China's draft controls target model weights
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Monday, November second.
In today's briefing we see China's draft export controls reaching model weights, SemiAnalysis ranking agents on a new benchmark where the best score is twelve percent, and AI-found bugs pushing back Postgres 19.
First up - Today in the big model news;
OpenAI
OpenAI is preparing a PR push after the midterms, according to TechCrunch citing Axios. Chris Lehane called the midterms a movie trailer for 2028 and said communities should hold companies accountable. Ideas floated include AI speed limits and renaming data centers as AI compute campuses. Amazon and Microsoft say they will stop using NDAs in local data-center negotiations. Siting and power have become a political liability for the build-out that capacity depends on. Capacity and pricing plans should assume local-permit friction, not just chip supply.
Google
Search queries at Google hit a record, but Search revenue growth slowed to twelve percent, to about sixty-three billion dollars. Alphabet shares are down roughly twelve percent over five days, with capex guidance of up to two hundred and five billion dollars. CNBC reports Google Cloud grew over one hundred percent to thirty-one billion dollars, a figure not cross-checked. If it holds, the weakness is specific to Search, and it is a monetization problem rather than a demand problem.
Anthropic
Anthropic has updated its usage policy, effective November twelfth. Where Claude drives autonomous physical hardware, a qualified operator must be able to stop it, and the equipment must hold a safe state if Claude disconnects. Claude also may not decide or recommend who to investigate, arrest or charge. Supported-region access is barred for entities incorporated, headquartered or majority-owned in unsupported regions. Anthropic calls the weapons and surveillance changes clarifications of past enforcement, while CNN reports a new ban on sustained, needless cruelty toward models. The ownership clause reaches any customer with an offshore parent or investor, so check corporate structure before the effective date. Robotics, drone and public-safety workflows need operator-stop and safe-state design confirmed by then.
Meta
Meta's ad revenue rose twenty-six percent to sixty-three point one billion dollars, while Google Search and other grew twelve percent to sixty-three point four billion. That is a fourteen-point gap, Meta's widest lead since 2021. Meta says advertisers following its AI assistant's recommendations saw almost five percent more conversions, and nearly ten million advertisers used at least one Meta AI tool. Shares are still down about four percent over five days, and capex guidance runs up to one hundred and forty-five billion dollars. Meta's evidence comes from performance-campaign automation, where the buyer measures conversions directly. Google's AI surface sits in front of the click, where nobody has yet shown the ad pays the way the old results page did.
In Local model developments,
There continues to be a shift in who controls access to frontier capability, and China is now moving on the exporter side. The commerce and science ministries issued a draft adding forty technologies to China's export catalogue, twenty-three of them AI.
Frontier-model technology is restricted if it meets any one of three tests: at least ten to the twenty-sixth training operations, at least two trillion parameters, or a global top-ten position at export time.
API, app and SaaS access for overseas customers is not covered. But local deployment of a restricted model outside China needs a licence, and interim pre-export registration is already in effect.
That is the part that matters for open weights: the cheap-weights path now carries licence risk.
The two trillion line would reach Kimi K3.1 Pro, at two point eight trillion parameters, whose weights are promised for November fourth. That link is an inference, not stated by any source.
Comments close November fifteenth. The Kimi release will show whether its licence terms survive the rule. Keep a tested non-Chinese fallback until then.
In the harness, tools and orchestration world;
SemiAnalysis published GDPx version one, a benchmark of two hundred and six agent tasks from fourteen data partners across eleven domains. Tasks are graded on the first attempt across three tries, with an eight-hour limit. Opus 5.5 leads at twelve percent, ahead of Gemini 4 Carbon at about eleven and GPT-6 Astra at about ten. Finance and legal are near zero under binary grading. Safety classifiers depress the Opus and Astra scores, so a second ranking excludes cyber, biology and alignment. The authors disclose that they co-founded an RL-environment startup that sold data to frontier labs, and that chip design is deliberately overweighted. No independent assessor has weighed in. Muse Wonder Full leads on cyber, and Opus leads on quant trading and chip design. Use it for domain-specific model choice, not as an overall leaderboard.
In AI Infra,
Postgres committer Robert Haas says version 19 will likely slip to November or later, after three major features were reverted late in beta. Back-patched commits are up seventy-four percent: four hundred and thirty-three on the version 18 branch in 2026, against two hundred and forty-eight on the version 17 branch in 2025. He credits a large number of people reporting bugs found by LLMs. LLM-drafted fixes are not always correct or standards-compliant, so human review is the bottleneck. His master-branch comparison is internally inconsistent and deserves caution. Detection is scaling faster than the people who verify and merge fixes. Budget triage and review capacity alongside agent-driven security scanning, because the queue lands on maintainers.
Quick hits from the consumer side;
Apple CEO John Ternus told Wired the company will not ship AI for AI's sake. Apple Intelligence features for China, with Alibaba's Qwen supporting selected ones, enter developer testing November fourth ahead of a December iOS 27.3 rollout, per a secondary digest.
That's the briefing. Have a great day, and don't forget to subscribe.