writing · llm research

frictions/attriti

Recurring ways language models fail, named and countered, from inside an extended writing and research project run as a standing testbed.I modi ricorrenti in cui i modelli linguistici cedono, nominati e contrastati, dall'interno di un esteso progetto di scrittura e ricerca condotto come banco di prova continuativo.

A model rarely fails loudly. It agrees, it returns something plausible, and the work drifts — three sessions later you are re-arguing a term that was settled weeks ago. Benchmarks cannot see this. On month three of a project, it is most of what you see.

Eight pieces, two originals each — EN and IT.
Licensed under Creative Commons CC BY-NC 4.0 · repository

It is the reason these phenomena are visible at all: most of them need weeks of accumulated context before they surface, which is exactly the condition an afternoon of testing cannot reproduce. It is also the limit of the evidence. This is one long case observed systematically, not a statistic, and nothing here is presented as a measurement.

Most of this work was done with Claude, and Claude is named wherever the episode is Claude's. Where the literature documents the same behaviour across vendors — positional degradation over long inputs, sycophancy under preference training, the limits of unaided self-correction — it is described as a family trait rather than one company's fault.

This inventory was produced by applying the protocols it describes. The observations, the analysis, the protocols and the final revision are mine; the drafting was done in dialogue with the systems under observation, under those same protocols — blind fields, forced refutations, pre-declared failure criteria.

articlessix

italianoenglish

italiano

I limiti della mia chat sono i limiti del mio mondo

Che cosa sopravvive alla fine di una sessione, che cosa no, e il documento di trasmigrazione che porta un progetto attraverso lo stacco.

english

The Limits of My Chat Are the Limits of My World

What survives the end of a session, what does not, and the transmigration document that carries a project across the gap.

italiano

Il modello vuole darsi ragione

Chiedi un test e ottieni un test che conferma l'ipotesi. Il campo cieco: come si progetta una verifica contro il modello anziché con lui.

english