@drew @kevin I don't know what "Nazi" translates to in this context, but I'm hoping it is neither Lenovo nor Framework.

@ariadne @whitequark sigh... on-prem models. the false "democratization" bollock. alas, the stolen valour (epistemic injustice) remains... "Sorry doesn't put fingers back on the hand, Marge!" And who pays for the recurring RLHF training cycles, and probably steals more data to stave off the model collapse?

@ariadne @whitequark This is where I get to plug an old friend's website.
https://www.spamradio.com/index/welcome_to_spamradio.html

@ariadne The burden-of-proof, appeal-to-novelty proposition that LLMs might have emotions statistically approximates to Geoffrey Hinton "getting high on his own supply".

@androcat @CynAq @lritter @swoonie @juergen_hubert I'm sure it crept into the thread somewhere, but the key point here is UNCERTAINTY. It's a lemma, a stepping stone. And I invoke the example of "what we're calling AI this week, depending on who you speak to" to illustrate the point. The GOFAI systems did handle it to some degree, the DNNs differ significantly in the technological approach used to emulating human cognitive capabilities and expression. When the "reasoning" of which one writes is actually a tie-break between an embedded token-based mapping, that surely does not resemble the weighing up of facts, at least to my mind, which leads back to the mistakes perhaps on the part of the lay person who cannot deal with uncertainty. But we do it all the time "under the hood" at a pre-conscious level. Just typing these words: uncertainty, in the neuro-chemical impulse of the fingers and their sense-feeling, not data, which can only ever be an approximation to what some might term qualia. There's something missing from the DNNs, that's obvious, but no one is putting their finger on it, hence the burden of proof arguments which arise on both "sides".

@CynAq @androcat @lritter @swoonie @juergen_hubert Having danced the dance of hypothesis across multiple chapters, I can tell you: an LLM couldn't actually do my job. And "the entirety of human thought" of which you write here is what compels me to respond, for the people who "boost" such things seem to fundamentally misunderstand that the sum total of human knowledge cannot truly be known, nor is it even VERBAL to begin with! To claim that the sum total of human knowledge is accessible to LLMs is the epitome of ignorance.
Dealing with uncertainty is the very apotheosis of abductive reasoning, the foundation of C.S. Peirce's philosophical school of Pragmatism. Humans do this all the time. 80s and 90s era GOFAI expert systems were capable of it to a degree, albeit with pre-coded heuristics. GPTs and DNN based systems and so forth use very computationally aggressive and energetically wasteful statistical approximation, with "inference" as a post-processing of the enumeration of said embedding. https://addxorrol.blogspot.com/2025/07/a-non-anthropomorphized-view-of-llms.html
There are definitely lessons to be learned from how the neural-net systems often approximate what would otherwise be a product of human reason badly, and that's the point: they cannot be said to be reasoning if one engages in functional decomposition of such systems, and leaves the "deus ex machina" interpretation to one side. The principle of logical parsimony at the heart of abduction (abductive reasoning) and the use of philosophical razors compel this; they require judgement and weighing-up of facts, which statistical systems do not do by definition. This is why I do not believe that these systems are capable of reason.
Knowing these things makes for stronger "human general intelligence" which, to be frank, is far more effective. Not infallible, but when LLMs go awry, they hit the bricks.

@lritter @MadCowKastor @swoonie @juergen_hubert "Clankers" understand nothing. They are not capable of engaging in dialectic and are algorithmic constructs which enumerate a Hilbert-space-like embedding via a net-of-neural-nets.

If straw man detection were a thing, that would be a killer app. But for the unreliable imitative narrators, which statistically approximate the outputs at the mercy of the system prompts and RLHF, reason is not possible.