Comments by "" (@obsidianjane4413) on "Apple DROPS AI BOMBSHELL: LLMS CANNOT Reason" video.

However. The LLMs map the patterns of human reasoning, because their training data reflects it. So while its not "novel" reasoning. Its a very close approximation that can pass Turing Tests.
1
BoominGame Well for that you don't use a LLM, you use a ML. LLMs are just (supposed to be) human-machine interfaces after all. True AGIs will probably be a synthesis of specialist ML algos, like human vision, hearing, and bodily housekeeping are all sub-processes below the conscious mind.
1
 BoominGame Machine Learning algorithms. Same neural net model can be used to train AIs to do things besides talk like a parrot.
1
Yet.
1
@7:00 The GSM8K is a bit of a red herring because LANGUAGE models are weak in calculating. So its not at all surprising. You are basically asking a English Lit. student to do maths. Its not at all an indicator of a LLM's ability to "reason". OTOH: No of course they don't "reason", ie; "think". But. The LLMs map the patterns of human reasoning, because their training data reflects it. So while its not "novel" reasoning. Its a very close approximation that can pass Turing Tests. @18:00 "AI systems can't fail when doing stuff" No one is going to ask an LLM to fly a plane or be a QC over a production line. Those are totally different machine learning systems. LLMs will always be "squishy" and inaccurate because human language is squishy and impresise.
1