Profile pic is from Jason Box, depicting a projection of Arctic warming to the year 2100 based on current trends.

  • 0 Posts
  • 10 Comments
Joined 1 year ago
cake
Cake day: March 3rd, 2024

help-circle
  • LLMs can be good at openings. Not because it is thinking through the rules or planning strategies, but because opening moves are likely in most general training data from various sources. It’s copying the most probable reaction to your move, based on lots of documentation. This can of course break down when you stray from a typical play style, as it has less to choose from in the options of probability, and only a few moves in there won’t be any more since there’s a huge number of possible moves.

    I.e., there’s no calculations involved. When you play a LLM at chess, you’re playing a list of common moves in history.

    An even simpler example would be to tell the LLM that its last move was illegal. Even knowing the rules you just told it, it will agree and take it back. This comes from being trained to give satisfying replies to a human prompt.







  • But he wasn’t. At least in the movie version, he and Banner had failed a few times, maybe more we didn’t see on screen. Something happened when Tony wasn’t there that sparked Ultron to become aware and catch Jarvis off guard. I’d give him credit for getting it 99% of the way there, same with Vision, but he didn’t make that final jump, it happened on its own.

    And Jarvis wasn’t AGI. Seems like it to us, but since Ultron was apparently the big moment of A(G)I in the MCU even with Jarvis being around all that time, he was just a very flexible and even self-aware scripting that would never do something on his own accord, only following Tony’s orders. I think even Ultron catches on to that in the brilliant few seconds of waking and realization with his “why do you call him Sir?”