4 ms·

genuine question here: why is using deep learning to do something useful impossible? I can think of a bunch of potential use cases for gpt 3 alone. or do you

by hejja 6y ago

genuine question here: why is using deep learning to do something useful impossible? I can think of a bunch of potential use cases for gpt 3 alone. or do you mean its impossible to build useful models from scratch because all the "easy" problems are solved? this also seems like a limited mind set. context: I'm a ML noob

laichzeit0 6y ago
> why is using deep learning to do something useful impossible? This is completely false. Here are some examples: Google translate. It's infinitely useful. It's not perfect, but it's good enough for me when I want to quickly check whether my translation into my second language is okay, or I'm not sure I got the meaning right of some translation. Second example: My home security cameras now uses object classification and only alerts me when there's movement in "high risk zones" and it's human. So many stupid false positives of shadows and stray cats completely gone. I'm pretty sure I can fine-tune it with examples of myself and my family and it will ignore them when it's reasonably confident it's them, but I couldn't be bothered.
- alkonaut 6y ago
  Yes. Google Translate is obviously a useful product, and obviously Google will have thousands of areas where they can apply AI (enough that they'd even write frameworks for it). But 99(.99)% of us work in mundane jobs doing boring CRUD apps. My post was meant in this context. You work in a company making an intranet product for the paper tissue industry. Your manager wants v3.0 to have some AI in it. No one remembers the fiasco when "Cloud" was added in v2.0
  - hejja 6y ago
    so what you're saying is... you either swallow the CRUD app red pill, or you live long enough to become the "AI a la carte" manager
  - heavenlyblue 6y ago
    Google Translate was good before the so-called "AI revolution".
    - jtjbdhsjjdnd 6y ago
      As far as I remember it was close to unusable. Even most basic translation to languages besides German, Spanish and French was a nightmare. That was at lest true during my second year at the university (~2010-2011).
    - simonw 6y ago
      Google Translate today is absolutely spectacular. A fun game used to be "translate this phrase English to French, then translate it back again and laugh at how meaningless it's made by the round-trip". That doesn't work any more.
    - rrrrrrrrrrrryan 6y ago
      From what I've read, the early translations were really rough (but on par with competing free products), until they fired all their linguists and brought in AI guys.
      - thu2111 6y ago
        Not exactly. Google Translate was always a statistical approach from day one. That was the reason it was better than the rules engines that dominated before that. The Translate guys really pioneered statistical machine translation. However it was all done with hand-crafted statistical functions and code. The new stuff using deep learning is the first time it'd have been referred to as "AI".
fat-chunk 6y ago
Most of the "useful" tasks AI has helped us with are problems that are domain specific and lend themselves well to machine learning (for example object detection in images and machine translation as another poster mentioned). However, what I think the poster meant was that just tacking on AI for the sake of it to a problem which most likely doesn't have an applicable use for it (which currently is the case for most existing IT projects or apps) is doomed for failure.
huffmsa 6y ago
It's not impossible. But there is a very real gap between "this is cool in a research paper" and "this is deep learning that works in real life". It's a large gap, covering everything from application topics, to data quality, to the need to actually run the damn think in a production setting with scalability, availability, error handling, etc. Production applications of deep learning aren't particularly glamorous, they're not the "next big thing" right now. Rather they're improvements of existing applications. Google's on device live captioning works really well, but still somewhat niche, and requires special / higher end SoC's to run.
- hejja 6y ago
  makes sense. thanks. models I have used seem to have their usefulness greatly outweighed by performance demands. scaling and economics are another question entirely. Perhaps we were spoiled with democratized web tech and it's wishful thinking to want everything to be that.
  - huffmsa 6y ago
    It will get there eventually. A lot of it is hardware / deployment constrained. Search by image and object detection and computer vision in general is cool and potentially useful, but right now, it's cumbersome as fuck to pull out your phone, find the Lens application, take a picture etc. Needs to be baked into a wearable / neuralink type setup. But self driving applications of CV work because the cameras are always deployed and running. But the hardware is expensive.

by hejja 6y ago

laichzeit0 6y ago

> why is using deep learning to do something useful impossible? This is completely false. Here are some examples: Google translate. It's infinitely useful. It's not perfect, but it's good enough for me when I want to quickly check whether my translation into my second language is okay, or I'm not sure I got the meaning right of some translation. Second example: My home security cameras now uses object classification and only alerts me when there's movement in "high risk zones" and it's human. So many stupid false positives of shadows and stray cats completely gone. I'm pretty sure I can fine-tune it with examples of myself and my family and it will ignore them when it's reasonably confident it's them, but I couldn't be bothered.

alkonaut 6y ago
Yes. Google Translate is obviously a useful product, and obviously Google will have thousands of areas where they can apply AI (enough that they'd even write frameworks for it). But 99(.99)% of us work in mundane jobs doing boring CRUD apps. My post was meant in this context. You work in a company making an intranet product for the paper tissue industry. Your manager wants v3.0 to have some AI in it. No one remembers the fiasco when "Cloud" was added in v2.0
- hejja 6y ago
  so what you're saying is... you either swallow the CRUD app red pill, or you live long enough to become the "AI a la carte" manager
- heavenlyblue 6y ago
  Google Translate was good before the so-called "AI revolution".
  - jtjbdhsjjdnd 6y ago
    As far as I remember it was close to unusable. Even most basic translation to languages besides German, Spanish and French was a nightmare. That was at lest true during my second year at the university (~2010-2011).
  - simonw 6y ago
    Google Translate today is absolutely spectacular. A fun game used to be "translate this phrase English to French, then translate it back again and laugh at how meaningless it's made by the round-trip". That doesn't work any more.
  - rrrrrrrrrrrryan 6y ago
    From what I've read, the early translations were really rough (but on par with competing free products), until they fired all their linguists and brought in AI guys.
    - thu2111 6y ago
      Not exactly. Google Translate was always a statistical approach from day one. That was the reason it was better than the rules engines that dominated before that. The Translate guys really pioneered statistical machine translation. However it was all done with hand-crafted statistical functions and code. The new stuff using deep learning is the first time it'd have been referred to as "AI".

fat-chunk 6y ago

Most of the "useful" tasks AI has helped us with are problems that are domain specific and lend themselves well to machine learning (for example object detection in images and machine translation as another poster mentioned). However, what I think the poster meant was that just tacking on AI for the sake of it to a problem which most likely doesn't have an applicable use for it (which currently is the case for most existing IT projects or apps) is doomed for failure.

huffmsa 6y ago

It's not impossible. But there is a very real gap between "this is cool in a research paper" and "this is deep learning that works in real life". It's a large gap, covering everything from application topics, to data quality, to the need to actually run the damn think in a production setting with scalability, availability, error handling, etc. Production applications of deep learning aren't particularly glamorous, they're not the "next big thing" right now. Rather they're improvements of existing applications. Google's on device live captioning works really well, but still somewhat niche, and requires special / higher end SoC's to run.

hejja 6y ago
makes sense. thanks. models I have used seem to have their usefulness greatly outweighed by performance demands. scaling and economics are another question entirely. Perhaps we were spoiled with democratized web tech and it's wishful thinking to want everything to be that.
- huffmsa 6y ago
  It will get there eventually. A lot of it is hardware / deployment constrained. Search by image and object detection and computer vision in general is cool and potentially useful, but right now, it's cumbersome as fuck to pull out your phone, find the Lens application, take a picture etc. Needs to be baked into a wearable / neuralink type setup. But self driving applications of CV work because the cameras are always deployed and running. But the hardware is expensive.