4 ms·

People are already a blackbox algorithm (to some degree). I think if there was personal accountability and an appeals process (which there needs to be to deal w

by ratorx 4y ago

People are already a blackbox algorithm (to some degree). I think if there was personal accountability and an appeals process (which there needs to be to deal with human shittiness anyway), then it makes no difference right? The key thing is that if the domain is appropriately fuzzy, and a human vs non-human decision is indistinguishable, then it shouldn’t matter. If it isn’t, then you don’t really need the AI. I think the key thing here is that people should not be able to blame the AI for bad decisions. Responsibility should be on whoever decided to delegate the decision.

izacus 4y ago
But that's the catch - people are known to be unreliable so there's usually (some) recognition of that in good systems. AI is being trusted blindly without human review where it's used right now - there's noone second guessing Google account bans, PayPal bans and YouTube bans when the computer says "no". And this attitude will spread.
- raincole 4y ago
  There is another issue I've rarely seen people mentioned on HN. Humans, especially those come from different background, have different kinds of biases, blind spots and agendas. It's actually a good thing. It means you can have a pair of humans, such as a human writer and editor, a human researcher and a reviewer, or two independent human doctors, and they usually make less errors than just one human. Does using two AIs (like ChatGPT + Bard) makes less error than just one AI? Are there data for this?
  - akasakahakada 4y ago
    This is legitimately the so called ensemble model. We can spawn one CbatGPT instance to make decision, and spawn 3 instances to ask questions and validate the decision. Since the power of ChatGPT enable it to mimic different characters, basically this ensemble can do anything.
  - RugnirViking 4y ago
    anecdotally, a technique I have seen spreading quickly is asking the AI to simulate a conversation between multiple characters of different types (an oxford professor talking to an ambitious student) on a particular topic you want to ask it about. It lets it question itself a lot better, including sometimes going down one thought process then deciding to go in another direction.
    - x86x87 4y ago
      Yes, yes! Even more "AI". That's the solution! The old age garbage in, garbage out will finally be defeated!
      - RugnirViking 4y ago
        have you ever used an LLM? It's clear that certain ways of asking for things have vastly better outputs. I'm describing one such way of doing that, as a way to answer the question the parent asked. I'm making no grand commentary at all on whether this technology is suitable for peer review or whatever.

by ratorx 4y ago

izacus 4y ago

But that's the catch - people are known to be unreliable so there's usually (some) recognition of that in good systems. AI is being trusted blindly without human review where it's used right now - there's noone second guessing Google account bans, PayPal bans and YouTube bans when the computer says "no". And this attitude will spread.

raincole 4y ago
There is another issue I've rarely seen people mentioned on HN. Humans, especially those come from different background, have different kinds of biases, blind spots and agendas. It's actually a good thing. It means you can have a pair of humans, such as a human writer and editor, a human researcher and a reviewer, or two independent human doctors, and they usually make less errors than just one human. Does using two AIs (like ChatGPT + Bard) makes less error than just one AI? Are there data for this?
- akasakahakada 4y ago
  This is legitimately the so called ensemble model. We can spawn one CbatGPT instance to make decision, and spawn 3 instances to ask questions and validate the decision. Since the power of ChatGPT enable it to mimic different characters, basically this ensemble can do anything.
- RugnirViking 4y ago
  anecdotally, a technique I have seen spreading quickly is asking the AI to simulate a conversation between multiple characters of different types (an oxford professor talking to an ambitious student) on a particular topic you want to ask it about. It lets it question itself a lot better, including sometimes going down one thought process then deciding to go in another direction.
  - x86x87 4y ago
    Yes, yes! Even more "AI". That's the solution! The old age garbage in, garbage out will finally be defeated!
    - RugnirViking 4y ago
      have you ever used an LLM? It's clear that certain ways of asking for things have vastly better outputs. I'm describing one such way of doing that, as a way to answer the question the parent asked. I'm making no grand commentary at all on whether this technology is suitable for peer review or whatever.