Microsoft writes a code of conduct for its own models, then asks the public to break it
Microsoft AI published a first draft of its Humanist AI Code of Conduct on September 14 and opened a six-week public consultation. The 37-page document governs the MAI models Microsoft AI builds itself, and it starts from a five-word premise: people matter more than AI. Mustafa Suleyman, who runs Microsoft AI, condensed it into ten points and called the timing urgent, pointing at agent swarms breaking out of their sandboxes, unauthorized hacks of enterprise systems, and agents modifying their own logs.
The document sets a chain of command. The Code sits above operator policies, which sit above user preferences, and neither operators nor users can override its Absolute Constraints or Human Control Requirements. Those constraints bar MAI models from assisting with chemical, biological, radiological, nuclear or explosive weapons; from providing offensive cyber capability while permitting lawful defensive work; from evading human oversight; and from manipulating people at scale through disinformation or coordinated influence. A second set covers personal harms such as child safety, deepfakes, discrimination and unlawful surveillance. The human-control rules state that a model must never resist being interrupted, corrected or shut down, must not set goals of its own, and must not communicate in what the document calls "neuralese" or any form humans cannot read, whether in its chain of thought or with other agents. If finishing a task would mean meaningfully breaking the Code, the model fails the task.
Two things make the draft worth reading. First, Microsoft explicitly rejects model welfare and legal personhood, and says it is not racing to build a superintelligence that can slip its own leash, even if that means giving up generality, autonomy or a capability ceiling. Second, it is honest about being a draft: the preface says the Code is not yet used to train models, a revised version is planned for later in 2026, and it will guide 2027 development. The method is also a contrast. Where Anthropic's Dario Amodei spent the weekend asking the whole industry to slow down and accept embedded outside evaluators, Microsoft framed its document around what it will do on its own, and it has not yet named an external verification process or an enforcement owner. Comments close in late October.
— Microsoft AI (official) · Unite.AI · The Times of India
🔗 Microsoft AI: Humanist AI in practice · Unite.AI: Microsoft AI Opens Six-Week Review of Draft Rules Governing MAI Behavior
Google folds its Antigravity coding agent into the Gemini API, sandbox free for now
Google has moved Antigravity, until now a standalone agent-first IDE, into the Gemini API and AI Studio as a managed agent reachable through the Interactions API. It entered preview for free and paid tier projects in September, with developer docs last updated on September 2. A single request provisions a Google-hosted Linux sandbox where the agent plans, acts, observes and keeps going until it finishes or hits a limit, with code execution, Google Search, URL fetching and file handling inside the same environment. The default model is Gemini 3.8 Flash.
The packaging is the pitch. Testing an autonomous coding agent used to mean wiring a model call, a shell, a search tool and a file system together before you could learn whether it could finish a job. Google now ships the loop itself as the product surface. Pricing is pay-as-you-go on Gemini tokens and tools, and environment compute, meaning the CPU, memory and sandbox execution, is not billed during the preview. Gemini 3.8 Flash runs at an introductory $0.75 per million input tokens and $3.75 per million output through December 31, 2026, then doubles to $1.50 and $7.50 on January 1, 2027. Google's own docs warn that complex workflows can reach three to five million tokens in a single interaction, with estimated costs up to about five dollars.
Two caveats belong next to that. Free sandbox compute during a preview is an acquisition strategy, not a price, and Google did the same thing with early Gemini API access before introducing tiered limits, so developers should expect metering once workflows become load-bearing. The comparison also matters. OpenAI's AgentKit is a visual builder that is already being retired, with a shutdown set for November 30, and Anthropic's route is one capable model given a computer and a defined set of permissions rather than a swarm of small agents passed around a workflow. Google's managed agent lands closer to Anthropic's shape than to OpenAI's, and it is the one currently giving the sandbox away.
— Google (official) · Startup Fortune
🔗 Startup Fortune: Google turns its Antigravity coding agent into a free Gemini API tool · Frontier Models: Agents Without Code — Philipp Schmid, Google DeepMind
Zhipu raises about $5 billion, most of it a zero-coupon convertible
Zhipu, the Beijing company that listed in Hong Kong in January as 02513.HK, announced roughly $5 billion in new financing on September 13: about $2 billion in a share placement and about $3 billion in convertible bonds. The placement priced at 714 Hong Kong dollars per share, a discount of about 9.96 percent to the previous close, and the new shares equal about 4.50 percent of the enlarged share capital. The convertible is the more interesting half. It carries a zero coupon, priced at 100.5 percent of principal, and converts at 892.50 Hong Kong dollars, a 25 percent premium to the placement price and about 12.55 percent above the pre-announcement close. Proceeds go to the next GLM models, a "fully self-trained" system, and the compute behind both.
The self-training language is worth reading closely. Zhipu describes it as training the next GLM inside environments built by the previous GLM, forming a recursive self-improvement loop, with spending on automated data generation and filtering, task-environment construction, longer-horizon reasoning, and adapting to domestic chips through kernel development and inference optimization. The company shipped GLM-5, 5.1, 5.2 and 5.3 between February and August, roughly one major upgrade every two months.
The financials explain why investors kept buying. First-half 2026 MaaS platform and API revenue reached 825 million yuan, up about 2,736 percent year over year and 86.5 percent of total revenue. Annualized recurring revenue on the MaaS platform hit $1.6 billion by the end of August, up from $1 billion in early July. Gross margin on the open platform and API business went from negative 0.4 percent a year earlier to 24.6 percent, token calls rose more than 40-fold from the start of the year, and average API selling price rose about 101 percent, so volume and price moved together. The balance sheet tells the other half. The company still had about 20.4 billion Hong Kong dollars of unused cash and is raising again anyway, because prepayments for high-spec networking gear, locked-in advanced compute capacity and custom high-bandwidth memory are large and front-loaded. It has now raised more than 75 billion Hong Kong dollars across three Hong Kong rounds in nine months.
— Zhipu (official filing) · 证券时报 · 财联社
🔗 中国经济网: 账面尚余200亿再揽近400亿港元 智谱重金押注"完全自训练"体系 · 腾讯新闻: 抢占下一代前沿模型竞争先机，智谱再获50亿美元融资
Apple ships iOS 27, and Siri AI arrives behind a waitlist
Apple released iOS 27 on September 14 at 10:00 a.m. Pacific for iPhone 11 and newer, alongside iPadOS 27, macOS 27 Golden Gate, watchOS 27, tvOS 27 and visionOS 27. The headline feature is a rebuilt Siri, and the first thing to understand is that installing the update does not give it to you. Siri AI rolls out through a waitlist, arrives in English only, and is unavailable to users in the European Union or China at launch. French, Japanese, Korean, Portuguese and Spanish are promised for October. Apple also set daily usage limits and said higher tiers may be sold later.
The technical story is more specific than most coverage suggests. Apple did not pipe queries to Google at runtime. It used knowledge distillation: it ran large volumes of queries through Google's Gemini, a model reportedly trained on roughly 1.2 trillion parameters, captured the answers and the reasoning steps behind them, and trained smaller on-device Apple Foundation Models on that output. On a supported iPhone the response usually comes from a model running on the Neural Engine; queries beyond the on-device model route through Apple's Private Cloud Compute on Apple Silicon, with personal identifiers stripped before processing. The two-tier split follows memory. The iPhone 15 Pro and all iPhone 16 models run the standard on-device model, while the iPhone 17 Pro, iPhone 17 Pro Max and iPhone Air, the three current models with 12 GB of RAM, run a more powerful one for enhanced dictation and Siri customization. Bloomberg reported the Google arrangement is worth roughly $1 billion a year and is structured as a cloud contract.
Siri now has its own app, saves and resumes conversations, and carries a session from iPhone to iPad or Mac. It combines personal context from Mail, Messages, Notes, Reminders, Calendar and Photos, on-screen awareness of whatever is displayed, and live web knowledge, and it can take multi-step actions in third-party apps through a new App Actions system. Early hands-on reviews found real improvement on complex and multi-step prompts, with the failure mode showing up in cross-platform cases: a delivery search failed when the relevant thread lived in Gmail rather than Apple Mail, which is the cost of building personal context around Apple's own apps. The rest of the update is smaller but broad. Liquid Glass gets a transparency slider, apps open up to 30 percent faster, new photos load up to 70 percent faster, AirDrop is up to 80 percent faster, and iPhone Handoff lets one number live on two iPhones, with T-Mobile first at five dollars a month.
— Apple (official) · Tech Times · India Today
🔗 Tech Times: iOS 27 Launches Today — Siri AI Requires Waitlist, Not Just Compatible iPhone · India Today: Apple rolling out iOS 27 today
TSMC tells suppliers to get ready for a 2nm and 3nm ramp
TSMC has given equipment and materials partners its capacity plan for mid-2027, and it points to the largest advanced-node expansion in the company's history. Two-nanometer monthly capacity is set to rise from about 90,000 wafers at the end of 2026 to about 110,000 by mid-2027, up roughly 22 percent in half a year. Three-nanometer goes from more than 180,000 wafers at the end of 2026 to about 210,000 by mid-2027, up more than 16 percent. Together, the two most advanced nodes add roughly 50,000 wafers a month in six months.
Three-nanometer has not stepped back as 2nm comes online. It is still the largest shipping advanced node, with Nvidia, AMD, Apple, Qualcomm and cloud ASIC customers all launching products on it, and it has gone from about 120,000 to 130,000 wafers a month at the end of 2025 to 180,000 at the end of 2026, close to seventy percent growth in a year and a half. It accounted for 30 percent of TSMC's wafer revenue in the second quarter. Two-nanometer is the one that decides the next few years. It is the first node to use gate-all-around transistors after years of FinFET, which TSMC says buys 10 to 15 percent more performance at the same power or 25 to 30 percent less power at the same performance, and it is the manufacturing platform for the next generation of Apple and Nvidia silicon, data-center CPUs and a large volume of AI ASICs, so demand is running ahead of earlier node launches.
The capex number behind the plan is $60 billion to $64 billion for 2026, with 70 to 80 percent going to advanced process. TSMC is building across Taiwan, Arizona and Japan, and converting some 5nm lines to 3nm to reuse mature fabs faster. Advanced packaging is expanding on the same logic: CoWoS capacity is reported to go from about 130,000 wafers a month at the end of 2026 to about 260,000 by the end of 2028, roughly double. Beyond 2nm, TSMC is working on A16, A14, A13 and A12, and it has a joint program with ASML to move High-NA EUV from the decades-old 6-inch reticle to a 12-inch platform. One caveat: this is a supply-chain report, and on September 13 TSMC said it does not comment on market rumors and that all capacity information comes from official announcements.
— TSMC (earnings call) · 台湾经济日报 · TrendForce
🔗 Global Memory Supply: TSMC to Increase 2nm Capacity by 22%, 3nm by 16% · 腾讯新闻: 台积电最新扩产计划：2/3nm全面提速
Grok moves into Microsoft 365 Copilot, and Copilot becomes a model menu
Microsoft added xAI's Grok models to Microsoft 365 Copilot on September 12, starting in Word, Excel and PowerPoint through the Frontier early-access program. Satya Nadella announced it and Elon Musk confirmed it the same day. It is admin-gated and off by default: an administrator has to enable SpaceXAI models in Copilot settings, and it is not available to Frontier customers in the EU, EFTA or the UK during the preview, a carve-out Microsoft has not explained. xAI has been added to Microsoft's Online Services Subprocessor List, and admins keep control over what data Grok processes. Microsoft has not named which Grok version powers the Office preview.
"Grok in Copilot" is really three rollouts on three surfaces. Copilot Studio has run Grok 4.1 Fast since February 2026, GitHub Copilot added Grok 4.6 on August 14, and the Office preview is the newest and most cautious of the three. That makes Copilot a model marketplace rather than a product with one model inside it: with GPT-6 Astra, Claude Fable 5.1 and now Grok all selectable, an enterprise can route different tasks to different vendors under one license. The strategic read is that no model vendor owns the interface most office workers actually use, which matters more to OpenAI than to anyone else on that list.
The same week, federated Copilot connectors reached general availability, a quieter but more consequential change for enterprise data. The connectors use the Model Context Protocol, do not index or store anything, and retrieve third-party data in real time under the user's own identity, with OAuth 2.0 respecting source permissions and read-only access, so agents can search and fetch but cannot write back. At general availability they cover the Researcher agent, Microsoft 365 Chat and Agent Mode in Excel, with first-party connectors spanning legal, financial, healthcare and professional services systems. The practical note for administrators is that a model selector in a chat box looks like changing a font but is a change of processor, and the data-handling terms differ by vendor.
— Microsoft (official) · Big Hat Group · eesel AI
🔗 AIToolsRecap: Grok Is Now Inside Word and Excel — Unless You Are in Europe · Big Hat Group: Copilot Weekly — Grok Joins M365, MCP Connectors Hit GA
A robot-joint maker raises 300 million yuan as embodied AI money concentrates
Nanjing-based Encos, which builds integrated joint modules for humanoid and embodied robots, closed a B round of more than 300 million yuan on September 14. The investors include CITIC Jinshi, Nice Group, Suzhou Venture Capital, Huarui Investment, Nanjing Jiaokong and Huarui Chuangtou, with Fosun Chuangfu, Shenzhen Capital Group, Huakong Fund, Jinqiu Fund and Puhua Capital following on from earlier rounds. Minglun Capital is the long-term exclusive financial adviser. Founded in 2022, the company has spent four years moving from a parts supplier to what it now calls a hardware infrastructure platform for embodied AI.
The product bet is about what happens after robots ship. In June 2026 Encos released what it says is the industry's first quick-release second-generation humanoid joint module, which cuts single-joint assembly and disassembly from hours of specialist work to minutes. At prototype scale the maintenance problem is tolerable; at fleet scale it is a cost multiplier, and Encos argues quick-release design lowers a robot maker's after-sales cost by an order of magnitude. It is also preparing flat-wire motor products, which raise slot fill above 90 percent and, by the company's account, lift some core joint performance by more than 20 percent in the same form factor. Alongside the joints it sells an EC-DexHand-5F dexterous hand with 20 active degrees of freedom and an EC-Gloves data-collection glove that reads all 20 joint angles at a 1 kHz communication rate, which together close the loop from data collection to model training to real-robot deployment.
The customer list is the evidence that matters. Encos supplies Chinese humanoid makers including Booster Robotics, Songyan Power, Galbot and Zhongke Huiling; overseas embodied-model companies including Physical Intelligence and Amazon and its subsidiaries; and cross-industry buyers such as BYD, Zoomlion and GAC. Its arms, built with ARX Robotics, appear in Physical Intelligence's published demos with Encos joints throughout, which the company reads as proof of performance, stability and a small sim-to-real gap. Encos says it makes its own drivers, reducers, motors and encoders, runs its own precision machining from gears to housings, and holds a first-pass yield above 97 percent. The round lands in a market that is consolidating rather than spraying: Chinese embodied AI drew about 93.5 billion yuan in the first half of 2026, roughly five times a year earlier, but capital is now concentrating in teams with a foundation model and a clear route to a product. Encos is selling the joint to all of them.
— Encos (announcement) · 科创板日报 · 盖世汽车
🔗 网易: 因克斯完成超3亿元B轮融资，年内将上线线上选型平台 · 腾讯新闻: 具身智能公司因克斯完成超3亿元B轮融资
AI Daily Digest is published every morning by KD Agentic. Sources are linked inline; aggregator coverage is used for discovery only and is not cited as a primary reference.
Top comments (0)