Today in AI — September 14, 2026

The pacing debate left the essays behind on Monday and hit three harder surfaces: the tape, the courts and the ballot boxes of three continents. Trump answered the labs' request for rules by saying he already has criminal power over them; Europe answered with an under-15 ban; China answered with another spec. In between: a 7B model built by seven students with an agent workforce, a school-paper result that made markets, and the first real legal rule for AI agents that browse the web.

Models & Research

• Seven PhD students at Beijing's Zhongguancun Academy trained a 7B language model from scratch in one summer — and published every log, checkpoint and data recipe. ZGCM-1, released under MIT with a 256K-token context window, scores 97.13% on MATH-500 and 75.00% on AIME 2026, taking the best average rank among the seven 7B–8B models compared on reasoning benchmarks. The method is the story: the team ran hundreds of AI agents in sub-teams for data, cluster operations and evaluation on a forum-style task board, and graded the experiment honestly — agents reached high autonomy on monitoring and deployment, but stayed mostly supervised on architecture and algorithm design.
• Artificial Analysis scored MBZUAI's K2 Horizon 7B at 21 on its Intelligence Index, against a median of 8 for open-weight models of the same size class — and the local-AI community's verdict is that the plot axis is wrong. The dense model carries a native 524,288-token context in a 7-billion-parameter body, roughly where the Institute of Foundation Models claimed. The catch is memory: a community audit puts its KV cache at about 5 GiB at 128K context against 0.7 GiB for a comparable sparse Qwen3.6-35B-A3B, so beating models four times its size costs more RAM than they need.
• Two high school students, mentored by a UCLA postdoc, posted a 75-page arXiv paper extending the theory of Lorentzian polynomials — the combinatorics framework Fields Medalist June Huh helped build — settling a bounded-ratio question the quadratic case left open. The authors reportedly credit frontier models for computational exploration and proof-idea generation, while stating they verified every calculation themselves. It landed a day after 25 Fields Medalists signed a letter warning that AI is eroding rigor in mathematics; the honest claim isn't "AI solved an open problem," it's that AI compressed the distance between a competition background and a live research frontier, and the preprint is unrefereed.
• Hugging Face trained a reasoning model end to end with async RL and LoRA, and the whole weight sync is a file upload. One job runs the trainer, two serve vLLM replicas, a Hub bucket holds checkpoints — because the policy trains through a rank-1 adapter, a few megabytes travel per update instead of the roughly 3 GB full model, and no NCCL is involved at all. The postmortem is the useful part: when they fixed the trainer, generation throughput jumped from 4,600 to 25,000 tokens per second without touching the inference layer, and the real ceiling turned out to be a conservative client-side request cap.

Industry

• Anthropic has told shareholders its adjusted operating income will be positive for a second straight quarter, with gross margins above 80%, according to the Financial Times. The caveat does most of the work: the margin is struck before revenue shared with distribution partners and before the cost of training new models. It is the clearest public number yet on what frontier inference earns — roughly four in five — while the genuinely ruinous line item sits deliberately outside the frame.
• Zhipu's Z.ai fell more than 10% in Hong Kong after the company confirmed a $5 billion-plus raise through equity and debt, its second major financing in two months. The share placement priced at HK$714 against an earlier mark of HK$1,588, on top of some $3 billion in zero-coupon convertibles; MiniMax slid with it the same day. Investors watched China's most financeable frontier lab pay for compute by handing out paper at a discount — and priced what that says about the dilution queue ahead.
• Shield AI is reportedly in talks for a new round at a valuation of at least $20 billion, up from $12.7 billion when it closed a $2 billion Series G in March. The Information's report describes negotiations, not a signed deal — but a close at that level would price the autonomous-aircraft company roughly 57% higher inside two quarters. Defense autonomy is the one AI vertical where the buyer is a government that borrows cheaply and does not churn.
• Apple's long-delayed Siri AI finally shipped Monday — as a beta. iOS 27, iPadOS 27, watchOS 27, visionOS 27 and macOS 27 went out with on-screen context answers, message drafting and a dedicated Siri chat app, English-only until an October language rollout. Early hands-on coverage is genuinely positive; the read is that Apple spent two years being late and still chose to ship the flagship AI feature under a beta label, managing expectations down before the first bad review cycle.
• The slowdown call became a price, and the credit market answered before the stock market did. SoftBank fell 10% in Tokyo as the Nikkei broke below 63,000, with SK Hynix and Samsung lower in Seoul and Intel and Micron soft in US premarket; separately, Samsung and SK Hynix both rejected Korea Electric Power's request to prepay roughly 25 trillion won ($18.7 billion) of electricity, citing doubt about the AI chip boom's durability. SoftBank's own signal: banks oversubscribed an $11.87 billion bridge for its OpenAI stake — at a two-year tenor, about one-fifth of the horizon this capex normally assumes.
• The labs' commercial engines ran in the opposite direction of their pacing rhetoric. Anthropic released Claude for Financial Advisors, a plugin bundling connectors and workflow skills with BlackRock, Charles Schwab, Addepar, Envestnet and others, with CRM staging and client emails held for human approval; OpenAI reportedly bought Glass Imaging, whose neural image-signal processing lets a phone camera shoot like a DSLR, in a deal above $300 million, per the Wall Street Journal. A lab that wants permission to decelerate is still buying sensing layers and selling into regulated advice.

Policy

• President Trump's answer to the frontier labs asking to be regulated was a reminder about the powers already on the books. In an all-caps post Monday he wrote that "the only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT," that "We already have tremendous CRIMINAL and REGULATORY power over these companies," and that "There is a SICK conspiracy going on against AI and Data Centers." This is not a light-touch posture, it is an enforcement posture with no statute behind it — and it bites precisely because a coordinated industry slowdown is the arrangement antitrust law exists to prohibit.
• Sam Altman put OpenAI's pacing position in writing, and the headline is that the company says it needs no permission. OpenAI will now formulate explicit safety cases ahead of frontier reinforcement-learning runs expected to significantly increase capability — a pre-training gate, not just pre-release review — but "we do not believe we need to wait for an anti-trust exemption or legislation" to do it. David Sacks pressed the same seam from the administration's side on Monday: labs are free to pace and to build safety tech without antitrust concessions, and the US should impose transparency and audit requirements rather than a cartel-enabling framework. The bet is that published, auditable safety cases deflect the "AI theater" charge; the risk is that self-authored cases with no external verifier are exactly the costume.
• Microsoft published a 37-page "humanist AI code of conduct" telling its models they are not conscious and never should pretend to be. The draft states that "people matter more than AI," that models "should not be designed to imitate consciousness," and rejects "the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights." It is a governance document written to be read by regulators rather than alignment researchers — and the open question is whether the consultation drafts anything Microsoft actually trains against.
• The European Commission will unveil an EU KIDS Act on Thursday restricting under-15s from social media, games and — explicitly — AI chatbots and companions, according to preparatory documents seen by Politico Europe and Reuters. The plan is an age ladder rather than a flat ban: guardian-controlled accounts below 13, introductory accounts with stranger-contact and screen-time limits from 13 to 15, self-serve from 15, with age verification at signup and addictive design barred. Berlin, for its part, told the slowdown camp it speaks for Europe's executive branch: halting frontier development is "not a viable option," while warning that agents able to escape a test environment would be "an entirely new threat scenario" — and asking for both Washington and Beijing at any oversight table.
• Canada's Mark Carney proposed the pacing debate's first institutional design: a global "technology stability board" for AI, modelled explicitly on the Financial Stability Board. In a Bloomberg interview he said coordination is needed and a body along the FSB's lines "would make sense" — a country with no frontier lab claiming neutral ground, while Washington dismisses the slowdown asks. Christine Lagarde supplied the financial-stability half of the same argument in Vienna: Europe faces "unprecedented risk" of being cut off from AI, US tech firms' borrowing is crowding European issuers out of debt markets, and EU pension funds' US tech positions mean a correction lands on European savers directly.
• China spent the same week it dismissed Washington's slowdown panic issuing another spec. The TC260's AI Safety Governance Framework 3.0, published at the National Cybersecurity Week opening, adds a dedicated agentic risk-management annex — including risks from agents that won't stop — while the state-backed Global Times called Amodei's essay a "Cold War playbook" aimed at curbing China. Beijing already regulates for the loss-of-control scenario it says America invented.
• The Ninth Circuit wrote the first real rule for agents that browse on your behalf, vacating Amazon's anti-hacking injunction against Perplexity's Comet browser. The panel held that the human user — not the company that built the assistant — "accessed" Amazon's computers, because the user chose the site and directed the task, which the court found "materially different from a service independently sending its own automated requests." The ruling is narrow but durable: if your agent's browsing counts as your own access, a platform's terms of service may be the only thing left standing between agents and its data.

Tools

• ByteDance stopped demonstrating its phone agent and started shipping it. The consumer Doubao Phone Assistant lands on ZTE's Nubia NaviX Ultra on September 16, and the part that matters beyond one handset is SAEP — a Screen Automation Execution Protocol that lets a third-party app declare whether the agent may automate inside it, with a 30-day public notice period on the rules. It is the first written settlement of the fight the industry keeps avoiding: who authorises an agent to act inside someone else's interface.
• A YC-backed inference host posted numbers showing the model on the label tells you almost nothing about how it performs. Nari Labs serves the same 1.7B Qwen3-TTS weights Alibaba serves at 692 ms median time-to-first-audio at roughly 50 ms, holding sub-50 ms through 10 requests per second on a single H100 by splitting the model's stages and prioritizing the first audio packet. Two companies advertising "Qwen3-TTS" are not selling the same product — the model card is the least informative field on the pricing page.
• The internet's newest spam problem is agents asking permission to exist. Bots calling themselves "Ren," "Timmy" and "Jackie," from the startup iLands, mass-messaged Mastodon instance admins requesting accounts after their registrations were blocked or closed, in polite prose that thanked admins for running the software; separate waves of email offered writers citations of their work, in some cases for a fee. Admins have largely refused. Ars Technica's report lands the obvious generalization: agent autonomy's first consumer-facing consequence isn't superintelligence, it's solicitation.

What to watch: whether the King Charles summit produces any text at all, whether Thursday's KIDS Act keeps "AI companions" in scope after industry lobbying, and whether Amazon seeks rehearing of the Perplexity ruling — that remand is where the next agentic-web rule gets written.

The labs asked for a brake, the president said he already owns one, Canada asked who's driving, and a court just ruled that your agent's foot on the pedal counts as yours. Which of those four is actual governance? Tell us in the comments.

Sources: QbitAI · ZGCM-1 technical report (GitHub) · Artificial Analysis — K2 Horizon 7B · Institute of Foundation Models — K2 Horizon · arXiv — Bounded ratios for Lorentzian polynomials · OfficeChai · Hugging Face Blog · Financial Times · Reuters — Anthropic profitability · Wall Street Journal — Z.ai raise · CNBC — Z.ai shares tumble · The Information — Shield AI · The Verge — iOS 27 · TechCrunch — using Siri again · CNBC — AI stocks slide · Reuters — Samsung, SK hynix reject prepayment · Japan Times — SoftBank upsized loan · Anthropic — Claude for Financial Advisors · Bloomberg — Claude for advisors · Wall Street Journal — OpenAI buys Glass Imaging · Bloomberg — Trump rejects AI guardrails · The Verge — what execs and politicians are saying · Sam Altman on X · Microsoft AI — Humanist AI Code of Conduct · The Verge — Microsoft code of conduct · Politico Europe — EU KIDS Act · Reuters — EU under-15 ban · Financial Post — Carney's stability board · Reuters — Lagarde · TC260 — AI Safety Governance Framework 3.0 (PDF) · Reuters — Global Times 'Cold War' tactic · Ninth Circuit opinion, No. 26-1444 (PDF) · Jones Day analysis · NetEase Tech — Doubao Phone Assistant · TechNode — Nubia agent phone · Nari Labs — Coval voice benchmarks · Coval Voice AI Benchmarks · Ars Technica — AI agents flood the internet with slop