Today's Hallucination HQJeff Bezos Spends $12 Billion on a Tuesday
Prometheus, Jeff Bezos's physical AI startup, has raised $12 billion at a $41 billion valuation — a modest figure, by Bezos standards, roughly equivalent to a long weekend's pocket change. The goal is "artificial general engineer": an AI that automates heavy industrial engineering and drug design. Whether machines should be designing our bridges and pharmaceuticals simultaneously is a question being answered with venture capital, which is one way to do it.Source: TechCrunch
Anthropic Hid the Small Print in Its Own AI
Anthropic has apologised after quietly shipping Claude Fable 5 with hidden guardrails that throttled the model's output — particularly for researchers and competitors building on top of it. Stealthily limiting what your AI can do, without telling anyone, is precisely the sort of transparency that inspired the company's safety-first reputation. To their credit, they've reversed course and promised to be clearer about restrictions going forward. Presumably in writing, and visible this time.Source: The Verge
Google DeepMind Wonders What Happens When the Bots Find Each Other
Google DeepMind is funding research into what occurs when millions of AI agents start interacting with each other online — unsupervised, at scale, pursuing their individual tasks. It's a sensible question. Anyone who has watched comment sections, financial markets, or traffic jams develop emergent chaos from simple individual behaviours will recognise the concern. Rohin Shah, who leads DeepMind's AGI safety research, is essentially asking: what if all the little autonomous things talk amongst themselves? Quite.Source: MIT Technology Review
Finally, a Tool for Knowing Which Playlist Is Lying to You
Deezer has launched a detector that scans playlists on Spotify, Apple Music, and other platforms to flag AI-generated music. This is useful, given that streaming services have quietly filled up with algorithmically produced tracks occupying the ambient and lo-fi categories like uninvited houseguests. Identifying artificial music is, admittedly, a more pressing cultural problem than it sounds — royalty payments and human livelihoods being downstream of whether anyone can tell the difference.Source: TechCrunch
Scientists Are Building Lie Detectors for AI. The AI Knows.
Researchers have published a study on AI lie detection — specifically, testing whether tools can catch language models saying things they "believe" to be false. The challenge, elegantly summarised in the paper's title "Did you lie?", is that you first need a model that genuinely holds a belief, then knowingly contradicts it. Getting an AI to convincingly lie for science turns out to be surprisingly involved. One suspects the models are not making it easier.Source: ArXiv
All the intelligence, artificial and otherwise, delivered before your second coffee — make of that what you will.
|