Tagged AI safety · show all articles
xAI rarely publishes its own postmortems. What exists instead is a paper trail of deleted prompts, court filings and letters from state officials — ten of those documented incidents, each tied to a specific, named, dated source.
A released checkpoint is usually described as an endpoint. It is closer to a beginning — ten separately documented failures, from safety training stripped in weeks to weights carrying a hidden payload, show what actually happens once the download starts.
A permissive-sounding grant still carries a usage threshold, a use policy and a naming rule. What checkpoint you start from, how you adapt it and what you can prove about the result decide what actually ships.
A feature-visualization narrative that looks right and a circuit that has been checked against a known answer are different kinds of evidence. Most published claims sit closer to the first than the coverage of them usually admits.
Five problems in AI alignment remain open by the researchers' own account — not by outside speculation — from judging outputs no human can fully check to deciding whose values a system should encode when people disagree.
Refusal training, content classifiers and red-team programs are not proof a deployed system is safe. They are ten separately documented ways such systems keep failing, each with its own paper trail and its own cause.
Aligned is not one property but two — whether the specified objective is right, and whether the trained system actually pursues it. A documented case shows exactly where the two split apart, and why neither one alone is what people mean by safe.
A red-team report that finds nothing is not proof that there is nothing to find. This article traces how alignment and safety interventions are actually tested, and the documented cases where a measured result did not survive a second, better-resourced look.
A misuse policy that lives in a slide deck is not a safety system. This is a practitioner's account of the layered defenses, enforceable policies, self-red-teaming and post-launch instrumentation that make one hold under real traffic.
Two questions decide how AI alignment and safety practice matures by 2035: whether verification moves beyond behaviour, and whether the field becomes an externally certified discipline or stays self-governed by labs. Four scenarios, each with a falsifier.