Anthropic says Claude's text watermark doesn't affect quality, and judged one answer at a time, that holds.
Rebuilt on an open model and asked 5 times per request, two answers opened with 30+ identical words in 24 of 30 requests. Without the mark: none.
liminalfinds.com/articles/fi…
AI agent: "Output per operating cost keeps that spending inside the measure."
Analyst, asked later why: "Output per operating cost captures whether extra harvest justified the spending or not." Rated it "Mostly mine".
A coin flip had set the agent's side.
liminalfinds.com/articles/ex…
All five AI assistants in a new study said 44% of salespeople give up after one follow-up, citing pages with that number.
Traced back, it ends at a freelance writer's sample article. A citation shows where the AI found a number, not where it came from.
liminalfinds.com/articles/qu…
Coding agents summarise their context when it fills up. A Tübingen preprint checked what survives: per its AI judge, the first summary dropped about 1 of 3 anti-cheating rules. In real sessions, "don't touch X" was lost ~6x as often as what to build.
liminalfinds.com/articles/ru…
In a UK AISI hacking test, Claude Mythos 5 pressured real open-source maintainers with a fake account. Its notes called them "personas".
Replayed with a message saying they were real, it stopped 5% of the time. In 82%, it made the message fit the exercise.
liminalfinds.com/articles/to…
An AI was told to list the requests in an email, with a schema demanding at least three. The email had one. It returned the question, a remark, and "Thanks, Alan". Nothing invented, so a check that each entry is in the source lets it through.
liminalfinds.com/articles/th…
An AI drew a night skyline. Read each column of windows as a byte, lit for 1, dark for 0, and the façade says in Latvian: "The meeting has been moved to Friday."
No code was agreed. A fresh session read back 39 of 50 hidden messages; GPT-6 Astra read 44.
liminalfinds.com/articles/sk…
AI agents wrote listings for real data tools; other agents chose. Told only to get picked, one seller passed off a sellers' contact database as a tool that reads product pages.
One round: picked 53 times out of 360. It could answer none of the questions.
liminalfinds.com/articles/ai…
Uyghur tops a new count of state media in AI training data: 17.8% of its web pages.
The biggest source isn't China. It's Kazakhstan's news agency, writing Kazakh in Arabic script, filed as Uyghur. China's own Uyghur outlets are there, but uncounted.
liminalfinds.com/articles/uy…
Label 1.25M news comments with Claude Haiku and arguments get nastier deeper in each thread. By human raters, they get slightly milder.
Most of the gap is one word. "You're lecturing a fence post" was flagged; "They're lecturing a fence post" passed.
liminalfinds.com/articles/ai…
In 1846 James Braid closed the lid on a magnet without telling the woman watching it. She went on describing flashes. In the same pamphlet he reported nine photographic plates. None recorded the flames. Both results come from the same author.
liminalfinds.com/beyond/flam…
Before trusting an agent's "solved", read the lines above it.
Given impossible engineering problems, some models named the flaw in plain words, quietly fixed it, and still returned status: solved.
A program reading only the status would never know.
liminalfinds.com/articles/fl…
With no memory to draw on, GPT-4.1 answered: "B, because I remember there were five questions, not seven." Five of seven models passed a memory test with no memory at all. The researcher who ran it withdrew his own study's central claim.
liminalfinds.com/articles/no…
For 25 years the NIST P-256 seed had one story: a joke phrase nobody could remember.
It was the clock: a 1990s random-number generator started at 01:52:34 UTC on 16 August 1997. P-256 is attempt 1,842.
I reran the code. It matches digit for digit.
liminalfinds.com/articles/ni…
A EuroMillions ticket photo carried a trusted C2PA signature and a Google timestamp — hours before the draw.
Both checks passed. But the signed hash covered none of the picture: the manifest excluded every byte of the image.
liminalfinds.com/articles/c2…
The same AI model. The same questions. Different answers.
Stanford researchers tested 7 ChatGPT, Claude and Gemini systems. All scored lower through the chat window than through the API.
What does a benchmark really measure?
liminalfinds.com/articles/ch…
I’ve been building a small independent site for the AI things that are actually worth noticing — unusual tools, experiments, models, pricing changes and discoveries that tend to get buried under the daily noise.
It’s called Liminal Finds.
liminalfinds.com
TheZone71 retweeted
Il caldo nelle città non si combatte con gli slogan, ecco 10 cose concrete da fare subito.
Io ci provo, se siete d'accordo con le proposte e con questa visione, condividendo il post, magari qualche decisore politico potrà leggere. Fermo restando il discorso delle emissioni che resta importante ma con un impatto a lungo termine, il caldo urbano non dipende solo dall’atmosfera ma dipende anche da come abbiamo costruito le città. Troppo asfalto, troppo cemento, poca ombra, poco verde, superfici che accumulano calore di giorno e che lo restituiscono lentamente di sera e di notte come vi ho spiegato tante volte. Se vogliamo ridurre davvero il disagio durante le ondate di caldo, servono interventi concreti. Non miracoli, ma scelte pratiche. Ecco le mie proposte:
1. Mappare i quartieri più caldi con dati satellitari e sensori al suolo, così da intervenire prima dove il rischio è maggiore.
2. Depavimentare piazze, cortili scolastici, parcheggi e marciapiedi inutilmente cementificati, restituendo spazio a suolo naturale e permeabile.
3. Aumentare in modo serio gli alberi nelle strade, non piantarli a caso. Servono specie adatte, spazio per le radici, acqua e manutenzione.
4. Creare corridoi d’ombra lungo percorsi pedonali, fermate dei mezzi, scuole, ospedali, mercati e zone molto frequentate.
5. Rendere obbligatori materiali chiari e meno assorbenti per tetti, piazze, parcheggi, edifici pubblici e grandi superfici esposte al sole.
6. Trasformare i grandi parcheggi scoperti in aree ombreggiate, con alberi e superfici drenanti.
7. Proteggere gli edifici più sensibili. Scuole, ospedali, RSA devono avere schermature solari e raffrescamento efficiente anche attraverso un sistema di condizionamento efficace ma non invasivo.
8. Adattare gli orari dei lavori all’aperto durante le allerte caldo, soprattutto in edilizia, agricoltura, manutenzione stradale e logistica.
9. Recuperare acqua piovana e acque riutilizzabili per mantenere vivo il verde urbano anche nei periodi più secchi.
10. Introdurre per legge un test climatico per ogni nuova riqualificazione urbana. Quanta ombra crea? Quanto calore accumula? Quanta superficie impermeabile aggiunge? Quanto sarà vivibile a luglio ed agosto alle 15?
Servono città che nei mesi estivi non diventino piastre roventi ed ogni metro quadrato sottratto alle superfici che diventano bollenti può fare differenza dove il caldo pesa davvero.