Nous Research released Arete-27B, a post-trained version of Alibaba's Qwen3.8-27B. The stated aim was to reinforce Western values on a Chinese-trained model. The method was narrow. The team ran two ep
An agent is mid-edit on design A when a message arrives on another channel: switch to design B. The router can wake the right worker. A shared folder can show which brief is live. The harder part is s
Around 30 October 2026, developers reading DeepSeek’s website JavaScript found application routes and key-management hooks for a Cyber model that had not yet appeared in any public menu. The reporting
Native multimodal does not mean dropping separate vision encoders; DeepSeek-V4.1-Flash still uses ViT + projector, and the real goal is an updatable world state.
In October 2026, a group of computer science and AI faculty published a Joint Statement on Slow Research, addressed to undergraduates who want to do research. The statement page lists 120 signatories,
Epistemic status: An argument about evidence standards, built on published psychology and interpretability papers. I take no position on whether any current model is conscious; scope and limits are in
Epistemic status: A close reading of the misbehavior and access-control sections of DeepSeek's DSec report, plus a proposed direction drawn from my own runtime work. The incidents are as DeepSeek repo
Epistemic status: A design note. The motivating failure is from a long-form narrative agent I ran. One ICLR 2026 workshop paper supplies the measurement. Scope is at the end. Related: Chen, Epistemic
Epistemic status: Same-day reading of OpenAI’s GPT-6 Sol / Luna launch and Anthropic’s Claude Opus 5.5 launch, plus one Cursor A/B paper-review sample of my own. All prices and benchmark numbers are f
