Today’s spread runs from the research lab to the city council: generative models that have to obey hard physical limits, a robotaxi that intervened on its own, and the electricity bill at the centre of the AI buildout.
Also in this roundup: a music model trained only on licensed catalogues, and fresh classroom evidence that simply banning AI backfires.
MIT’s HardFlow keeps generative models inside hard limits
MIT researchers have a fix for generative models that stray outside hard limits: enforce the constraints on the finished answer rather than at every intermediate step. Their HardFlow algorithm treats sampling as a trajectory-optimisation problem, letting the model explore freely while still landing on a feasible output. Across robot manipulation, maze navigation and image-editing tests it satisfied every constraint and beat existing methods on solution quality at comparable compute. Because it is plug-and-play at deployment, pretrained models need no retraining.
Source: MIT News AI
Waymo’s robotaxi called the cops on armed passengers
A Waymo robotaxi in San Francisco pulled itself over and summoned police after detecting a firearm in the cabin. Two juveniles were arrested and taken to juvenile hall, and a loaded AR-style ghost gun was recovered. The police report did not identify who was driving, but Waymo confirmed to the Los Angeles Times that the car was its own, describing the episode as a terms-of-service violation involving a weapon. Earlier this year one of the company’s vehicles faked a mechanical fault in order to stop.
Source: The Verge
Data-center builders switch to fighting utilities on consumers’ behalf
Big tech’s data-center developers are changing tactics in the battle for local approval. Amazon, Microsoft, Oracle and peers are offering municipalities and regulators sweeter financial packages, while siding with ratepayers against utilities that want the public to shoulder part of the cost of new electricity loads. The companies still expect years of local opposition, since activists are unmoved by the money and increasingly organised, but they are now putting concrete benefits on the table instead of just sharpening their messaging.
Source: The Information
ElevenLabs ships a sharper music model with licensed-only training
ElevenLabs has pushed out Music v2.5 for its AI music generator, available through both its app and its API, with free and paid tiers. Listeners picked the new version over its predecessor in a blind test of roughly 48,000 comparison pairs. ElevenLabs says the model learned only from licensed music, a claim that matters to customers weighing rights exposure.
Source: The Decoder
Two-year study finds banning AI hurt the students it protected
A law professor ran a two-year classroom experiment comparing three regimes: a blanket AI ban, unguided use, and structured training. The cohort working without AI placed last in both years. The researcher, who had expected unfettered access to backfire, now writes that the initial assumption was wrong. The result undercuts the simplest version of the case for prohibition in teaching.
Source: The Decoder
Sources
About the Author
