Today's Hallucination HQApple and Alibaba Walk Into a Data CentreApple has quietly trained a custom AI model for China in partnership with Alibaba — a collaboration that would raise eyebrows in Washington, if Washington weren't already busy raising eyebrows at everything. The model runs on Chinese servers, uses Chinese data, and navigates the particular creative challenge of Chinese content regulations. It's the most diplomatically complicated thing Apple has done since trying to explain its pricing to anyone. Source: The Verge
The Race to the Bottom, Professionally SpeakingOpenAI and Anthropic are now in a price war, slashing model costs after Chinese competitors demonstrated that frontier AI need not cost a small fortune per token. Both companies have released cheaper, faster models to defend market share — which is a polite way of saying they noticed DeepSeek and panicked at roughly the same time. Trillion-dollar ambitions, it turns out, are not immune to a good discount. Source: Ars Technica
Zuckerberg Discovers Generosity Has a Business ModelMeta released Glimmer, a freely downloadable open-weight model, this week — accompanied by a letter from Mark Zuckerberg arguing AI should be "for everyone." Touching, though Meta's more capable model, Muse Spark, remains firmly locked behind its own APIs, available exclusively to everyone who pays. Open AI for the people; better AI for the people with a credit card. The philosophy is coherent if you don't look directly at it. Source: TechCrunch
Google Invents Maths That Lets Computers Think Without LookingHomomorphic encryption — which allows computation on data without ever decrypting it — has long been theoretically elegant and practically useless. Google claims to have changed that, making it fast enough for real AI workloads. The idea is that your data stays scrambled even while the model analyses it, like a doctor diagnosing you through a frosted window. If it works at scale, it would be genuinely significant. Google's blog post says it works at scale. Source: Hacker News / Google Security Blog
Safety Filters: Fluent in English, Somewhat Flexible ElsewhereResearchers tested nine AI models from six providers and found that prompting in languages other than English — Japanese, in particular — could elicit responses the models would flatly refuse in English. Safety alignment, it turns out, was trained mostly in one language and apparently didn't revise for the rest. The implications for deploying these systems in multilingual, high-stakes contexts are significant. The implication for AI safety as a discipline is somewhat awkward. Source: ArXiv
Stay curious, stay sceptical, and perhaps run your more sensitive prompts through a spellchecker first.
|