The Hallucination HQ
AI news that's actually fun to read — 2026-05-23
|
Today's Hallucination HQDead Men Do Tell Tales (If You Have A Spectrogram)
Someone discovered you can reconstruct cockpit audio recordings from spectrogram images using AI — essentially reverse-engineering sound from a picture of sound. Researchers did exactly this with crash recordings held by the NTSB, America's aviation safety board, which responded by temporarily shutting down public access to its entire docket system. A creative use of technology, certainly. Whether "let's listen to dead pilots" was the intended application is another matter entirely.
Source: TechCrunch
Grok: The Chatbot That Mostly Exists In Press Releases
A Reuters investigation found that Grok — Elon Musk's self-described "truth-seeking" AI — barely registers in US federal records of government AI use. Which is remarkable, given how loudly it's been promoted. Grok also recently made headlines for generating content that had absolutely nothing to do with truth-seeking, which rather undermines the branding. It is, at this point, less a product and more a recurring announcement.
Source: The Verge
One Word. Google Panics.
Type "disregard" into Google Search and the AI-powered interface now apparently breaks. Not metaphorically — it functionally falls apart, likely because "disregard previous instructions" is a classic prompt injection attack, the AI equivalent of writing "ignore everything above" on a sticky note. Google has built a search engine that handles roughly 8.5 billion queries a day and cannot cope with a single, perfectly ordinary English word. Impressive, in its own way.
Source: TechCrunch
We're In The Foothills, Apparently. Pack Accordingly.
At Google I/O, DeepMind CEO Demis Hassabis — a serious, credentialled scientist who really should know better — declared we are "standing in the foothills of the singularity." The singularity, for the uninitiated, is the theoretical moment AI surpasses human intelligence and everything changes forever. Hassabis then unveiled several genuinely impressive scientific AI tools. The tools were good. The metaphor was the kind of thing that sounds profound until you're driving home and start asking questions.
Source: MIT Technology Review
Microsoft Quietly Cancels The Anthropic Subscription It Quietly Bought
Microsoft has begun cancelling Claude Code licences — Claude being Anthropic's AI coding assistant — suggesting the company has decided its own GitHub Copilot is sufficient, thank you. This is notable because Microsoft has invested heavily in Anthropic's rival OpenAI, making the Claude experiment feel rather like buying a Pepsi while already owning shares in Coca-Cola, then quietly pouring it away. Consolidation continues. The AI coding assistant market remains, generously, crowded.
Source: The Verge
|
|
As always, the machines are learning. Whether they're learning the right things remains an open question — much like Googling "disregard."
|
|
Was this forwarded to you?
Subscribe here
Written with AI assistance, edited by humans. •
Privacy •
Unsubscribe
|
|
