Comments by "clray123" (@clray123) on "OpenAI’s new “deep-thinking” o1 model crushes coding benchmarks" video.

  1. AI is a 21st technology where you give it some rigorously-defined requirements which it will thoroughly and automatically reason about, to finally randomly fuck them up, so that your model checker written in 1986 can have some fun finding all the bugs.
    151
  2.  @TheFinalsTV  Oh yeah, but don't forget the managers will get replaced first, all according to your logic.
    45
  3.  @TheFinalsTV  Too bad defining accurate reward functions to train RL on is just as hard, if not harder, than actually implementing rule-based algorithms.
    34
  4.  @TheFinalsTV  Good luck with using the benchmark-beating bots as your developers, manager bro. I hope you have many benchmarks and math olympiads to implement.
    9
  5.  @TheFinalsTV  I said it based on the past 1.5 years of experience with AI marketing bullshit. Also my own experience with training LLM and RL models.
    5
  6.  @TheFinalsTV  Good luck with firing your employees then (if you have any). "They will remember that."
    3
  7. Looks like we have another fully automated new and improved benchmark solver.
    2
  8.  @3thinking  Ask yourself how many "stakeholders" you know who prefer to or even are interacting with the AI today? Then ask yourself: why do they shun interaction with the AI today? Once you have it figured out, you will also understand the reasons why they will not be interacting with the AI tomorrow or next year.
    2
  9. Chain-of-Thought withheld by OpenAI: 1. Holy fuk it's another puzzle invented by dim-witted AI reseachers 2. Lemme look it up in my big hash table of solutions real quick 3. Damn it is not there yet 4. Let's pretend that I am thinking about it 5. Meanwhile send it through function call to Autonomous Indians who like to handle such crap 6. ??? 7. Waiting for AI response.... 8. Got it! The solution is 42.
    1
  10. ​ @johndewey7243  No, for RL to work you have to define the final reward function, and specifying what it means for tests to pass correctly is just as hard as writing the damn code. You have probably not done anything with RL yourself in the past to know what it is like. Unfortunately, the AI is just as unable to write correct tests as it is unable to write correct code which makes them pass. Not to mention how inefficient it is to let a machine figure out what to change by trial and error (which RL boils down to if you don't have a good guiding reward function to dole out intermediate rewards, only the "final test passed" reward). Not to mention that this whole AI bullshit presupposes that all information to optimize on is already provided in some form (which may be the case for little benchmarks and math olympiad tasks). But in real life a great majority of information is not written down, but e.g. contained in heads of people who define the requirements AND ALSO if not mostly in heads of people who fulfill these requirements, based on their long-term experience in the domain and with the org they are working for. Unless you are talking about some junior code monkeys who you can hire and fire at will as replaceable parts. So once again good luck convincing these senior people to write everything down so that the AI can "grok" it. You're gonna have to pay for that, and those tasked with eliminating their own jobs in favor of AI are gonna sabotage your stupid efforts - because unlike AI they have brains to figure out what is strategically important for them and what you are trying to do to them. I predict that in the future it will not even work for artists and other low-key "creative" professionals who will refuse to work in certain setups or demand huge markup for "training" the imitating/copying/filling in the gaps algorithms (which is the only true "capability" of today's "AI"). Already the fact that you are falling for this AI scam (which is mostly targeted at pulling as much money from clueless investors as they can before the house of cards inevitably collapses, as it has with "AI" several times before) indicates that you might not be the brightest bulb out there. So again I wish you all the best with your efforts with outsmarting people who can actually do their jobs.
    1
  11. Yaya, bla, bla, we've heard it before.
    1