Here is long-time reader TB’s e-mail response to what I said yesterday. (TB works in AI and/or adjacent fields; I didn’t press him on his employment details.) I told TB that I need a better understanding of what goes on ‘under the hood’ of these amazing engines to be sure that I am not ‘blowing smoke’ when I assert that they are not truly intelligent but merely simulate human intelligence. I have bolded a couple of key sentences and made some minor edits. It is wonderfully clear and — as far as I can judge — accurate.
……………………………..
You are 100% correct in your post in my view. I read Nagel’s most recent defence of the reality of first person consciousness a few years ago. He could have made his argument in about half the words, but he is right. Dennett and friends claim that consciousness is an illusion; there is a theory that the mind plays this trick whereby all stimulus from without gets converted to a sort of realtime stream of imagery / impressions that we mistake for consciousness. This ‘illusion’ argument seems to forget what ‘illusion’ means – first person experience. Each person’s illusion is his own. And then we realise that calling it an ‘illusion’ is meant to make us dismiss it as not real. So I think your closing comments are exactly right – the deniers want to argue that the first person experience we really are having is a) really there (something has to be there, since something is fooling us into thinking there is something there), but it’s an illusion, so ‘not real’, while simultaneously telling us b) that it (i.e. conscious experience) is not really there at all.
A few things on AI. You have undoubtedly understood that there is something called a Large Language Model (LLM). LLMs are trained on the whole internet, and in addition vast numbers of (generally illegally obtained) PDFs of physical books that the AI companies obtain, destroy the spine, and scan. All of this is the ‘training base’. Each sentence of every document, web page etc, is ‘tokenised’ and each each token is ’embedded’ in a high dimension vector space, such that concepts that frequently co-occur or appear to be used as near synonyms in similar documents, end up as vectors that are close, i.e. whose normalised mathematical projection (cosine) is close to 1.0.
Tools like ChatGPT work as follows: they take your prompt, prepend a bunch of other info, including things it has learned about you from prior use (name=vallicella; professional philosopher; tradition = analytic / anglo / German; location = Phoenix AZ, . . . etc, plus a summarised version of prior queries), and it returns text that it scores highest as being an appropriate response (i.e. ‘prediction’ of what is likely to come next in this kind of exchange). There is other functionality wrapped around the core, that will prevent it getting into profanity, discussions of the kind that KKK or Hamas types might enjoy, and also pro-woke censorship.