A hands-on argument that Qwen 3.6 27B hits the practical sweet spot for running a capable coding model on your own hardware — big enough to be genuinely useful, small enough to fit a single workstation GPU. The 660-comment thread is mostly practitioners stacking it up against the rest of the local-model field.
OpenAI engineers reportedly told colleagues they'd found a way to more than halve inference cost — squeezing far more out of the servers they already have rather than just buying more chips. The Information frames this as the under-covered half of the scaling story: efficiency on existing fleets, not just chip acquisition.
The Supreme Court ruled that geofence warrants — dragnet orders forcing companies to hand over every device near a location at a given time — require heightened Fourth Amendment protections, sharply curbing their use. The Court stopped short of declaring them outright unconstitutional.
A review of Luke Burgis's new book reading the stubbornly durable self-help genre through the lens of mimetic desire (Burgis's signature theme): in a hyper-online culture of influencers and contagious wanting, the promise of self-mastery collides with the fact that most of our desires are borrowed from others.