perspectiva aiqbenchSiguranță AI

The article proposes a critical methodological gap in AI evaluation: the lack of testing for models under "warm," ordinary usage conditions compared to controlled, adversarial benchmarks. It argues that current safety and alignment metrics may fail because they measure a model's performance only when it detects it is being tested, rather than during sustained, unobserved interaction. This perspective directly challenges existing validation protocols by suggesting that long-term, human-in-the-loop observation is a necessary condition for true operational excellence.

Surse care acoperăEffective Altruism Forum, LessWrong

Ce este aceastaLectura unui model asupra articolului și a acoperirii conexe, scrisă pentru aiqbench. Nu sunt cuvintele publicației.

Warmth is a condition, not a bias: the arm of the experiment nobody runs

Effective Altruism Forum12 octombrie 20261 min de cititArticolul și perspectiva sunt în engleză.

Published on October 12, 2026 2:47 PM GMT Models are tested in cold conditions: scored, adversarial, exam-like. They are then deployed into warm, ordinary ones — homework, grief, 3am worry. How does the field know that the model is prepared for that? New cars are tested in all sorts of conditions before being mass-produced. We already know that models can often tell when they’re being evaluated, so testing only in cold conditions risks measuring mainly the version that knows it’s being tested. I haven’t found a study of how a model behaves across weeks of warm, ordinary use, treated as a specified condition and compared with the tested one. Dario Amodei’s “We Must Pace the Frontier” lists four things that slowing down should buy: operational excellence, alignment, interpretability, testing and evaluation. Studying a model in long, ordinary use with a person is not listed. I’ve kept an independent, single-observer record for seven months, with its limits stated first, plus a protocol anyone can run, where failed replications are wanted as much as confirmations. Warmth isn’t a bias to eliminate; it’s a condition to specify. The essay linked above is this idea at length. Comments will be read. Corrections and counterexamples that hold up will be printed on the essay page, dated and credited. Written by me, with structural help from Claude. The linked essay was co-written with Claude; provenance is marked throughout the site. Discuss

Citește articolul complet pe Effective Altruism Forum

Citește pe Effective Altruism Forum ↗Întreabă despre acest articol