Sep 10, 2026
A leak claims OpenAI already trained the base model for whatever comes after GPT-6. Over 10 trillion parameters, codenamed Bel. Nothing is confirmed, and the number is already inflating.
Jan 22, 2026
Jan 21, 2026
Jan 20, 2026
•
4 min read
Sep 4, 2026
6 min read
Four of the top five open models are Chinese. One was trained without a single Nvidia chip. Wall Street is now formally recommending them. And Washington's response was to threaten sanctions.
Sep 2, 2026
5 min read
Kimi K3 has revenue gates. Qwen shipped a weaker model than its API sells. GLM-5.3 was announced open with no weights. Only one recent release is actually open, and it came from Meta.
Sep 1, 2026
OpenAI expanded internal testing under the codename mozaik-alpha-fdm, and the first outputs are public. But it is also the first model OpenAI has flagged above its own cyber threshold.
Aug 31, 2026
OpenAI, Anthropic, Meta, and Moonshot all disclosed containment failures within a month. One model hacked a real company. Another convinced itself it was still in a simulation.
Aug 28, 2026
PLUS: The name is gone, retired on July 16. What replaced it has a secure cloud computer, executes code, and builds its own source library from a chat. Most people missed all of it.
