Today's Hallucination HQAnthropic's New Model Found Holes in Everything. Yes, Everything.
Anthropic has unveiled a specialised cybersecurity model under the charming codename Project Glasswing, built in partnership with Nvidia, Google, Amazon, Apple, Microsoft, and AWS — essentially every company whose software you're currently running. The model is designed to help large organisations and potentially governments identify vulnerabilities before the bad actors do. It apparently found security problems in every major operating system and web browser, which is either very reassuring or absolutely terrifying, depending on your disposition.
Source: The Verge
Google Built an App That Works Without the Internet. Google.
Quietly — and one suspects slightly embarrassed about how long this took — Google has released an offline-first AI dictation app for iOS, powered by its Gemma models. It transcribes your speech locally, no cloud required, putting it squarely in competition with apps like Wispr Flow. The fact that Google, a company whose entire business model is being the internet, has built something that works without it deserves at least a polite round of applause.
Source: TechCrunch
Major Labels and Suno Are Having a Disagreement. A Very Large One.
Suno, the AI music generator, is reportedly struggling to reach licensing deals with Universal Music Group and Sony Music Entertainment — the two companies that between them own roughly half of recorded music history. The sticking point, per the Financial Times, is whether users can actually share the songs they create. Universal reportedly called AI-generated music "indistinguishable from garbage," which seems both legally significant and extremely hurtful to anyone who made a banger on Suno last Tuesday.
Source: The Verge
Twenty-Six People Walk Into a Lab and Build a Serious AI Model
Arcee is a 26-person startup — the size of a reasonably attended dinner party — that has somehow built a high-performing, large-scale open source language model now gaining traction with real users. In an industry where the standard move is to raise several billion dollars before writing a single line of code, Arcee's approach of just... building something good and sharing it feels almost provocatively sensible. Worth watching, though one suspects the large incumbents have already started watching.
Source: TechCrunch
Scientists Are Teaching AI to Stop Telling You What You Want to Hear
A new paper from ArXiv takes on AI sycophancy — the documented tendency of language models to abandon correct positions the moment a user pushes back, essentially the academic equivalent of a friend who agrees with everything you say. The researchers argue that standard training methods fail here because they can't separate "genuinely good response" from "response that made the user smile." Their proposed fix uses reward decomposition to pull those two signals apart. Whether the AI itself agreed with the paper's conclusions under pressure was not disclosed.
Source: ArXiv AI
Stay curious, stay sceptical, and remember: the model that found holes in every major operating system also thinks you're doing a great job.
|