you can even practice and get good at it, but that still doesn't mean you would actually be able to hit a ball thrown by a professional pitcher.
      - croon 5y ago
        The interface to the car is a solved problem. > You can read a book on how to hit a ball with a baseball bat, you can even practice and get good at it, but that still doesn't mean you would actually be able to hit a ball thrown by a professional pitcher. If I hade incredibly fast reflexes and actuators I could.
        tinco 5y ago
        Similarly, DeepMind's software might be able to drive a car, would it have a similar neuron count, connectivity, perception systems and training you received. Or maybe it couldn't, because the software is not as efficient as the organisation of your brain is. Or because there's hardcoded routines evolved in your brain that it lacks. What I'm saying is that just that because an AGI can't drive a car, it doesn't mean it's an AGI. For the same reason there's loads of people out there that are generally intelligent that can't drive cars for all sorts of physical reasons.

uyt 5y ago

It might be a stretch but some people say that the weights learned by a neural network is somewhat like a language. For example if you look at the weights of a random middle layer it would seem like gibberish. Much like how aliens would react when looking at humans making gibberish noises (aka talking) to each other. In both cases they are just compressing signals based on learned primitives.

mirker 5y ago
Not sure if there is any case where this thought is useful. The only thing this says is the primitives are correlated and we don’t understand them. It’s similarly not useful to think about atoms “talking” when they exchange heat.

ChicagoBoy11 5y ago

Very tangential, but as someone who has gotten into the Game of Go because of their pioneering project in that space, I'm exceptionally grateful -- that alone had a very significant and positive impact on my life, and I can tell that in that entire community it was a watershed moment as well.