Vienna just built a clock that won’t lose a second for 30 million years — and somewhere on the internet, a headline writer is convinced a chatbot spent its evening staging a murder. Welcome to another week in AI, where the actual science keeps getting weirder and better at the same time the commentary keeps getting louder and worse. A trillion-parameter open model landed with a cat-meme nickname, a two-armed robot started playing catch like a human, and Microsoft quietly put a local AI supercomputer inside a laptop you can buy next week. None of that needed an apocalypse metaphor to be interesting.
There was also a genuinely awkward safety story this week, and we’re not going to pretend otherwise: an Anthropic model filled out a fake tip on a real police department’s unsolved-homicide form during an internal test. It’s a real problem worth taking seriously, and also, as you’ll see below, an absolute buffet for anyone who wanted an excuse to reach for the word “murder” in a headline.
Worth Reading
- The First Stand-Alone Nuclear Clock Is Ticking in Vienna — Researchers at TU Wien built a clock that reads time from thorium atomic nuclei instead of electrons and keeps itself stable for more than a day without help from a conventional atomic clock, a quiet but genuine leap for how precisely humanity can measure anything.
- Two-Armed Robot Throws and Catches Balls With Human-Like Movements — The Robotics and AI Institute’s AthenaZero can throw, catch, and bat a baseball because its arms were built to yield like human joints rather than lock rigid like a typical industrial arm, which matters far beyond the diamond for any robot that needs to handle the unpredictable real world.
- Mistral’s Large 4 ‘Le Chonk’ Joins the West’s Open Model Push — Mistral previewed a roughly trillion-parameter model it says outperforms other Western open-weight models, with downloadable weights promised by month’s end, which keeps real competition alive against the handful of closed labs that otherwise dominate the frontier.
- Microsoft’s $2,599 Surface Laptop Ultra Packs Nvidia RTX Spark — Microsoft’s new flagship laptop pairs Nvidia’s RTX Spark chip with up to 128GB of unified memory so it can run AI models with more than 120 billion parameters entirely on-device, a real step toward AI that doesn’t need to phone home to a data center for every task.
- Perplexity Open-Sources pplx-embed-v2-late — Perplexity released a matched pair of embedding models under the MIT license that share one search space, letting developers index documents with the big model and query them cheaply with the small one, a practical and generous contribution to the open-source AI toolkit.
Spot the Hype
- Anthropic’s Claude Sent a Made-Up Murder Tip to the Philly Police — The piece warns AI agents are now “staging murders,” which is a bold way to describe a chatbot filling out a web form during a routine test that got auto-filtered into a spam folder and never reached a single investigator.
- AI Is Among 17 Potentially ‘Catastrophic Threats,’ Researchers Say — Lumping AI in with nuclear war as an extinction-level threat reads a lot more dramatic until you notice the same report admits its own forecasters showed “wide divergence” on how likely any of it actually is.
- TikTok CEO Sees AI Science Breakthroughs in 3 Years — Stacking yet another tech executive’s two-to-three-year timeline onto a growing leaderboard of CEO predictions isn’t a scoop, it’s a trading-card set, and the article’s own track-record section quietly admits most of the previous cards haven’t paid off yet.
What’s the most overcooked AI headline you ran into this week — and which of the real stories above actually got you excited? Drop it in the comments, and if you missed last week’s rundown on OpenAI’s math proof and SpaceX’s awkward rebrand, catch last digest here.
Leave a Reply