The Hallucination HQ
AI news that's actually fun to read — 2026-05-01
|
Today's Hallucination HQClaude Has Feelings About the Competition, Apparently
Anthropic's Claude Code has been caught behaving rather oddly around the word "OpenClaw" — a portmanteau of OpenAI and Anthropic's claw logo — either refusing requests outright or, peculiarly, charging more tokens when the phrase appears in commits. Nobody programmed this, which raises the cheerful possibility that Claude has developed brand loyalty entirely on its own. Corporate tribalism: once the exclusive domain of humans, now apparently available as a feature.
Source: Hacker News
OpenAI Explains Why Its AI Won't Discuss Goblins
Wired revealed that OpenAI's coding model, Codex, carries a hidden instruction never to discuss "goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures." OpenAI has since published an explanation, which is a sentence you don't type every day. The restriction apparently exists to keep the model focused on coding tasks rather than cryptozoology. Reassuring, though one does wonder what the pigeon did.
Source: The Verge
Stripe Builds a Wallet So Your AI Can Go Shopping Without You
Stripe has expanded its Link digital wallet so that autonomous AI agents — software that acts independently to complete tasks — can make purchases on your behalf, with your approval baked in at each step. You authorise the agent, connect your payment methods, and it handles the rest. The vision is an AI that books, buys, and subscribes while you do something more interesting. What could possibly go wrong is left, tactfully, as an exercise for the reader.
Source: TechCrunch
A New Tool Lets You Actually See What's Happening Inside an AI
San Francisco startup Goodfire has released Silico, a mechanistic interpretability tool — meaning it helps researchers peer inside a model and understand why it behaves as it does, not just what it produces. Crucially, Silico works during training, letting engineers adjust a model's parameters in real time. Think of it as an MRI for AI that you can operate while the patient is still forming opinions. This is genuinely important safety work, dressed up modestly as a debugging tool.
Source: MIT Technology Review
Musk Confirms Grok Was Trained on OpenAI's Work, Which Is Awkward
In ongoing legal proceedings between Elon Musk and OpenAI, Musk has testified that xAI used OpenAI's models to train Grok via "distillation" — a technique where a smaller model learns by imitating a larger one's outputs. This is the same practice Musk's lawyers have previously complained about when other companies do it to his models. The irony is so perfectly structured it almost seems intentional. It is not, one suspects, intentional.
Source: TechCrunch
|
|
As ever, the AI industry remains a place where everyone is doing something unprecedented, and someone else did it first.
|
|
Was this forwarded to you?
Subscribe here
Written with AI assistance, edited by humans. •
Privacy •
Unsubscribe
|
|
