OK, … another tiny little idea that may enhance the understanding of what today’s ‘AI’ actually implements, reverse engineering interpretation-wise:
A proper definition of a Context for your Gen tooling turns out to be a partial collection of most fuzzy (since text-based and bereft of formal logic) Expert System clause statements. Think LISP for B1-B2 language levels. Or Prolog but #notafan somehow I never got the hang of that one; difficult to find the logic of the logic.
Yes, that’s quite an elliptic description but still: Isn’t context-writing (including output style definition and whatever one does in vain to prevent slop and hallucinations of the deepest undertow kind) a tie-down of permissible sources for (long-chain) token matching à la (long-list! i.e., numerous-argument-) expression/function argument matching-by-exclusion ..? A slight upgrade of SHRDLU that is looking on from around the corner,
at the loss of about-all exactness and logic – where we see that the failings (of understanding the potential due to non-fundamental tech limitations) of earlier systems have been overcome but at what cost… The baby of ultra-powerful machines capable of firmly logic-based consciousness, now lost to intellect-destroying babble.
As if The 2nd of Thermodynamics is made of silicon now. All Information that humankind built, contra that, Information equalling Energy in any science model, now being destroyed; turned into uninformative hence useless waste Slop, fake ‘facts’ and zero-logic reasoning. Wetware humans are getting dumber when they come into contact with it.
Ah yes, the Author’s rampaging against the dumbing-down of the world as we knew it.
But ..! There’s a positive aspect still to be found: Be absolutely radically totalitarian with your Context definition so your tools will be able to work only on seriously tied-down subject-matter, relevant source info. This may lead to both lack of precision in data when certain data elements or -relations are weakly defined in the historian data set(s) and enhanced precision when actually real relations are (possibly) established in a causal sense, being able to reduce noise by stipulating exact relations.
Induction; wasn’t that the original aim? From there, the ‘laws’ that are found, can be fed back into the context; the Expert System builds itself and can than be questioned as an intelligent, fuzzy-language assistant-to the human operator. Staying away from, staving the fractallucinistics while building better Einstein/Abe tooling.
For once, I do have Hope. No Cash, no trump, but whatever. And we all have:
Showing how ‘developments’-over-time are absolutely worthwhile refinements.
