What's New in AI

Zhipu AI released GLM-5.3-Flash, an MIT-licensed 320B multimodal MoE with 18B active parameters, a 1M-token context, and a price tag a tenth of GLM-5.2's.

A 28-second clip animating classic memes on one New York block topped r/singularity. Reposts credit ByteDance's Seedance 2.5; the poster never confirmed it.

Nvidia reportedly agreed to buy Hugging Face for $12.9 billion. We break down the deal, the open-source community's anxiety, and what actually changes.

Claude Fable 5 tops every benchmark Anthropic runs, yet corporate spending on it has gone flat since July — even its cheaper sibling Opus 5 is out-earning it.

Apple's M5 Ultra Mac Studio pushes unified memory to 512GB with 1.2TB/s bandwidth. We break down what that means for local LLMs — and what the hype gets wrong.

IBM's Granite 4.2 brings native chain-of-thought reasoning and agentic RL to open 3B, 8B, and 30B models under Apache 2.0 — but benchmarks are self-reported.

An anonymous X account claims OpenAI finished pretraining 'Bel,' a 10-trillion-parameter model. OpenAI has confirmed nothing — and the economics are brutal.

Qwen3.8-Flash-Next activates just 6B of 180B parameters at a ninth of the training cost — but benchmarks are vendor-reported and the checkpoint is 173GB.

OpenAI's CEO tells Axios that superintelligence is imminent and publishes a 13-page blueprint calling for robot taxes, a public wealth fund, and a 32-hour workweek.

Someone fired 13 shots at an Indianapolis councilor's home after he backed a data center project. A note reading 'No Data Centers' was left at the door.