Five stories from today: a draft Chinese law that puts a number on "frontier model," Google's news-licensing pilot doubling in a month, and a cheap DeepSeek update outrunning bigger models on agentic work.
- China drafts export controls that define a "frontier model" and carve out open weights
- OpenAI retreats from the browser and moves into Chrome's sidebar
- Google's AI news pilot passes 400 publishers as CNN, FT, Axios and Forbes sign on
- DeepSeek's "minor" Flash update beats larger rivals on agentic coding
- Google shows magic state cultivation on Willow, discarding 92% of attempts
- Tools worth your time
China drafts export controls that define a "frontier model" and carve out open weights
China's commerce and science ministries opened a consultation on 16 October that would add 23 AI technologies to the country's export-control catalogue and, for the first time, put numbers on what counts as a "frontier model."
The draft defines the category as a model with at least 2 trillion parameters, 10^26 FLOPs of training compute, or a place in the global top ten by ranking — thresholds that Kimi K3 and Qwen3.8-Max reportedly already cross. It would fast-track open-weight releases with automatic licensing, exempt models that are already public, and ban AI that can autonomously write or alter malware. Notably, it writes the open-weights carve-out into the control regime rather than treating open release as a loophole to close.
This is a draft, not law: comments run to 15 November, and Beijing has floated similar restrictions before without enacting them. Worth watching is less the thresholds themselves than the precedent of a government encoding a FLOPs-and-parameters definition of "frontier" into trade rules, which others may borrow.
OpenAI retreats from the browser and moves into Chrome's sidebar
OpenAI split its ChatGPT Chrome extension, which it says has more than 7 million users, into two modes: a Chat mode that reads across open tabs, and a Work mode that can act on them — sending messages, filling forms, downloading files and grouping tabs. Users can also tag @chrome directly inside ChatGPT on web and desktop.
The move partly recovers reach OpenAI lost when it retired its standalone Atlas browser. It is a quieter bet than owning the browser outright: rather than fight Chrome for the window, OpenAI is taking the sidebar and the agentic actions, and leaving Google the chrome around it. Whether users want an assistant with write access to their open tabs is the open question.
Google's AI news pilot passes 400 publishers as CNN, FT, Axios and Forbes sign on
Google said on 15 October that its paid News AI licensing pilot now covers more than 400 publications, roughly doubling in a month. New signings include CNN — its first such deal since the Paramount Skydance takeover — alongside the FT, Axios, Forbes and a set of US local outlets.
Licensed content feeds AI-generated "story previews" and customisable audio briefings in Google News. Each preview carries a disclaimer that the copy is AI-generated and "may include mistakes." Google declined to disclose terms, so the central unknown is what any of these publishers are actually being paid. The disclaimer is also instructive by contrast: it sits as a line to be read past, where a visible machine-written mark treats the authorship as a fact about the piece rather than fine print.
DeepSeek's "minor" Flash update beats larger rivals on agentic coding
DeepSeek released V4.1-Flash-1017 on 17 October under the same name, price and architecture as its predecessor, but with double-digit gains on agentic benchmarks — Terminal-Bench 4.0 up about 11 points — and faster responses with shorter outputs.
DeepSeek frames it as a Pareto improvement: roughly 77% on DeepSWE at under $0.20 per task, against a reported $2–10 for GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5. The weights are MIT-licensed on Hugging Face. Alongside the model, DeepSeek opened its in-house Search and Fetch APIs to outside developers for the first time at $0.75 per 1,000 calls, free in several harnesses until 31 October. The benchmark gains are DeepSeek's own figures and worth independent confirmation, but the cost gap, if it holds, is the part competitors will feel.
Google shows magic state cultivation on Willow, discarding 92% of attempts
A Science paper published on 15 October reports that Google Quantum AI demonstrated "magic state cultivation" on its 105-qubit Willow processor, cutting errors roughly 40-fold and reaching 99.99% logical fidelity on the T-states needed for universal quantum computing.
The headline number comes with heavy conditions. The protocol discarded about 92% of attempts, keeping only around 8%, and a reinforcement-learning agent spent roughly four hours tuning control parameters first. This is hardware validation of one known bottleneck, not a working quantum application: fidelity still needs improvement by orders of magnitude, and the discard rate is a cost someone has to pay. Useful progress on a component, read as such rather than as a finished capability.
Tools worth your time
- Eney by MacPaw — a proactive desktop and iPhone assistant that reads context across apps, files, calendar and meetings and prepares the next step before asking for approval, built on MacPaw's own stack including a memory layer it calls Mnemos. For Mac and iPhone users who want on-device action rather than another chatbot.
- SuperLocalMemory — an open-source, local-first shared-memory layer (SQLite-backed, MCP-compatible) that lets Claude Code, Cursor, Codex and ChatGPT share one synced context space on hardware you control, with bi-temporal fact tracking and role-based access. For developers running multi-agent setups who want data sovereignty.
- Mercury CLI — a terminal coding agent whose beta.35 adds keyless web search with citations, an on-device embedding model for recall by meaning, and session memory that loads pinned rules at start. A smaller independent project, so treat it as one to watch rather than a settled pick.
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
