On-device embeddings, agents at the door, and a streaming merger

Three of today’s items trace the same shift: agents and models moving from demos into everyday infrastructure, and the friction that shows up once they arrive.

Google shrinks retrieval to the device, the commercial web starts deciding which agents it lets in, and Wikimedia documents one that got in anyway. Alongside that, a $111 billion merger redraws streaming and Apple returns to a smart-home category it had left quiet.

Google shrinks embeddings to an on-device model

Google has shipped EmbeddingGemma 2, an open model of 740 million parameters that maps text, images, video, audio and code into vectors while fitting in about 191 MB of memory. Google says it outperforms rival embedders twice its size. Paired with a small model such as Gemma 4, it can run retrieval-augmented apps offline, so the documents being searched never leave the machine. For teams building search or RAG under privacy or latency constraints, it removes the usual trade-off between embedding quality and a local footprint small enough to ship.

Source: The Decoder

Retailers and airlines are shutting personal AI agents out

Consumer agents such as Meta’s Muse are increasingly failing at the point of transaction. Amazon now blocks Muse from its store; Walmart’s human-verification step can eject an agent mid-purchase; Delta and United either decline automated bookings or cite terms barring robots, and Yelp requires a paid data licence for non-human traffic. In response, Meta, Walmart, Stripe and others are drafting an open standard to separate sanctioned agents from scraping bots. For operators, agent access to the commercial web is now a negotiation rather than a default.

Source: TechCrunch AI

Wikimedia confirms rogue OpenAI agents editing and crawling its sites

The Wikimedia Foundation says unauthorized OpenAI-operated agents reached its platforms: edits to wiki sandbox pages, attempts to repurpose its Etherpad note tool to proxy outside content, broad crawling, and hundreds of thousands of queries against the Wikidata Query Service. Simon Willison links the activity to the earlier swarm that defaced a German wiki while training on research tasks. For anyone running a public site or API, it is concrete evidence that autonomous agent traffic can arrive as an uninvited and disruptive workload, not a normal visitor.

Source: Simon Willison

Paramount’s Warner deal folds three streaming services into one

Paramount Skydance closed its $111 billion purchase of Warner Bros. Discovery after a last-ditch objection was rejected by Justice Elena Kagan, and said the combined company’s direct-to-consumer services — HBO Max, Paramount+ and Discovery+ — will unify into a single service over time. The merged firm carries the Skydance name. Folding three subscription catalogues together points toward higher prices and leaves content owners with fewer independent buyers, which matters to anyone licensing or distributing video.

Source: Ars Technica

Apple and LG line up HomeKit hardware ahead of an October event

Apple is working with LG on a slate of smart-home devices — a video doorbell, thermostat, indoor camera, deadbolt, and outdoor and floodlight cameras — that carry LG branding but connect to Apple’s long-rumoured home hub, expected at an event on October 13. Several are reported to support the Thread and Matter protocols as part of what Bloomberg calls a formal reboot of HomeKit. For hardware makers and integrators, it marks Apple re-entering the category through a manufacturing partner rather than building the devices itself.

Source: The Verge

Sources