How Far Can AI Reason? The limits of AI exposed

Поділитися
Вставка
  • Опубліковано 31 жов 2024

КОМЕНТАРІ • 35

  • @pubfixture
    @pubfixture 10 днів тому +3

    So, a Marvel movie.

    • @code4AI
      @code4AI  8 днів тому

      Nice idea. The most complex thought of humans (USA), that provided the training data for AI pattern recognition, is a Marvel movie? Imagine the training data would come from India or China? How would it differ?

  • @MoreThanLoveHeatherDRichmond
    @MoreThanLoveHeatherDRichmond 10 днів тому +5

    The Architect, The Devourer, and The Observer are the same being. :-)

    • @code4AI
      @code4AI  8 днів тому +1

      If it would only be that easy ...

  • @rousabout7578
    @rousabout7578 10 днів тому +2

    Philosophical AI questions and discussions, interesting. I can see people will like this. Perhaps one channel for the technical and scientific side of AI, another channel for philosophy and human interest discussions...

  • @Jorn-sy6ho
    @Jorn-sy6ho 10 днів тому +2

    I find it philosofically sound. A good basis for further theory.

  • @OccamsPlasmaGun
    @OccamsPlasmaGun 10 днів тому +2

    LLMs suffer from Dunning Kruger like the rest of us. Did really well until it tried to introduce new logic by hand-waving away a paradox. I don’t like anthropomorphizing LLMs, but that's a very human response.

    • @TheRealUsername
      @TheRealUsername 10 днів тому +1

      Am I the only one who thinks these models have no "intelligence" at all ? Not even o1 which is a reasoning structure retriever rather than a "reasoner", I will change my mind when Transformer stops to be a "lazy" N-gram statistical predictor and is able to do out-of-distribution pattern extrapolation.

    • @preston_is_on_youtube
      @preston_is_on_youtube 10 днів тому

      @@TheRealUsername what does that scenario look irl and as a youtube commentary? asking for a friend

    • @mattmmilli8287
      @mattmmilli8287 10 днів тому

      @@TheRealUsernamenot yet they don’t. For programming you can basically pair up with it and explore and create novel ideas though which is really neat

  • @heffpeople
    @heffpeople 10 днів тому +1

    Great job Felix!

  • @ariaden
    @ariaden 10 днів тому +1

    Interesting example on what humans may accept as "complex" (as in well-defined but hard to comprehend).
    Example of an actually complex thing: Organizational structure of European Union.
    Example of a problem too complex for current LLMs to solve correctly (without multi-round prompting): Multiply two 4-digit hexadecimal numbers 0x21df * 0x2497 = ?

    • @code4AI
      @code4AI  8 днів тому

      An Texas Instrument calculator can solve your complex task. Which is just a tool to an Agent, therefore solvable.

    • @ariaden
      @ariaden 8 днів тому

      ​@@code4AI For humans, the task solvable with pen and paper, so it should be solvable for LLMs even without tools.
      But, have you seen this task actually solved by a current LLM (with or without tools but without specialized prompting) yet?
      If not, I posit that my task proves the answer to "How Far Can AI Reason?" question is "it depends, sometimes not that far".

  • @DevinDTV
    @DevinDTV 9 днів тому

    You forgot that LLMs just predict language. This isn't actually at the fringe of its abilities. This is just what its algorithms determined would be an appropriate human response to the question.

    • @code4AI
      @code4AI  8 днів тому +1

      No, it does not predict language, it only predicts the next token ...

  • @JellySword8
    @JellySword8 4 години тому

    This just sounds like video game development...

  • @markc5025
    @markc5025 10 днів тому +1

    Great video!!! Anything self-referential gives me headache!

  • @pensiveintrovert4318
    @pensiveintrovert4318 10 днів тому +1

    Here is the most complex problem. How can you tell the difference between the normal LLM hallucination(aka B.S.) from a profound problem?

    • @carlhealy
      @carlhealy 10 днів тому +1

      The same way we determine the veracity of our own hallucinations, the scientific method.

    • @pensiveintrovert4318
      @pensiveintrovert4318 10 днів тому

      @@carlhealy seems like the whole point of making LLMs do the work becomes moot.

    • @carlhealy
      @carlhealy 10 днів тому

      @@pensiveintrovert4318 The objective is enabling AI to do the same type of epistemic reasoning that we humans employ to understand the world around us. The AI should be able to justify ideas with evidence and sound reasoning. That is how we separate the BS from the profound.

  • @MusingsAndIdeas
    @MusingsAndIdeas 10 днів тому

    Orthogonal rules means that the axioms governing the Devourer are completely independent of the axioms of Epsilon.

    • @MatthewKelley-mq4ce
      @MatthewKelley-mq4ce 10 днів тому

      Presumably they aren't completely independent though.

    • @code4AI
      @code4AI  8 днів тому

      Independent across all time frames in the multiverse?

    • @MusingsAndIdeas
      @MusingsAndIdeas 8 днів тому

      @@code4AI I mean, orthogonality by definition means they are independent. Across all time frames? I'm confused, do you mean across all the ways we could slice them up? I don't think it was ever assumed that the metrics would even have just one that differed in sign.
      To answer your question, that is an undecidable statement. The definition of orthogonality is based off of axioms that here are self-referential, so there's no way to decide.

  • @AnnonymousPrime-ks4uf
    @AnnonymousPrime-ks4uf 9 днів тому

    This is still inconsistent. The observer was supposed to regulate both the actions of the architect and devourer. If the observer is only relative it can't regulate the actions of both at the same time. So it doesn't provide a consistent solution to the paradox.

    • @code4AI
      @code4AI  8 днів тому +1

      Well , you have there an assumptions, that limits your reasoning. Think about theoretical physics the wave - particle dualism and its mathematical formulae.

  • @brendawilliams8062
    @brendawilliams8062 10 днів тому

    4:00 is it that spin glass has it’s bigger obstacle with I chemistry

  • @ravivegan5442
    @ravivegan5442 10 днів тому +1

    doesn't it sound like incorrect version of hindu philosophy? why incorrect ? because it attempts to mix bible and matrix too in hindu philosophy. while bible in half truth and information wise matrix contains inaccurate information.
    the architect can be vishnu, observer can be bramha and devouver can be shanker the destroyer.

    • @code4AI
      @code4AI  8 днів тому +1

      I never thought about religion and AI pattern before ...

  • @Zbezt
    @Zbezt 10 днів тому +1

    Hahahahahaha so basically is it saying itll troll the speculative equilibrium in a sense of neoism thats perfect XD atleast it understands that it can be corrected as the neccesity arises

  • @luiztomikawa
    @luiztomikawa 10 днів тому +1

    Lol elbow_salami