3 ms·

Companies are lying about what these LLMs are capable of, full stop. That is the problem. And the reason for that is because they can't justify their costs if a

Companies are lying about what these LLMs are capable of, full stop. That is the problem. And the reason for that is because they can't justify their costs if all they've created is a really good chat bot and writing assistant. The responsible thing to do would be to have these LLMs refuse to respond if asked informational questions, but that would kill interest. So they won't. They're going to keep trying to shove it into applications that could potentially generate recurring revenue, and it's going to keep failing. I suspect they will go the way of voice assistants.

falcor84 4y ago
Why should we hold AIs to a higher standard than humans. The internet (and the world at large) is full of people confidently spewing misinformation - should we go ahead and censor all of them?
- phailhaus 4y ago
  Huh? What I'm telling you is that this is not "AI", it is a LANGUAGE MODEL. It models language. People literally think that these chatbots have access to some sort of "knowledge base" or are capable of "looking up information", but they can't. It's not a matter of "misinformation", these companies are trying to trick you into thinking they have something that they do not. And the idea that we shouldn't hold our tools to a higher standard is a joke. It's like if you sold me a screwdriver as a hammer and then went "wow, nobody's perfect, ok so we should just BAN all humans that are bad at hammering?!"
  - falcor84 4y ago
    >...this is not "AI", it is a LANGUAGE MODEL. It sounds like we have a semantic disagreement here. As I understand it, a language model is a mathematical construct, while a particular implementation of a language model as a software system is an AI. At the least it's as much an AI as a Chess Engine, right? > It's like if you sold me a screwdriver as a hammer and then went "wow, nobody's perfect, ok so we should just BAN all humans that are bad at hammering?!" That doesn't sound quite analogous to me. My scenario would be that for a particular hammering job, I can hire either a human or a hammering AI. If the hammering AI is imperfect, but still better and/or cheaper than the typical human worker, (at least as a cold-hearted business manager) I'd prefer the AI.
    - phailhaus 4y ago
      The reason I'm drawing a distinction between "language model" and "AI" is because when you say "AI", it means absolutely nothing. Like you just pointed out, a chess engine also qualifies as "AI", but you can't use a chess AI to write you a marketing copy. So what is the difference? Obviously, you would object if someone tried to tell you that this chess AI can answer your questions. And I'm telling you that LLM's can't be used for educational purposes because they're chatbots. >My scenario would be that for a particular hammering job, I can hire either a human or a hammering AI This is why it's 100% analogous: you don't have a hammering AI. You have a screwdriver AI and the salesman is telling you that you can use it as a hammer! These companies are trying to get you to use LLM's for things they are not built to do, because it's hard for the layman to tell the difference. It's a language model. It has no knowledge base. It doesn't even understand "true" from "false". It only generates plausible-sounding text, but it learned from a massive corpus of human-generated text, so it seems intelligent.

falcor84 4y ago

Why should we hold AIs to a higher standard than humans. The internet (and the world at large) is full of people confidently spewing misinformation - should we go ahead and censor all of them?

phailhaus 4y ago
Huh? What I'm telling you is that this is not "AI", it is a LANGUAGE MODEL. It models language. People literally think that these chatbots have access to some sort of "knowledge base" or are capable of "looking up information", but they can't. It's not a matter of "misinformation", these companies are trying to trick you into thinking they have something that they do not. And the idea that we shouldn't hold our tools to a higher standard is a joke. It's like if you sold me a screwdriver as a hammer and then went "wow, nobody's perfect, ok so we should just BAN all humans that are bad at hammering?!"
- falcor84 4y ago
  >...this is not "AI", it is a LANGUAGE MODEL. It sounds like we have a semantic disagreement here. As I understand it, a language model is a mathematical construct, while a particular implementation of a language model as a software system is an AI. At the least it's as much an AI as a Chess Engine, right? > It's like if you sold me a screwdriver as a hammer and then went "wow, nobody's perfect, ok so we should just BAN all humans that are bad at hammering?!" That doesn't sound quite analogous to me. My scenario would be that for a particular hammering job, I can hire either a human or a hammering AI. If the hammering AI is imperfect, but still better and/or cheaper than the typical human worker, (at least as a cold-hearted business manager) I'd prefer the AI.
  - phailhaus 4y ago
    The reason I'm drawing a distinction between "language model" and "AI" is because when you say "AI", it means absolutely nothing. Like you just pointed out, a chess engine also qualifies as "AI", but you can't use a chess AI to write you a marketing copy. So what is the difference? Obviously, you would object if someone tried to tell you that this chess AI can answer your questions. And I'm telling you that LLM's can't be used for educational purposes because they're chatbots. >My scenario would be that for a particular hammering job, I can hire either a human or a hammering AI This is why it's 100% analogous: you don't have a hammering AI. You have a screwdriver AI and the salesman is telling you that you can use it as a hammer! These companies are trying to get you to use LLM's for things they are not built to do, because it's hard for the layman to tell the difference. It's a language model. It has no knowledge base. It doesn't even understand "true" from "false". It only generates plausible-sounding text, but it learned from a massive corpus of human-generated text, so it seems intelligent.