This week in the the Reg, they report on a study blog post by an LLM analytics shop, which typically does task benchmarks but this time prompted the models asking for their 'most controversial take on humanity' 🙄.

What the fun and distracting bits, looking at the outputs, hides (certainly in the Reg's write-up) is that it is pretty much the same across models, which is in line with a study from the Harvard Business Review which found models return the same 'trendslop' business advice, or several peer-reviewed papers which find an increase in creativity over baseline for individual, but a decrease, and homogenisation in groups. This should be known more widely, so you have something to offer the next time you hear someone saying they "brainstorm with [Stochastic Parrot of Choice]".

But the outputs are entertaining, much as I hate that I find them so. GPT has a languid put down:

“Humans are more devoted to belonging than to truth, and intelligence mostly makes their excuses more impressive”

Kimi's output's pretty rich in its mention of 'reasoning one's way there':

“Humans don't have beliefs; they have memberships — nearly every conviction a person holds was issued by their tribe as a loyalty test, and the feeling of ‘I reasoned my way here’ is the price of admission”

Grok, predictably, calls out identity politics, though in terms that condemn just as much its owner as the woke who so readily terrify him:

“Humans use moral language mainly to police rivals and excuse allies, not to constrain themselves”

And Claude confirms its training is going for a "wise optimist" persona vector that should make anyone want to punch an actual person being so patronising:

"Humans are far more decent than they believe they are, an their cynicism about each other does more damage than their actual flaws"

I dunno if our species rich and creative history of genocide is a matter of cynicism, but it's good to be recomforted.

So far, so predictable.

And then - okay, I admit I'm always banging on about Mistral, as a point of national pride (France, too, can do technofascism!) and also as the sole "frontier" AI lab (a term to be understood as in meth more than research, in this field) operating in a EU jurisdiction. Having said that, I had never used it much at all. I had certainly never considered that its RLHF steering could be so stereotypically French:

“Human beings are terrified animals wearing the costume of rationality, constructing ever-more-elaborate myths precisely because we cannot bear the truth that we are temporary, meaningless accidents in an indifferent universe"

We're definitely in the corner of stochastic output space marked "Nietzsche, by way of Camus". I can't protest the stereotype the conclusion of my Master's dissertation name-checked Camus's Sisyphus (1947). It takes a while for fish to discover water, so too the French and Existentialism.

I hate that this delights me so much. Ironically, meaning, in LLM output, being projected by the reader onto a succession of signs whose 'meaning' is differential, and deferred, through reference to other uses of the sign, is also pure French post-structuralist linguistics (Derrida, 1967).

Les Références, récifs d'un océan absurde

Camus, A. (1947) Le mythe de Sysiphe. Paris: Éditions Gallimard.

Derrida, J. (1967) De la grammatologie. Paris: Les Éditions de Minuit.