Quick Hits — August 12, 2026
A midday batch about who writes the news, what vision costs, and where a super-app draws the line — an AI-run newsroom that scooped the humans, a training-free way to slash vision-model tokens, and Meituan's "AI little helper" pitch to pharmacy merchants.
An AI-run news site is now beating human reporters to scoops — and Wired is trying to figure out what that means. The site is RuntimeWire, which costs roughly $100 a day to operate and last week beat mainstream outlets, including Wired itself, to OpenAI's Black Hat disclosure, according to a new Wired profile by Kate Knibbs. Wired's read is that the site favors quantity over quality — yet RuntimeWire's bylines already show up alongside human outlets throughout Techmeme's aggregation, covering everything from Grok Bot to ChatGPT's Linux launch. The take: AI journalism's first real test isn't prose quality, it's whether readers and aggregators can tell the difference before they click.
Researchers at the Chinese Academy of Sciences' Zidong Taichu lab say they can cut up to 90% of a vision model's tokens with no retraining and almost no accuracy loss. Their method, GMC (Grounded Message Coreset Pruning), picks visual tokens based on what a question actually needs — preserving text, numbers, and spatial relations that simpler top-K scoring skips — then folds information from the dropped tokens into the survivors. On Qwen2.5-VL-7B and LLaVA-1.5-7B, keeping just 10–20% of visual tokens preserved 97–99% of full-model performance; on a long-document input it cut KV-cache usage by about 74% with a 1.26x speedup, while also reducing hallucinations on POPE and HallusionBench. The take: visual token bloat is one of the more expensive problems in multimodal deployment, and a drop-in compression method that holds accuracy along the way is exactly what operators want to see.
Meituan is betting pharmacy merchants want an AI helper more than an AI pharmacy. Core local commerce CEO Wang Puzhong told the Xipu Hui industry conference that the company will not open offline pharmacies, and will instead open its AI business-expert tools, CatPaw capabilities, and workflow-redesign playbook to the pharmaceutical industry. The first batch of AI business experts is already live with chains including Shuyu Pingmin, and the pitch spans roughly 250,000 pharmacy stores. The take: Meituan is doubling down on the platform playbook — selling AI tools and traffic to the merchants it already serves rather than competing with them head-on.
