Raw LLM Responses

I'm only about 11 minutes in, but I'm a bit skeptical about some of what this guest is saying. At 11:37 he says "This- You could call this the very beginning of reasoning, they're called reasoning models; philosophers can bicker all day about wether it's real reasoning" Except that's clearly not true?? You don't have to be a philosopher; that's simply just not what's happening. It's disingenuous to even imply that it could be considered 'reasoning'; these models are not capable of introspection, and they have no internal frame of mind or logic process; they're still just predictive text output models; albeit with some trickery. 'Reasoning' models simply work by prompting themselves for input using custom alignment, training, or prompting before generating a reply to the user's query/input. It's not much different from you posting a query to a 'non predictive' LLM with an appendded section at the end saying "before you answer, expand on my question". You're still getting purely predictive output; they just filter the initial prompt through a sub layer of pre-emptory predictive output before giving the the 'primary' (second) output to the user. On one hand this can 'enrich context', but on the other hand it could potentially increase hallucination because it's increasing the amount of predictive text output needed per query. To simplify this explanation, some models allow you to toggle on visibility of 'reasoning'. It often starts with a phrase like "hmm, I should expand on this query; the user is asking about (whatever). It would be important to clarify what they mean (ECT, ECT)" This is not reasoning: it's still predictive text output. The only difference is the model is turned inwards first, with training and alignment being used to guide the model to generate an output *approximating* waxing on the input before generating a secondary output. There's an important distinction here between actual reasoning and what it's doing, because it can still hallucinate, and it still lacks actual awareness. TLDR; No, they're not actually reasoning; they're using a trick where the model is internally tuned to give two outputs in return for every query, with the first being pretended with the instructions "generate an output *as if* you were thinking, then use that to reply to the user", and then the second output is what the user gets. There's no special sauce or magic; it's still all trained predictive text output. No logic, no reasoning.