Comments by "Gilad Barlev" (@GSBarlev) on "Firefox Has A PERFECT Use For AI Text Generation" video.

  1. I'd love to see this running in lynx.
    3
  2. Does inference happening in "a few seconds" on a laptop means maxing out the CPU for those few seconds? I'd assume so. Now let's say you're loading a page with 20 images. Is that a full minute of your processor going full bore?
    2
  3. It feels like this is missing a punchline like: "And now my system won't even boot!" But in all seriousness—I'm both really glad for you and super curious regarding your proompts—like, what were you asking it to help you do?
    2
  4. b) "A few seconds" per image of your processor maxing out can chew through a laptop battery pretty quick c) The image-to-text encoders are trained on CC datasets. GPT-2 is trained on CommonCrawl, which contains copyrighted works. Don't get me wrong—I'm mostly in favor of this feature being available. But I'm glad it's off by default, and, subject to accessibility workflows, it should only process images the user explicitly requests.
    1
  5. While I think these use cases are noble, I do want to caution that a model being "open source," "permissively licensed" or "open weight" is only half the issue—if the training dataset is not open as well, there's a near certainty that the models were trained on copyrighted material. It's amazing that the computer vision models were trained on widely available open datasets, but they're coupling that capability with GPT-2, which was trained on CommonCrawl, which just indiscriminately scraped the web. CommonCrawl claims their dataset constitutes "fair use," but time will tell whether that's legally defensible—let alone morally justified. There's also the very real fact that transformer models are taking system resources and draining your battery / driving up your power bill. If you're vision impaired, I'm sure that's a fair trade-off. But for other people who neither need nor want alt text, performing this inferences should be opt-in and disabled by default. Anyway, that's my 2¢. Ultimately, Mozilla is being a lot more ethical here than pretty much anyone else. But every proposed application of GenAI should be met with at least some degree of skepticism.
    1
  6. ​ @Poldovico Yes, for this specific use case, I agree. But I also understand the perspective that there is no ethical use of system built unethically, and, at the very least, it should be a conversation.
    1
  7.  @Poldovico  Really bad example, since the current medical consensus is that the results of those studies have little—if any—scientific value (the results and conclusions were wrong and likely set back our understanding by decades). Also pretty bold of you to trot out that example to someone with my pfp and last name, but 🤷‍♂️
    1
  8.  @Poldovico  Again, I agree with you about the specific. However, I still think it's something we should always keep in mind in order to ensure that the benefits outweigh the potential harms (which, again! in this case they do!)
    1