OpenAI pitches "Private Intelligence" for sensitive business work. In a post on X on October 30, OpenAI announced OpenAI Private Intelligence, a bundle of enterprise data protections for businesses using its most capable models — including Zero Data Retention for eligible API workflows and Private Inference, a confidential-computing option still in preview. Its pitch, in a few words: "Privacy should advance as fast as intelligence." The company says it built the protections with some of its largest customers. For business readers: if client confidentiality has kept your most sensitive work off frontier models, this is OpenAI's answer — but note the eligibility and preview qualifiers before assuming full coverage. Posted by @OpenAI on X, October 30; product details as summarized from OpenAI's Private Intelligence page.

Altman: robots should build the data centers. On a podcast appearance reported October 30, OpenAI's CEO said the company will aim its robotics program at building data centers — robots running chip factories, building turbines and solar panels, and building the data centers themselves — calling self-replicating data centers "within grasp." He also said OpenAI has made its own chip and is learning to build its own data centers. For business readers: the company buying the most compute in the world is now talking about automating the construction of that compute — a signal about where AI infrastructure costs may head. As reported from Altman's appearance on the Cheeky Pint podcast.

Mistral opens the weights of its trillion-parameter Large 4. Mistral released full open weights for Mistral Large 4, a 1-trillion-parameter multimodal model with 52 billion active parameters, aimed at agentic coding and defensive cybersecurity work — meaning organizations can run it on their own infrastructure rather than through Mistral's hosted service. The release follows an October preview and lands in a crowded open-model week. For business readers: another frontier-class model you can self-host, with the security, cost, and operational ownership that choice brings. Announced October 29; as reported.

One opt-out for AI training, backed by all four labs. The four companies announced on October 30 that they will honor a single robots.txt token — __usage_ai-training — so a website can opt out of AI training with all of them in one line, instead of blocking each crawler separately. The signal comes from an IETF draft written by Anthropic's Kevin Kelley. Timelines differ: Anthropic says it has already started, OpenAI says "the coming months," Microsoft targets February 1, 2027, and Google says 2027. Meta's communications head Andy Stone signaled support on social media without committing Meta to honor it. Why it matters here: this site runs a robots.txt too — one line instead of four is a real saving for every small publisher. Announced on the companies' official channels; reported by The Verge.

Altman on the delayed Astra release. In posts on X reported this week, OpenAI's CEO called Astra a "powerful model" and said the company still intends to make it generally available — but its cybersecurity capabilities mean "we need a little bit longer to do this safely," adding "but hopefully not too long." No release date was given. For business readers: the flagship model in our estimator's price table is being held back on safety grounds, by the company's own account. As reported from Altman's posts on X.

The rival labs say they have been coordinating on safety. The three labs confirmed this week that they spent several weeks coordinating on AI safety, jointly unveiling cyber-focused models, safeguards, and access programs — a public break from their usual compete-don't-collaborate posture. Details of how the coordination continues have not been fully disclosed. Reported October 29; treat specifics as provisional until the labs publish their own full accounts.

Karpathy's warning to agent builders. In remarks circulating on X on October 30, Andrej Karpathy — co-founder of OpenAI, now building at Anthropic — argued the biggest mistake in AI right now is forcing agents to work before mastering the harness around the model: the prompts, tools, and checks the model runs inside. His claim: OpenAI made that mistake in 2016 and it cost years. It matches this site's own advice — the setup around an agent matters more than the agent's brand name. Quoted from a talk clip shared on X; wording as reported by viewers.