8 votes

Inside the white-hot center of AI doomerism: Anthropic

16 comments

  1. [11]
    RobotOverlord525
    Link
    With the ubiquity of ChatGPT as the chatbot/AI, I thought it was interesting to see behind the curtain a bit on one of its competitors. Especially because the author was giving a somewhat critical...

    With the ubiquity of ChatGPT as the chatbot/AI, I thought it was interesting to see behind the curtain a bit on one of its competitors. Especially because the author was giving a somewhat critical look at the company.

    They worry, obsessively, about what will happen if A.I. alignment — the industry term for the effort to make A.I. systems obey human values — isn’t solved by the time more powerful A.I. systems arrive. But they also believe that alignment can be solved. And even their most apocalyptic predictions about A.I.’s trajectory (20 percent chance of imminent doom!) contain seeds of optimism (80 percent chance of no imminent doom!).

    It's not certain that chatbots really are the path to AGI, but if they are, it's interesting to see one of them being developed specifically with "AI safety" in mind. (I actually wrote an unpublished novella about that several years ago, so it was especially funny for me to see.)

    The "Constitutional AI" idea is also interesting, particularly for someone who grew up reading Asimov and his Three Laws of Robotics.

    6 votes
    1. [10]
      bioemerl
      Link Parent
      For these companies safety is just a fancy buzzword for corporate control. You shouldn't trust them and you shouldn't accept them having control over AI

      For these companies safety is just a fancy buzzword for corporate control. You shouldn't trust them and you shouldn't accept them having control over AI

      9 votes
      1. [9]
        skybrian
        Link Parent
        I can understand not trusting them, but what do you suggest as an alternative? They control what they built.

        I can understand not trusting them, but what do you suggest as an alternative? They control what they built.

        5 votes
        1. [8]
          bioemerl
          Link Parent
          From a commercial perspective? There's not much I can offer them. From my perspective? Keep it open source and keep people being able to do whatever they want with it. Having this technology...

          From a commercial perspective? There's not much I can offer them.

          From my perspective?

          Keep it open source and keep people being able to do whatever they want with it. Having this technology locked behind any companies is unacceptable and their little web end points that you have to use to access their models shouldn't be accepted by any consumers. We should demand that those models run on our local machines as well.

          4 votes
          1. [3]
            skybrian
            Link Parent
            If it's too dangerous for one company to have it, why isn't it also too dangerous for every company to have it? And every individual too?

            If it's too dangerous for one company to have it, why isn't it also too dangerous for every company to have it? And every individual too?

            5 votes
            1. [2]
              bioemerl
              Link Parent
              The exclusive access and control is the problem, not the fact they have access to an AI.

              The exclusive access and control is the problem, not the fact they have access to an AI.

              2 votes
              1. skybrian
                Link Parent
                Whether access restrictions are good or bad depends on what it is. Widespread access to something dangerous like, say, fentanyl, is widely considered a bad thing. We could think about a product on...

                Whether access restrictions are good or bad depends on what it is. Widespread access to something dangerous like, say, fentanyl, is widely considered a bad thing.

                We could think about a product on two scales: danger and usefulness. Access to AI might be important but there’s widespread disagreement on how dangerous it is and how useful it is.

                2 votes
          2. [4]
            moriarty
            Link Parent
            Even if you open source them, these models require tens of thousands of GBs of fast ram over hundreds of GPUs, not to mention the PB amounts of corpora to train on. It is well beyond the...

            Even if you open source them, these models require tens of thousands of GBs of fast ram over hundreds of GPUs, not to mention the PB amounts of corpora to train on. It is well beyond the capabilities of even very large enterprises to train these, let alone individuals.

            2 votes
            1. [2]
              skybrian
              Link Parent
              This is true of the big ones but not the small ones that run on a laptop. It's unclear how useful they are. People write about using them but I haven't seen rave reviews or anything. They will...

              This is true of the big ones but not the small ones that run on a laptop. It's unclear how useful they are. People write about using them but I haven't seen rave reviews or anything. They will probably improve somewhat, but nobody knows how much.

              2 votes
              1. moriarty
                Link Parent
                The inference would run on a laptop. Training the model (as in what you use the code for) would still require massive amount of infrastructure. I'm not really sure what you can do with the...

                The inference would run on a laptop. Training the model (as in what you use the code for) would still require massive amount of infrastructure. I'm not really sure what you can do with the inference model itself though, it's just a set of billions of weights.

                That said, there are plenty of transformer based models that have been open sourced. The architecture for ChatGPT is probably not significantly different - the probably added a ton of optimizations to make it run faster, cheaper and with less memory, but the foundation is likely the same as models that have been opened. The main difference is its size.

                4 votes
            2. bioemerl
              Link Parent
              Time will eventually render those models runnable by individuals.

              Time will eventually render those models runnable by individuals.

              1 vote
  2. [4]
    SpinnerMaster
    Link
    I have access to Claude and its aboslutely nowhere near as useful or helpful compared to ChatGPT.

    I have access to Claude and its aboslutely nowhere near as useful or helpful compared to ChatGPT.

    1. skybrian
      Link Parent
      Claude 2 is available to everyone now at claude.ai. (That probably has to do with the timing of this news article.) Is that the one you tried?

      Claude 2 is available to everyone now at claude.ai. (That probably has to do with the timing of this news article.) Is that the one you tried?

      3 votes
    2. moriarty
      Link Parent
      Really? In what way? What did you try to run on it? The newer Claude (1.3?) supposedly has a context length of 100k tokens (compared to ChatGPT's 8k?) so it should definitely be better at some...

      Really? In what way? What did you try to run on it? The newer Claude (1.3?) supposedly has a context length of 100k tokens (compared to ChatGPT's 8k?) so it should definitely be better at some tasks which require parsing huge documents.
      I haven't tried Claude 2 yet but it is rumored to be very good.

    3. RobotOverlord525
      Link Parent
      I haven't tried it yet, but from what little I played with ChatGPT, I struggle sometimes to think of a good use case for them. I don't trust the accuracy of these enough to depend on anything it...

      I haven't tried it yet, but from what little I played with ChatGPT, I struggle sometimes to think of a good use case for them.

      I don't trust the accuracy of these enough to depend on anything it tells me, so it has to be something quasi-creative. And when it comes to the realm of creativity, its output is, by nature, incredibly derivative.

      It did help my daughter come up with a backstory for her D&D character, though.