From David Bentley Hart today – a comment on the nature of LLMs, and a criticism of the sort-of "analytically propositional thinking" that gives rise to the impoverished conclusions of thinkers like Richard Dawkins, whose methods all-but oblige him to conclude that Claude is "conscious."
This is a paid Substack. I'm sharing only a small selection here. I highly recommend subscribing, if you find this kind of thing interesting.
"LLMs in general seem to me an obvious occasion for remarking not only the vast difference between mental agency and machine function, but also a certain tacit metaphysical presupposition in the practice of analytically propositional thinking that too often condemns it to the status of a secondary and ultimately inadequate methodological predilection rather than a serious style of rational reflection. Viewed from the perspective solely of the Turing test, there is no need to discriminate between what might be called tokens of genuine semantic meaning with intentional content and tokens simply of syntactic order and statistical frequency. As a system of cumulative autoregressive data, an LLM is a kind of ouroboros, consuming its own tail in a perpetual cycle; it contaminates virtual space continuously with its own aimless generation of textual echoes and then continuously ingests and metabolizes what it has produced into an ever vaster collection of ‘information’, in the purely Shannon-bits sense of that word, which means into an ever more errant flow of misinformation. This may prove to be an incorrigible defect of the system precisely because there is no intentionality there that will ever have the power to grasp such words as 'truth' or 'falsehood' as actual values rather than as mere periodic syntactic constants. There is no act of thinking in an LLM, of necessity but also on principle, since meaning is precisely what must be set aside in order for the system to process its data. There is not even anything that might plausibly be called a linguistic act in its functions, since intentional content obviously cannot be quantified in terms of statistically measurable syntactic recurrence. The system merely exploits the predictable sequences of that recurrence, setting it off against a background of absolute entropic randomness.
In the end, the appearance of semantic intention is one that arises out of total semantic vacuity, which is why much of the virtual world will soon be AI talking to AI—or not talking at all, in fact, but merely appearing to do so—since the organic substrate of all that psychopathic babble will have become superfluous. In that sense, LLMs have produced the most rigorous analytic reduction of language possible precisely by evacuating it of meaning and converting it entirely into formal and structuralist rules of order."