4 ms·

Geoffrey Hinton Has a Hopeful Plan for Keeping Future AI Friendly

11 points by danboarder 3y ago

ChatGTP 3y ago
“Some people think, hey, there's this ultimate barrier, which is we have subjective experience and [robots] don't, so we truly understand things and they don’t,” says Hinton. “That's just bullshit. Because in order to predict the next word, you have to understand what the question was. You can't predict the next word without understanding, right? Of course they're trained to predict the next word, but as a result of predicting the next word they understand the world, because that's the only way to do it.” When Geoffrey says they understand the question, are these models really understanding the question or isn't it just "transforming" in the sense that some type of lookup is happening where it says, "given something that looks like this, I should say x next". Is he just arguing that he believes humans do the same thing? Personally I don't think we do this because somehow we now we're wrong long before we give an answer, even when people give false answers? People give false information out in order to try feel important, be helpful or to avoid being redundant etc.
- ETH_start 3y ago
  The range of questions it can answer is far beyond the number of permutations it can be trained on. The generalizability can only come from genuine intelligence, and it shouldn't be surprising that it has this given artificial neural networks are modelled after biological ones which we can attest produce understanding/intelligence.
  - ChatGTP 3y ago
    I’m not sure this is correct at all. For example if “genuine intelligence(c)” is what it is about, the why do these LLMs have more trouble with languages than English ? Imo the permutations it’s exposed too would matter here. What about hallucinating? It’s just giving you the closest statistical match which is wrong in some cases, not knowing it doesn’t know.
    - ETH_start 3y ago
      Can you elaborate on your point about languages other than English? How does the underperformance of LLMs when dealing with said languages suggest my point about genuine intelligence is wrong? What do you think is the implication of this disparity in comprehension/expression between languages? Hallucinations to me suggest instances of low intelligence and poor heuristics. ChatGPT is far from perfect. It is not perfectly generalizable. But when ChatGPT works well, and doesn't hallucinate, it is demonstrating clear signs of intelligence.
      - ChatGTP 3y ago
        Can you elaborate on your point about languages other than English? Honestly I couldn’t be bothered, this topic has been covered already a few times this week.

11 points by danboarder 3y ago

“Some people think, hey, there's this ultimate barrier, which is we have subjective experience and [robots] don't, so we truly understand things and they don’t,” says Hinton. “That's just bullshit. Because in order to predict the next word, you have to understand what the question was. You can't predict the next word without understanding, right? Of course they're trained to predict the next word, but as a result of predicting the next word they understand the world, because that's the only way to do it.” When Geoffrey says they understand the question, are these models really understanding the question or isn't it just "transforming" in the sense that some type of lookup is happening where it says, "given something that looks like this, I should say x next". Is he just arguing that he believes humans do the same thing? Personally I don't think we do this because somehow we now we're wrong long before we give an answer, even when people give false answers? People give false information out in order to try feel important, be helpful or to avoid being redundant etc.

ETH_start 3y ago
The range of questions it can answer is far beyond the number of permutations it can be trained on. The generalizability can only come from genuine intelligence, and it shouldn't be surprising that it has this given artificial neural networks are modelled after biological ones which we can attest produce understanding/intelligence.
- ChatGTP 3y ago
  I’m not sure this is correct at all. For example if “genuine intelligence(c)” is what it is about, the why do these LLMs have more trouble with languages than English ? Imo the permutations it’s exposed too would matter here. What about hallucinating? It’s just giving you the closest statistical match which is wrong in some cases, not knowing it doesn’t know.
  - ETH_start 3y ago
    Can you elaborate on your point about languages other than English? How does the underperformance of LLMs when dealing with said languages suggest my point about genuine intelligence is wrong? What do you think is the implication of this disparity in comprehension/expression between languages? Hallucinations to me suggest instances of low intelligence and poor heuristics. ChatGPT is far from perfect. It is not perfectly generalizable. But when ChatGPT works well, and doesn't hallucinate, it is demonstrating clear signs of intelligence.
    - ChatGTP 3y ago
      Can you elaborate on your point about languages other than English? Honestly I couldn’t be bothered, this topic has been covered already a few times this week.