2 – 8 November 2026
Anthropic talks about AI that improves itself, a seal to prove which model answered, DeepSeek builds plugins on request
Anthropic writes in its IPO filing that it may soon reach a threshold of AI-accelerated research. It also releases an open tool to prove which model produced an answer, and DeepSeek shows how to build your own plugins by chatting.
Anthropic says it is close to a threshold of “AI that helps build AI”
On November 2, Anthropic filed an amended version of its IPO paperwork (S-1/A). According to TechCrunch, the risk section says the company may soon reach the threshold its own safety policy calls “dramatically accelerated progress attributable to AI”. Anthropic says no current model has crossed it, that it is still assessing the data, and that it does not plan to let Claude design the next generations alone without human oversight. It is considering changes to its policy. On its measurements page, which dates from August, Anthropic says Claude “leads” 26% of its internal AI research work, up from under 1% in February, without being autonomous on any part of it. An intern who starts marking the papers: the teacher keeps the red pen.
What it changes: Nothing changes in how you use Claude today. These are Anthropic's statements in a financial document, and TechCrunch notes that some readers will also see them as a communication argument. We could not open the SEC filing itself: the exact wording remains to be checked.
Claude version history TechCrunch — Anthropic warns that it's very close to RSI capability, aka self-building AI (November 2) Anthropic — Measurements for understanding the pace of AI development inside frontier labs
FlashProve: a mathematical seal to prove which model answered
On November 2, Anthropic released FlashProve, an open prototype (Apache 2.0 license) that lets you check that a reply really comes from a specific model, using a cryptographic proof that reveals nothing about the model itself. The work was done with researchers from EPFL, MIT and other teams. It is still very heavy: tested on Llama 4 Maverick, it needs about 72 times more graphics-card computing time, each proof weighs about 18.8 MB for a 1,024-token window, and checking takes around twenty seconds on a processor. A wax seal on every reply, except the seal weighs more than the letter.
What it changes: You will not see this in a chatbot any time soon. The idea is that one day we could check which model really answered. Anthropic itself presents it as a research prototype, with the limits listed in its article.
Anthropic — Progress on provable inference for large language models (November 2)
DeepSeek Harness: building plugins by chatting
On November 2, DeepSeek showed the “Creator Mode” of Harness, its open agent platform (MIT license, in public preview worldwide). You describe the tool you want in plain language, for example a scientific paper reader, a poster maker or a dashboard of your conversations, and the agent writes the plugin, installs it and checks it. Harness comes as an app for Mac (Apple chips) and Windows, as a web version, and the code is on GitHub. DeepSeek warns that its core plugins and interfaces are still evolving. It is a bit like ordering custom furniture by describing it to the carpenter, except the carpenter sometimes needs several tries.
What it changes: Handy if you like tinkering, and nothing changes for everyone else. As always with a program that installs extensions on its own, it is wise to check what it added.
DeepSeek — Harness TechNode — DeepSeek showcases Creator Mode for building custom plugins through conversation (November 3)
Apple Intelligence tested in China, with Alibaba's Qwen
According to TechNode, which cites Apple's Chinese-language site, Apple announced on October 31 that Apple Intelligence will be tested in mainland China from November 4. Alibaba's Qwen model will be built into Siri, Writing Tools and Visual Intelligence. The features arrive first in an iOS developer beta, with a wider launch planned for December with iOS 27.3. Same phone, different brain depending on the country.
What it changes: No effect in France, where Siri AI remains unavailable for now. We could not open Apple's page: the dates come from TechNode.
This week’s analysis
Edition opened on Wednesday, November 4, with news from Monday 2 and Tuesday 3. It will be completed on Friday. The trend: labs talk more and more about how AI takes part in building itself, and about ways to control what comes out of the machines. Read these statements for what they are, announcements from the companies themselves.