I am being systematically cancelled on HN (as if I am a fucking russian – I am NOT – or something), but sometimes it feels ridiculous.
LLM don’t “do maths” by definition (which in this context is the code). LLM “predict” the next symbol after “2+2=” and separately and independently (in principle) predict token-by-token of why that is true.
There is absolutely no “reasoning” or “understanding” of any kind whatsoever, just as a parrot would do.
Also this:
The ridiculous part is that everything I said here is factually correct:
- the artifact (the actual data structure) has been trained to output estimated probabilities (they call them “predictions”) with deliberately adding noise to avoid overfitting.
- the process of querying this “structure” (they call it “prompting”) is based on selecting a most probable next token, again, with some randomness to sometimes select the second best “path”.
- the process of “tuning” the model is just re-assigning (updating) probabilities to seemingly correct paths, while there is no way to determine correctness or even adequacy.
- one more time – there is no way to determine the correctness within a model (in principle) since it requires a separate, different kind of a process – the kind of an experimental verification or of a mathematical prof – which isn’t available (and can’t be) in the code or in the data anywhere.
The point is that the process of validation of any knowledge is a higher-level to a mere information processing, and information alone cannot, in principle, be the source of Truth. Truth requires a validation, which has been known since the beginning of time.
Now, dear HN, please, kindly fuck off.