Comments by "Mikko Rantalainen" (@MikkoRantalainen) on "Lex Clips" channel.

  1. If you see M1, look for better, right?
    2
  2. Some argue that having great grasp of trade-offs is the single most important sign of a great software architect. If we assume that's true, Carmack definitely achieves that criteria.
    2
  3. I find it interesting that we still don't truly understand what IQ even means. It seems that (some?) researchers are still thinking it only means doing the same thing faster when doing the task requires some amount of processing. However, I would consider true IQ as ability to more complex stuff than somebody with lesser IQ. That would require more and more progressively harder tests until the test subject no longer can complete any. I don't know if this would be called IQ in research or something else. There are mental tasks that less intelligent person cannot do, no matter how many hours you give to complete the test.
    2
  4. That's a great idea. Make a program that generates a copy of itself with some small amount of mutations and run multiple copies (like millions or billions) in an environment where a successful program gets more copies. The evolution will then automatically take care of the rest. The only big question is how much computing power would you require in total? Training biggest LLM systems like GPT-4 already require computing resources where the computation part alone costs 100 million USD.
    1
  5. Not even computers can evaluate the full game so we cannot know if the computer program actually evaluates the position correctly using heuristics (AI or otherwise). However, computers are already much stronger in Chess so we must simply trust the computer to be probably correct – or at least more correct than humans.
    1
  6. If it was simply about giving the turn to opponent, surely you would want to push only one square forward. The fact that it's a two square move, it must be about tempo. I would guess it's about tempo getting the pawn closer to promotion for the end game and it's still possible to defend the pawn if needed so it's not a gambit.
    1
  7. If you look at the position that Hikaru is talking about around 8:40, the issue is speficially the situation where king side knight has already moved over to other side of the board. That's what makes that move so weird to still make sense for a computer.
    1
  8. The most interesting question is when to start throwing more computing resources into it vs when to keep improving the algorithm. It seems that 175 billion weight system like ChatGPT is already big enough to be too hard to fully train – that is, if we could spend way more computing power and keep the same network and architecture, it would perform better. However, even trying to try that once might cost half a billion USD. Who is wealthy enough to try?
    1