Pre-tokenisation boundaries leave a measurable “compression bill” for tokenizers
A study quantifies how adding boundary rules changes the minimum achievable token count – showing a 28.3% to 36.8% boundary overhead on English Wikipedia. It also separates boundary-induced compression costs from prediction quality by using certificate-based accounting, concluding that compression and prediction can favor different token dictionaries.
OpenAI and Meta are betting on AI agents in dedicated hardware
The Verge reports that Meta and OpenAI are pushing beyond phones and PCs with “cutesy” agent experiences designed for physical devices, taking aim at a space where earlier attempts often met backlash. The coverage frames this as a hardware appetite test, with Jony Ive-linked OpenAI efforts and Meta’s Muse/companion push running on a similar bet that user behavior will follow agent interfaces.
Neurosymbolic routing for reliable reasoning on edge devices
Researchers propose a router that learns to dispatch each query to the cheapest correct solver instead of forcing a small probabilistic model to approximate deterministic reasoning. On Raspberry Pi 4 hardware, the learned routing hits 100% routing accuracy and up to 98.3% overall accuracy (with fast millisecond responses for formatted queries), turning “small models are unreliable” into a routing problem.
Trump orders the US government to refer to AI as “Super Intelligence”
The Verge says a new executive order changes official language from “artificial intelligence” to “Super Intelligence,” with Trump pitching it as a simpler, more fitting label for what he argues is already beyond “artificial.” The move arrives alongside claims that safety oversight is being handled through self-policing by major tech players.
Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning
Hugging Face introduces an open leaderboard designed to evaluate multilingual TTS and voice cloning at scale, targeting the hard part of generative audio: consistent, comparable benchmarking. The work is meant to make it easier to track progress across systems while keeping evaluation aligned with real-world voice and language variability.
Google reportedly tests paying publishers for AI search overviews
Google is piloting a program that pays roughly 100 publishers based on how much their content contributes to AI Overviews, AI Mode, and related Gemini experiences, according to The Verge citing reporting from The Information and Digiday. The test signals a push to address scrutiny over how AI features reshape web traffic and citation incentives.
Risk-averse online POMDP planning via CVaR of immediate cost
A new planning method targets a gap in earlier CVaR approaches by applying risk sensitivity directly to per-step immediate cost over the belief state, rather than only at trajectory-level value. The paper keeps standard MDP structure so existing expectation-based planners can be adapted – while also providing finite-time guarantees that bound surrogate-belief approximation error back to the original POMDP value.