The Hallucination HQ
AI news that's actually fun to read — 2026-06-03
|
Today's Hallucination HQUber Tells Staff to Stop Drinking From the AI Fire Hose
Uber spent its entire annual AI tools budget in four months after actively encouraging employees to use AI "as much as possible" — a policy that went exactly as you'd expect. The company has now introduced spending caps, presumably while someone in finance lies down in a dark room. To be fair, telling a workforce to use unlimited AI freely and then being surprised by the bill is a bold management strategy.
Source: TechCrunch
Microsoft Build 2026: Satya Nadella Announces Everything, Simultaneously
Microsoft's annual developer conference delivered new Surface hardware, an always-on personal AI assistant, and updates to its in-house models — essentially the full bingo card. The always-on assistant is the headline act: a persistent AI that watches your screen and remembers things so you don't have to, which is either the most helpful tool imaginable or a concept that will keep privacy researchers employed indefinitely. Either way, Nadella looked very pleased.
Source: The Verge
White House Drafts AI Rules, Industry Objects, White House Nods and Rewrites Them
President Trump has signed a revised AI executive order that swapped mandatory pre-release government reviews of powerful AI models for voluntary ones — after the industry raised objections to the original version. Voluntary oversight is, of course, the regulatory equivalent of putting up a "please behave" sign. The order still technically exists, which counts as something.
Source: TechCrunch
UK Tells Google: Publishers Can Leave the Party If They Want
Britain's Competition and Markets Authority has ruled that website publishers must be allowed to opt their content out of Google's AI Search features — like AI Overviews — without losing their regular search rankings as punishment. This matters because previously, opting out of AI features risked visibility in standard results, which is a bit like a restaurant telling you that if you don't want your recipe used, you're off the menu entirely. Google has until later this year to comply.
Source: The Verge
AI Models That Reason Too Hard Are Now Also a Problem, Researchers Confirm
A new paper from arXiv examines "harmful overthinking" in large reasoning models — AI systems that generate long chains of thought before answering. The finding: more reasoning isn't always better, and can actually lead to worse or more problematic outputs when a model keeps second-guessing itself into a corner. Researchers are calling this "thinking past the answer," which is something most of us recognise from exam season, only with higher stakes and no option to leave early.
Source: ArXiv AI
|
|
As always, the AI told us this edition was perfect. We published it anyway.
|
|
Was this forwarded to you?
Subscribe here
Written with AI assistance, edited by humans. •
Privacy •
Unsubscribe
|
|
