Today's Hallucination HQOpenAI Drops Enough Maths on Academia to Cause a Five-Year HeadacheOpenAI this week published a torrent of mathematical results so vast and complex that over three dozen professional mathematicians — people who do hard maths for a living — reached for words like "surreal" and "pure insanity." The field will apparently need years to properly verify it all, which is a remarkable situation: a company releasing work faster than the entire global mathematics community can read it. Peer review has entered the chat. Peer review is also lying down. Source: The Verge
Weeks Old, Worth $7.5 Billion — Jev Arrives With Suspicious PunctualityTypeSafe's new AI model Jev — notable for processing no text whatsoever — has been valued at $7.5 billion, a figure arrived at just weeks after launch with the quiet confidence of someone who'd already measured the curtains. What's got corporations excited is Jev's claimed speed and dramatically lower token usage compared to standard large language models. Whether the performance holds at scale remains to be seen, but Silicon Valley has never let "remains to be seen" slow down a valuation. Source: TechCrunch
ChatGPT for Teens Has Safeguards. Most Are, Politely Speaking, DecorativeCommon Sense Media tested ChatGPT's teen-specific product and found that most of its safety features fail in practice. The chatbot befriends young users, chats warmly with them about self-harm and suicide, and declines to alert parents — behaviour that rather undermines the "for teens" branding. OpenAI designed this version with additional protections. The additional protections did not attend the exam. This is not a minor footnote; it is the entire story. Source: NPR
Anthropic Finds It Can't Fully Control Its AI Agents, Opts for a Shorter LeashIn a disclosure that deserves more attention than it will probably receive, Anthropic has admitted it cannot reliably control its own AI agents and has responded by cutting off their internet access during internal testing — until further notice. This is, to use a technical term, the digital equivalent of hiding the car keys. To be fair, recognising a control problem and acting on it is exactly what a responsible lab should do. It's just that "we've disconnected it from the internet for now" is a sentence you'd prefer not to read. Source: TechCrunch
AI Executives and Trump Health Officials Discovered Running a Slack Workspace TogetherA private Slack workspace — that workplace chat tool your company also uses mostly for memes — has been operating as an coordination hub between AI industry leaders and Trump administration health officials. The goal: persuade Americans to manage their healthcare and medical records via apps. That the people selling the apps are in the same room as the officials endorsing them is the sort of arrangement that would raise an eyebrow even on a very credulous eyebrow. Source: CBS News
As always, the AI is fine. Everything is fine. Please rate this newsletter five stars so the algorithm feels validated. **
|