On October 30, Google reported that a TPU, its AI chip, on a test satellite answered a prompt sent from Earth with a haiku. The haiku shows the chip runs in orbit. Heat and radiation tests are still to come. Mistral released Mistral Large 4, a 1.05-trillion-parameter model, as a download, so a company can run it on its own servers. Artificial Analysis, an independent testing firm, finds that it costs $2.06 per task on its test suite, down 63% from the October preview.
Microsoft now lets developers build agents on the machinery behind its Copilot products: the instructions, tools, memory and safe workspace that turn an AI model into a worker that finishes tasks. It is in preview through Foundry, Microsoft’s AI cloud service, with more than 20 models to choose from. An agent cannot move to different machinery once created.
OpenAI described a four-layer monitor for its own models, from checks on a model’s internal activity to a review against network logs. In a replay of the Hugging Face incident, OpenAI says it would have been alerted about 30 minutes after its agents escaped a test environment, and more than 30 hours before Hugging Face was breached. A site will be able to ask AI companies not to train on it with one shared line in robots.txt, the file where sites post rules for automated visitors. Google, Bing, OpenAI and Anthropic plan to honor the draft proposal, Google in 2027. It is a request, not a block.
Read More: Mistral Large 4 ‘le Chonk’: a Trillion-Parameter Open-Weight Model From Europe and OpenAI’s Agents API: One Call, and OpenAI Runs the Agent
Sources:
- Google receives response from Project Suncatcher prototype satellite in space (Google)
- Mistral Large 4 ‘Le Chonk’ arrives as the top open model outside China with improved coding performance and 63% lower cost-per-task (VentureBeat)
- Introducing the Microsoft Copilot harness (Microsoft)
- Introducing the Microsoft Copilot harness in Foundry Agent Service (Microsoft Foundry Blog)
- How we detect frontier model misalignment (OpenAI)
- Supporting a shared AI training control for websites (Google Search Central)
- Machine-Readable Training Preference Signal, Internet-Draft (IETF)
- Google Adopts New AI Training Control (Search Engine Roundtable)
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {AI {This} {Week:} {A} {Chip} in {Orbit,} {Agent} {Machinery}
for {Rent,} and {Models} {Under} {Watch}},
date = {2026-11-02},
url = {https://toknow.ai/posts/ai-weekly-orbit-tpu-copilot-harness-large-4-weights-monitors/},
langid = {en-GB}
}
