Apple: ‘Reasoning’ AIs fail hard if they actually have to think

The term "reasoning model" is as gaslighting a marketing term as "hallucination". When an LLM is "Reasoning" it is just running the model multiple times. As this report implies, using more tokens appears to increase the probability of producing a factually accurate response, but the AI is not "reasoning", and the "steps" of it "thinking" are just bullshit approximations.