Comments by "miraculixx" (@miraculixxs) on "Internet of Bugs" channel.

  1. I agree. Keep telling people exactly this. Surprisingly the same people who would never take as little as 1% failure from any process are suddendly very lenient towards GAI. The say things like "oh, we shouldn't expect too much from a new technology, even if we can't get 100% reliability". The problem often is that people think this is just the beginning on a path to perfection.
    28
  2.  @michaelbarker6460  interesting approach using multiple models, thanks. I doubt though that it adresses the problem, fundamentally having multiple models agree just means its a majority decision. If 2 of 3 models think the answer is A, 1 model thinks the answer is B, and B is correct, the end result is A and it is still wrong. Re. RAG, that also doesn't really fix hallucinations it just adds more complexity, that is hallucinations can now come from a subset of one if your documenta, or even a mix of many. No guarantees for correctness. I think if the use case is information retrieval, i.e. find the most relevant documents and parts there in, well do that. No LLM needed (vector db / search engine will do), perhaps using templated responses.
    4
  3. Thank you. Fully agree
    1
  4. 2:35 Wait how did you screenshot that code and it came out as editable in your IDE?
    1
  5. Thank you for this. More people need to see this.
    1
  6. 4:55 That's a pretty good summary of how outsourced codes have been coming back from the far eastern parts for 20+yrs. There is code, there are tests, they pass, except none of it is actually useful. Management loves it because it's cheap and fast.
    1
  7. Good for you. It is no proof however that Claude can code in general, it just shows that in your specific case you found a way to use Claude as a tool to get the outcome you wanted. That's great and shows you're really smart. Re Claude, News at 5.
    1
  8. 💯
    1