OpenAI publishes 722 mathematical manuscripts

In its [announcement](https://openai.com/index/sharing-ai-progress-in-mathematics/), OpenAI reports publishing 722 manuscripts from an unreleased frontier model, grouped into 372 result families. Its [repository](https://github.com/openai/math) includes Lean formalizations for many, but not all, manuscripts: those provide computer-checkable proofs, while unformalized results may contain errors.

Mistral Large 4: API first, weights planned for October 27

Axios

Mistral's new multimodal model has one trillion parameters, with 49 billion active at a time. It is initially available through a moderated API; select partners get a less restricted version for testing, and public weights are planned for October 27 after further training and safety tests.

Lambert challenges the case for banning open models

Interconnects AI via Techmeme

Nathan Lambert argues that GLM-5.3's open release has yet to produce major attacks despite Anthropic's warnings about its cyber capabilities. His essay examines the trade-offs behind restrictions on open weights.

Personal Agent Protocol targets agent-to-business interactions

CNBC via Techmeme

Meta and Sierra have published the Personal Agent Protocol with Stripe and Shopify, as well as Walmart and other companies. The group's stated aim is to standardize and secure how personal agents interact with businesses.

OpenAI demonstrates decisions from a fixed set of choices

OpenAI · video, 5m34s

The Decisions API uses GPT-6 Luna to select one predefined answer from text or image input. OpenAI demonstrates server-side responses under 100 milliseconds for routing and control tasks; the API complements full reasoning models for more complex workflows.

Strands Decider brings constrained choices to local hardware

Strands Agents

Strands Decider is an open-source model with two billion parameters that selects among supplied options. The [release post](https://strandsagents.com/blog/introducing-strands-decider/) reports a median decision latency of about 115 milliseconds on an RTX 3090, with latency increasing as tasks get longer.

Finland orders Google to stop two data-center builds

CNBC via Techmeme

Google's subsidiary Tuike Finland Oy must halt construction in Muhos and Kajaani by October 23 under an order from Finland, pending environmental impact assessments. The order affects two specific projects in Google's Finnish expansion.

FT reports SpaceX is seeking $40 billion for Nvidia chips

Financial Times via Techmeme

SpaceX is seeking $40 billion in bank loans and investment-grade debt to buy Nvidia chips, according to the Financial Times, which cites sources. Apollo is leading the proposed deal; the report describes financing being sought, rather than a completed transaction.

Temasek warns about an unwinding of the AI trade

CNBC

Singapore's state-owned investor Temasek identifies an unwinding of the AI trade as the biggest risk facing markets. Its warning addresses the financial exposure accompanying the infrastructure buildout.

Check the balance before inference consumes it

AI Engineer · conference talk, 20m8s

A conference talk on validating entitlements before consumption, reserving credits, and settling actual usage asynchronously. It explains why shared agent budgets need concurrency control, idempotency, and an auditable ledger.

Resolve AI's approach to incident investigation

AI Engineer · conference talk, 21m5s

Resolve AI describes a production harness that combines model routing, scoped context retrieval, least-privilege access, and evaluations. Root-cause claims must follow a verifiable evidence chain; incomplete chains trigger lower confidence and alternative diagnostic paths.

Langfuse connects production traces to improvement loops

AI Engineer · conference talk, 17m6s

Langfuse demonstrates a changelog-writing agent that uses review feedback to update evaluation criteria and test proposed changes. Humans retain control over dataset alignment and modification review to limit overfitting.

Tessle separates autonomy from deployment trust

AI Engineer · conference talk, 20m31s

Drew Knox distinguishes how often agents need corrections from how much deployment a team trusts them to automate. His conference talk separates fast checks during drafting, expensive checks after a pull request, and broader feedback loops that improve both.