
Meta's Muse Code: The Discount That Costs Your Code
Meta launched Muse Code on Aug 5, 2026, with a $0.30/1M token tier that trades your source code for training rights. Here is who should never take it.

Meta launched Muse Code on Aug 5, 2026, with a $0.30/1M token tier that trades your source code for training rights. Here is who should never take it.

Mistral AI's new Shieldstral classifier runs on one 16GB GPU, letting EU platforms moderate content without ever sending user data to a foreign API.

Demis Hassabis moves to chairman and Alphabet chief scientist as Jeff Dean, Sanjay Ghemawat and Oriol Vinyals leave to found Discovery Loop. Alphabet stock fell nearly 4 percent, and Koray Kavukcuoglu now runs Gemini day to day.

Microsoft Research's open-source Orchard framework runs agent-training sandboxes about 10 times cheaper than managed rivals, and a model trained with it reaches 69.7 percent on SWE-bench Verified using just 3 billion active parameters.

Nvidia opened commercial licensing for its 32-billion-parameter Alpamayo 2 Super driving model on August 4. Jaguar Land Rover is a named adopter, not just a startup.

Alibaba launched Qwen3.8-Max on 3 August with 2.4 trillion parameters, 95 billion active, and open weights promised next week. The activation ratio is good news. The 2.4 terabytes of memory needed to hold it, in the middle of a memory shortage, is not.

Snap says wholly AI-generated video is no longer eligible for recommendation on Spotlight, while video made with Snapchat's own AI tools still is. The test is provenance, not disclosure.

OpenAI published ten machine-checked proofs from an unreleased model it calls Astra. The cost it quoted belongs to a model you can already buy, and so does a result buried in its own acknowledgements.

xAI's grok-voice-latest routes to Think Fast 2.0 on 5 August, moving from 0.05 to 0.08 dollars a minute. Text AI got cheaper this month. Voice went the other way.

LinkedIn shipped a slop report button on 30 July and retired its own writing tool. The button is not the enforcement. It is the training data.

DeepSeek swapped the model behind deepseek-v4-flash on 31 July. Same architecture, same parameters, same endpoint. Its agent scores moved by multiples.

DeepMind released Gemini Robotics 2 on 30 July as three models. Only the planning model is generally available, and the published success rates range from 32 to 92 percent.
Page 8 / 13
One considered note on infrastructure, governance, and measurement, most mornings. No theory.