4 ms·

The first AI winter was because of a book about Perceptrons, was perceived as being extremely critical of them, which killed interest. It took decades for neur

The first AI winter was because of a book about Perceptrons, was perceived as being extremely critical of them, which killed interest. It took decades for neural nets to finally prove their effectiveness. The second was because the benefits of "Expert Systems" were oversold, and didn't prove to be as beneficial as expected. There will be a heavy correction in the market caps of companies that bet heavy on AI, as the promises don't live up to the hype. However, in this case, there clearly is benefit to using LLMs, the cost of training has proven to be prohibitive. Slower and more prudent approaches, with better hardware will allow progress to continue. If nothing else, it'll be a mild AI winter.

pdimitar 2y ago
> However, in this case, there clearly is benefit to using LLMs Legitimate curious question: what are the clear benefits exactly? I mean stuff that objectively improves society, not the plaything that ChatGPT has mostly become (there are exceptions to this as well, of course, I had a CFO telling me lately he used ChatGPT to find mistakes or bad contract clauses).
- mikewarot 2y ago
  Uses for LLMs and Neural Nets in general include: Searching documents in discovery, sentiment analysis, summarization, flushing out ideas, 1st level support, writing code, doing speech->text, text->speech, translation. There are many things we can do now that were previously expensive or cost prohibitive. This is about the same as the computerization of records. No longer to we have to pay people to keep rooms full of files in order to aid retrieval and long term record storage. --- but I see you're asking about non-cost things We can now read documents that are translated from other languages, lowering those barriers to learning about other cultures. It's now easier to interact with the information available to all via the internet. It also makes some evil things easier, but that comes with every technology.
  - pdimitar 2y ago
    OK, I can partially agree with some of the points but... writing code? Any tangible proof on that one? I have spoken with many colleagues who were adamant that the LLMs can at best provide some boilerplate generation and nothing much beyond that. One or two mentioned it could find a bug in their Python code. > This is about the same as the computerization of records. No longer to we have to pay people to keep rooms full of files in order to aid retrieval and long term record storage. Eh, have you read the "Bullshit Jobs" book? It's a social and economic problem, rarely has been a technological one... TL;DR people want headcount, not efficiency. :/
    - FergusArgyll 2y ago
      > I have spoken with many colleagues who were adamant that the LLMs can at best provide some boilerplate generation and nothing much beyond that. Why don't you try it out yourself, eh? there are a lot of free options
      - pdimitar 2y ago
        Because my time is valuable, dude, and having to use an entire weekend -- likely much more than that even -- is not very feasible. And because I trust my circle of programmer acquaintances, they are extremely pragmatic folk who usually gets called to fix messes left by juniors, wannabe, poseurs and "AI" enthusiasts. Still, I am not entirely close-minded on "AI" but I don't think it's unfair to say that if it was that good then our area would have been hugely disrupted by now. And it does not seem like it is.
    - setopt 2y ago
      > writing code? Any tangible proof on that one? Personal experience? Not just me, but e.g. all the comments on HN on ChatGPT, Copilot, et al.? I’ve been using it for about two years now to draft new code, after having programmed for two decades. It provides value to me in that some tasks that previously would take me a whole day now can be done in a couple of hours, maybe one hour of throwing ball with ChatGPT and then one hour of refactoring the code myself. I have plenty of colleagues who have similar experiences, especially when using it for tasks that can be “tricky to write but easy to verify” (e.g. visualizations). Just keep in mind that (i) the paid version is miles ahead of the free version for coding, and (ii) using an AI assistant is like any other skill, it requires some training to learn what tasks they’re good at and how to talk to them.
- Yawrehto 2y ago
  I don't use them that much, but a relative does, and they've told me that they like using one that can do recipes (they have a household with very specific food restrictions) and another one. I don't remember what it's called, but it sounds vaguely like Magic School Bus, and it's good for doing forms and rewriting things to be less, say, sweary (they're a grader and the stories they tell about students submitting stupid things...)
  - pdimitar 2y ago
    Ohhh, those are actually awesome use-cases, thanks for bringing them up to my attention!
kypro 2y ago
> There will be a heavy correction in the market caps of companies that bet heavy on AI, as the promises don't live up to the hype. This might be your point, but I think in most cases the disappointment will be entirely financial. Some, perhaps even most of the tools and products being developed are pretty cool, but in the majority of cases they're basically just AI wrappers or nice to haves. It's like launching an email service or a website that tells you the current time in different countries in the late 90s or something. Email might be great, and your service might be really popular, but that doesn't mean there's any money to be made because anyone can implement POP3. Same with a website that tells you the current time in different countries – nifty tool, but not a real profit making business.
- meiraleal 2y ago
  It will be really interesting if the for-profit OpenAI fail then making the nonprofit OpenAI achieve its initial purpose.

pdimitar 2y ago

> However, in this case, there clearly is benefit to using LLMs Legitimate curious question: what are the clear benefits exactly? I mean stuff that objectively improves society, not the plaything that ChatGPT has mostly become (there are exceptions to this as well, of course, I had a CFO telling me lately he used ChatGPT to find mistakes or bad contract clauses).

mikewarot 2y ago
Uses for LLMs and Neural Nets in general include: Searching documents in discovery, sentiment analysis, summarization, flushing out ideas, 1st level support, writing code, doing speech->text, text->speech, translation. There are many things we can do now that were previously expensive or cost prohibitive. This is about the same as the computerization of records. No longer to we have to pay people to keep rooms full of files in order to aid retrieval and long term record storage. --- but I see you're asking about non-cost things We can now read documents that are translated from other languages, lowering those barriers to learning about other cultures. It's now easier to interact with the information available to all via the internet. It also makes some evil things easier, but that comes with every technology.
- pdimitar 2y ago
  OK, I can partially agree with some of the points but... writing code? Any tangible proof on that one? I have spoken with many colleagues who were adamant that the LLMs can at best provide some boilerplate generation and nothing much beyond that. One or two mentioned it could find a bug in their Python code. > This is about the same as the computerization of records. No longer to we have to pay people to keep rooms full of files in order to aid retrieval and long term record storage. Eh, have you read the "Bullshit Jobs" book? It's a social and economic problem, rarely has been a technological one... TL;DR people want headcount, not efficiency. :/
  - FergusArgyll 2y ago
    > I have spoken with many colleagues who were adamant that the LLMs can at best provide some boilerplate generation and nothing much beyond that. Why don't you try it out yourself, eh? there are a lot of free options
    - pdimitar 2y ago
      Because my time is valuable, dude, and having to use an entire weekend -- likely much more than that even -- is not very feasible. And because I trust my circle of programmer acquaintances, they are extremely pragmatic folk who usually gets called to fix messes left by juniors, wannabe, poseurs and "AI" enthusiasts. Still, I am not entirely close-minded on "AI" but I don't think it's unfair to say that if it was that good then our area would have been hugely disrupted by now. And it does not seem like it is.
  - setopt 2y ago
    > writing code? Any tangible proof on that one? Personal experience? Not just me, but e.g. all the comments on HN on ChatGPT, Copilot, et al.? I’ve been using it for about two years now to draft new code, after having programmed for two decades. It provides value to me in that some tasks that previously would take me a whole day now can be done in a couple of hours, maybe one hour of throwing ball with ChatGPT and then one hour of refactoring the code myself. I have plenty of colleagues who have similar experiences, especially when using it for tasks that can be “tricky to write but easy to verify” (e.g. visualizations). Just keep in mind that (i) the paid version is miles ahead of the free version for coding, and (ii) using an AI assistant is like any other skill, it requires some training to learn what tasks they’re good at and how to talk to them.
Yawrehto 2y ago
I don't use them that much, but a relative does, and they've told me that they like using one that can do recipes (they have a household with very specific food restrictions) and another one. I don't remember what it's called, but it sounds vaguely like Magic School Bus, and it's good for doing forms and rewriting things to be less, say, sweary (they're a grader and the stories they tell about students submitting stupid things...)
- pdimitar 2y ago
  Ohhh, those are actually awesome use-cases, thanks for bringing them up to my attention!

kypro 2y ago

> There will be a heavy correction in the market caps of companies that bet heavy on AI, as the promises don't live up to the hype. This might be your point, but I think in most cases the disappointment will be entirely financial. Some, perhaps even most of the tools and products being developed are pretty cool, but in the majority of cases they're basically just AI wrappers or nice to haves. It's like launching an email service or a website that tells you the current time in different countries in the late 90s or something. Email might be great, and your service might be really popular, but that doesn't mean there's any money to be made because anyone can implement POP3. Same with a website that tells you the current time in different countries – nifty tool, but not a real profit making business.

meiraleal 2y ago
It will be really interesting if the for-profit OpenAI fail then making the nonprofit OpenAI achieve its initial purpose.