Prompt Markers, Reasoning Ceilings, and One-GPU Agents

Prompt Markers, Reasoning Ceilings, and One-GPU Agents

A researcher decoded hidden Claude Code prompt markers and Anthropic pulled them. Developers mapped GPT-5.5 reasoning tokens clustering at 516. DeepReinforce shipped MIT-licensed Ornith-1.0, targeting one 80GB GPU.

Custom Silicon, Ford's Reversal, and Computer Use

OpenAI unveiled its first inference chip with Broadcom, Ford rehired 350 engineers after an AI push fell short, and Google built computer use into Gemini Flash.

Open Weights, Code Review, and Agent Identity

US curbs pushed buyers to open-weight models, two reports argued AI coding's hard part is no longer writing it, and agents got identities of their own.

Model Halts, Agentic Git, and Open Reasoning

A US order halted Anthropic's new models, AI agents rewrote Git in Rust, and Hugging Face open-sourced a DeepSeek-R1 reasoning recipe.

Coding Costs, Apple on Gemini, and Open-Weight Gains

AI coding tools may cost providers 10x what users pay, Apple rebuilt Siri on Google's Gemini, and DeepSeek's V4 reportedly beat GPT-5.5 on precision.