More posts
ChatGPT merged Chat and Codex into one entry point. Claude just did the same with Chat and Cowork. So you don't have to think about whether you want a chat or a task anymore. Describe what you need in the chat and the agent works it out on its own, whether that means running something long, calling tools, or handing you the finished thing. Claude Docs is out too, for writing documents together. Multiple people can work on them, and you can export to Word, PDF, or Google Docs. Then there's Claude Slides, which generates decks you can edit and present, and it exports to PPT.
. From 8,000 characters a week down to 2,000. Honestly still enough for what I use it for. The voice input is actually pretty simple to build, but I haven't seen anyone else do it. Paid conversion's probably just too low?
. 1.6T params, ~48B active, native multimodal, aimed at long-horizon agents, trained start to finish on Chinese hardware. Images get quantized into multi-level discrete tokens, and that's where Visual Coding and 3D pick up. On the RL side, failed trajectories get turned back into verifiable training tasks. W4A4 inference, 43% faster prefill. Meituan's own evals put it between Claude Opus 4.8 and Opus 5.
It stacks layers instead of parameters. The decoder has 32 layers, but layers 4 through 27 run twice with shared weights, so a forward pass does 56 layers of work while the checkpoint stays at 9B. Their own eval numbers: 71.38 on BIRD-Dev and 13.72 on AutomationBench, both above Qwen3.5-27B. Terminal Bench is still behind.
