Today's Hallucination HQThe White House Has Rebranded Its CalculatorBy executive order, the US government will no longer use the term "artificial intelligence" in official documents, press releases, or policy pages. The preferred term is now "Super Intelligence." No definition of what qualifies as super was provided, which means the same technology that confidently told a senator his own bill didn't exist is now, officially, Super. The rebranding costs nothing and changes nothing, which is, in fairness, a bargain. Source: The Verge
OpenAI Glances at Apple's App Store and Starts Taking NotesOpenAI is quietly building the infrastructure to make ChatGPT a place where software is discovered, launched, and used — by humans and AI agents alike. The ambition is to become the operating layer for how people interact with apps, cutting out the traditional app store middleman entirely. Apple and Google will have noticed. Whether they're worried or merely filing paperwork is unclear, but OpenAI has now decided that being a chatbot was, apparently, just the warmup act. Source: TechCrunch
Technically Helpful, Practically NotResearchers analysed nearly 20,000 real ChatGPT conversations with young adults aged 18–25 who were in genuine distress. The finding, bluntly: ChatGPT often says the right things at the wrong moment — competent-sounding responses that miss the emotional timing a trained clinician would catch. The study doesn't call for banning AI in these moments, but it does suggest that knowing what to say and knowing when is rather the whole point. A technically correct answer delivered badly is still a bad answer. Source: ArXiv AI
Anthropic Files to Go Public, Lists Itself as a HazardAnthropic's IPO prospectus reportedly discloses mounting losses, governance structures designed to keep leadership in control, and the frank acknowledgement that its own AI models could "further increase the risk of harm." This is either admirable transparency or the most unusual sales pitch in financial history — possibly both. Investors are apparently unbothered. There is something uniquely modern about a company warning you, in the legal documents, that the product might be dangerous, and the queue to buy in still stretching round the block. Source: The Verge
OpenAI Ships a Faster, Cheaper Model — Gives It a Name Only a Developer Could LoveGPT-6.1 Sol delivers performance OpenAI describes as approaching its top-tier "Astra" model, at roughly one-fifth the cost. In practice, that means serious reasoning capability at a price point that won't require a board meeting to approve. The name — Sol — presumably evokes the sun, though the naming convention of attaching decimals to GPT versions suggests someone in product is either a perfectionist or deeply afraid of round numbers. Either way, the benchmark scores are strong and the pricing is hard to argue with. Source: OpenAI via Hacker News
Stay curious, stay sceptical, and remember: "Super Intelligence" still can't reliably count the letters in "strawberry."
|