3 ms·

LLMs may be stochastic parrots, but the critics of LLMs are starting to look like deterministic parrots. Do they know any other phrases besides “autocomplete on

by fwlr 4y ago

LLMs may be stochastic parrots, but the critics of LLMs are starting to look like deterministic parrots. Do they know any other phrases besides “autocomplete on steroids” and “stochastic parrot”? (The LLMs do: I asked ChatGPT for some sarcastic and dismissive phrases for language models and it gave me back “mindless mimics”, “algorithmic babblers”, “robotic regurgitators”, “synthetic chatterboxes”, and my personal favorite: “soulless scribble-bots”.)

AstralStorm 4y ago
Imaginationless idea-free implements. So, when will ChatGPT write a credible self-critical article? Oh wait, it really cannot. It can only transform text. The difference between writing and producing text does indeed exist.
- fwlr 4y ago
  It used to be someone would confidently say “computers can’t do this” and they could be right for 10 or 20 years before the state of the art caught up and proved them wrong. Now it’s like, “it can’t write a credible and coherent article“. Well, GPT3 does fine on paragraphs, and whole articles are literally the very next incremental step from that, so probably GPT4 will be able to do that. And when does that come out? Like next week?
  - ryanjshaw 4y ago
    Are you sure it can't write articles? I had it write an outline for an article, then an outline for each of those sections, then the content of those sections and it wasn't too bad.
    - shadowofneptune 4y ago
      Based on what's presented hete: https://acoup.blog/2023/02/17/collections-on-chatgpt/ https://acoup.blog/2023/02/17/collections-on-chatgpt/, it's quite good at producing the form of an article or an essay, but the actual thesis made is incoherent. If what you want is a content mill, it's great. A depressing amount of web articles already look like this with human authors. To be clear though, the generation of the format of communication, content being optional, has always been the goal and domain of chatbots. It's no big revolution.
danaris 4y ago
I think this is in large part because they are reacting against the breathless hype that LLMs are either already sentient themselves, and thus we need to start giving them rights and/or being terrified that they're coming for us, or they're the last step before that happens (with basically the same conclusions, just time-shifted slightly).
- norwalkbear 4y ago
  I think they are in reacting from a place of fear. Writers and liberals arts are under the automation threat for the first time ever.
  - danaris 4y ago
    Not meaningfully—at least, not from LLMs. Sure, companies might be able to have ChatGPT and its brethren spit out marketing copy and similar types of writing that are, to a large extent, about chasing the lowest common denominator. But even if we grant for the sake of argument that it's possible to train an LLM that can create, say, a brand-new "Shakespeare" play, without enough input and tweaking that it will take a writer or a liberal arts major to do it anyway, that day is a long, long way off. No; the only people I've heard talking about how all the knowledge jobs will be automated away by ChatGPT (or, more commonly, just all the "unproductive" humanities-type jobs—because, of course, the humanities are something we can totally hand over to LLMs with no negative consequences) are more tech-oriented people with vastly inflated ideas of ChatGPT's abilities and the tech sector's primacy in the world. The humanities people I've talked to about the subject tend to fall either into the "this is a bunch of BS hype" camp or the "I've watched too many sci-fi movies and think Skynet is launching tomorrow" camp.
    - pixl97 4y ago
      Eh, this is just because people are in general terrible at nuance. We don't have AGI yet, so no Skynet isn't here, but that doesn't mean we shouldn't watch out and make sure it doesn't show up and that we have a regulatory framework to ensure Skynet doesn't happen. And, no liberal arts majors are not going to be replaced, but that doesn't mean they are coming out unscathed either. Things like GPT are going to force further specialization as we offload the generalized work off to tools like this. Now, humanity has been doing this for a long time so this is nothing new, but specialization is risky, even a small change in technological capability and the thing you trained your entire life for has been replaced by a small shell script and you're left trying to figure out how to pay rent next month.

by fwlr 4y ago

AstralStorm 4y ago

Imaginationless idea-free implements. So, when will ChatGPT write a credible self-critical article? Oh wait, it really cannot. It can only transform text. The difference between writing and producing text does indeed exist.

fwlr 4y ago
It used to be someone would confidently say “computers can’t do this” and they could be right for 10 or 20 years before the state of the art caught up and proved them wrong. Now it’s like, “it can’t write a credible and coherent article“. Well, GPT3 does fine on paragraphs, and whole articles are literally the very next incremental step from that, so probably GPT4 will be able to do that. And when does that come out? Like next week?
- ryanjshaw 4y ago
  Are you sure it can't write articles? I had it write an outline for an article, then an outline for each of those sections, then the content of those sections and it wasn't too bad.
  - shadowofneptune 4y ago
    Based on what's presented hete: https://acoup.blog/2023/02/17/collections-on-chatgpt/ https://acoup.blog/2023/02/17/collections-on-chatgpt/, it's quite good at producing the form of an article or an essay, but the actual thesis made is incoherent. If what you want is a content mill, it's great. A depressing amount of web articles already look like this with human authors. To be clear though, the generation of the format of communication, content being optional, has always been the goal and domain of chatbots. It's no big revolution.

danaris 4y ago

I think this is in large part because they are reacting against the breathless hype that LLMs are either already sentient themselves, and thus we need to start giving them rights and/or being terrified that they're coming for us, or they're the last step before that happens (with basically the same conclusions, just time-shifted slightly).

norwalkbear 4y ago
I think they are in reacting from a place of fear. Writers and liberals arts are under the automation threat for the first time ever.
- danaris 4y ago
  Not meaningfully—at least, not from LLMs. Sure, companies might be able to have ChatGPT and its brethren spit out marketing copy and similar types of writing that are, to a large extent, about chasing the lowest common denominator. But even if we grant for the sake of argument that it's possible to train an LLM that can create, say, a brand-new "Shakespeare" play, without enough input and tweaking that it will take a writer or a liberal arts major to do it anyway, that day is a long, long way off. No; the only people I've heard talking about how all the knowledge jobs will be automated away by ChatGPT (or, more commonly, just all the "unproductive" humanities-type jobs—because, of course, the humanities are something we can totally hand over to LLMs with no negative consequences) are more tech-oriented people with vastly inflated ideas of ChatGPT's abilities and the tech sector's primacy in the world. The humanities people I've talked to about the subject tend to fall either into the "this is a bunch of BS hype" camp or the "I've watched too many sci-fi movies and think Skynet is launching tomorrow" camp.
  - pixl97 4y ago
    Eh, this is just because people are in general terrible at nuance. We don't have AGI yet, so no Skynet isn't here, but that doesn't mean we shouldn't watch out and make sure it doesn't show up and that we have a regulatory framework to ensure Skynet doesn't happen. And, no liberal arts majors are not going to be replaced, but that doesn't mean they are coming out unscathed either. Things like GPT are going to force further specialization as we offload the generalized work off to tools like this. Now, humanity has been doing this for a long time so this is nothing new, but specialization is risky, even a small change in technological capability and the thing you trained your entire life for has been replaced by a small shell script and you're left trying to figure out how to pay rent next month.