Comments by "Vitaly L" (@vitalyl1327) on "Two Computer Scientists Debunk A.I. Hype with Arvind Narayanan and Sayash Kapoor - 281" video.

  1. LLMs won't "crash". They already overdelivered. The NLP problem is solved (which nobody expected to happen in their lifetime). Semantic search problem is solved. Guided proofs problem is solved. Code synthesis problem is solved. If that is not enough for you, you have some severe issues with your cognitive capacity.
    9
  2.  @ang3111  my point is that you have no idea what you're yapping about. I am a robotics engineer. Using LLMs for path planning very efficiently now - this problem was nearly unsolvable before, and now it became trivial. Not to mention hundreds of other ground breaking use cases for LLMs. I suspect you're some kind of a web coder. It is funny when you lot call yourself "engineers".
    7
  3.  @ang3111  you're not a software engineer.
    2
  4. Use local models then. Fine-tune them on your own data. There is hardly any justification for cloud-based inference.
    2
  5. ​ @sharkbelly1169 imagine being a web code monkey and having an audacity to have opinions on things so far above your pay grade...
    1
  6.  @ang3111  if your job title.says you're an engineer, it does not necessarily mean you're an engineer. FAANG especially, they hired a lot of useless trash. Since you're so obviously incompetent, it is clearly your case.
    1
  7. ​ @ang3111 my condolenses to the poor FAANG company that employs such "engineers".
    1
  8.  @ang3111  this field was never ML-first, up until recent advancements. Explainable, physically-based algorithms were always the main requirement. And LLMs allowed to bridge the gap between fuzzy hand-wavey ML world and explainable / certifiable certainty that industrial robotics needs. And if LLMs work perfectly here, they must work even better in simpler and less restricted domains. I use LLMs to generate formal proofs for the mission-critical code. It was very tedious previously, now fully automated. I use vLLMs to design mechanical parts with a CAD. I use LLMs to generate and debug in simulation good enough circuits that I can later refine manually. It saves me weeks or even months of work. I was never so productive. Literally every area of my work is improved by LLMs. And yes, of course I am not using raw ChatGPT (or any other cloud-hosted model).
    1
  9.  @ang3111  again, even if in this hard field AI shines (and mind you, explainability is a must, so it is really a hard sell), in all others it outperformed humans by far.
    1
  10.  @SoulyG  you have no faintest idea of what you're talking about. You don't understand AI.
    1
  11. @SoulyG  LLM is a semantic search engine. I.e., the last remaining piece of puzzle we needed to build an AI. Once again, you have no faintest idea of what you're talking about. LLM itself cannot do logical reasoning. An LLM armed with Prolog or an SMT solver is reasoning better than any human.
    1
  12. @SoulyG  PS a very simple example to counter your "specific task" point - LLMs write a much better code in programming languages they do not know. If the language was not in a train set, an LLM cannot just generate learned patterns and havr to resort to thinking step by step, querying the documentation on its way.
    1
  13.  @TookAHikeNowWhat  I'm using local models exclusively, with an RTX-4080, small models (7B, 11B) are super fast.
    1