Shanghai AI Laboratory released Atria Dawn Preview — 744B-parameter agentic MoE, top scores on five of sixteen benchmarks
Shanghai AI Laboratory released Atria Dawn Preview, a 744B-parameter agentic MoE built on GLM-5.2 and trained via a Verifiable Experience Pipeline that grounds tool use in executable environments, reporting competitive results across 16 benchmarks with top scores on five. Atria Dawn is the largest Chinese open-weight model release since Kimi K3’s 2.8T parameters in July — and it arrives the same week Xi announced the BRICS AI Open Source Zone.
Meta settled child harm claims for $18 billion — larger than last week’s $16.68B figure, with platform change commitments
Meta settles landmark state child harm claims for $18 billion and promises changes to its platforms. The settlement grew from the $16.68B figure reported last week — the final number is $18B, making it definitively the largest social media harm settlement in history. The “promises changes” language is the detail that matters for product teams: Meta committed to specific platform design changes as part of the settlement, not just financial payment.
Microsoft released emergency Windows patches closing two critical CVEs — and Anthropic’s compute commitments hit $517B, triple the earlier estimate
Microsoft on September 14 released emergency out-of-band updates to fix Remote Desktop Services instability, Hyper-V failures, and multichannel USB Audio breakage introduced by September 8 Patch Tuesday. The OOB also closes CVE-2026-62721 and CVE-2026-85921. Patch immediately if you run Windows infrastructure. Separately, Anthropic’s compute commitments hit $517B, covering 14.8GW of capacity in the 11 months through August 2026 — nearly triple the ~$180B in server-leasing spend it previously told investors.
Map the geopolitical risk in your AI vendor stack — before it maps you
Prompt: Trump attacked Amodei personally and warned AI companies he holds "criminal and regulatory power" over them. China called pacing a Cold War playbook and Xi launched a BRICS AI Open Source Zone. China's exit-ban decree targeting AI and chip researchers took effect today. The NSA/CISA/FBI advisory named six Chinese AI labs for extracting billions of tokens from US frontier models. Shanghai AI Lab released a 744B-parameter model the same day. The AI vendor decision is now also a geopolitical decision. Our current AI stack: [list every model and vendor — OpenAI, Anthropic, Google, Meta, DeepSeek, Alibaba Qwen, Mistral, or others — and what we use each for]. Help me map the geopolitical risk across three dimensions: 1. The administration risk — Trump's Truth Social post signals that AI firms whose public positions conflict with the administration's are political targets. For each US-based vendor we use: what is their public position on pacing, safety regulation, and China competition? Anthropic drew direct fire this week. OpenAI and Altman backed pacing but have not yet drawn equivalent attack. For government-adjacent workloads: does our vendor's political exposure create procurement risk we need to plan for? 2. The Chinese vendor exposure — the NSA/CISA/FBI advisory named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for unauthorised model distillation. China's exit-ban decree now locks AI researchers inside the country. The BRICS Open Source Zone is an explicit alternative ecosystem. For any Chinese-origin model we use: what data flows through it, what is the regulatory exposure if US export controls tighten further, and what is our fallback if access is restricted as Fable 5 was in June? 3. The BRICS zone opportunity — Xi's BRICS AI Open Source Zone is not just a political statement. It is a distribution network for open-weight models across countries representing 40%+ of global GDP. If your product serves markets in BRICS countries: what does a world look like where the AI models your users access are increasingly from the BRICS ecosystem rather than US frontier labs? Is that a risk to plan for, an opportunity to build toward, or both?
OpenAI Pays Contractors to Read Real ChatGPT Conversations
According to an investigation by 404 Media, OpenAI is using human contractors to review real ChatGPT conversations as part of its model training and improvement process.
- The internal initiative, reportedly called “Project Lily,” involves contractors reading, summarizing, and evaluating ChatGPT conversations.
- Contractors are reportedly paid through third-party staffing firms and assess multiple AI responses with the goal of improving the responses ChatGPT gives its users.
- Reviewers are instructed to flag overly flattering responses (sycophancy), unnecessary AI-style language (“AI-speak” and “emoji misuse”), and claims that suggest the model has human experiences (such as “As a chef, I like to…” or “I know what that’s like.”)
- OpenAI says usernames are not provided to reviewers, and conversations pass through an automated privacy filter designed to remove identifying information.
- However, OpenAI acknowledges that its privacy filter can make mistakes, including missing uncommon identifiers or ambiguous personal details.
- 404 Media reports that reviewers can also see a “user memories summary” in some cases, potentially providing additional context about a user’s previous interactions, which can inadvertently reveal their general location, profession, or personal life context.
OpenAI isn’t the only AI lab doing this. Anthropic confirmed that it uses human reviewers when users have turned on the “Help improve Claude” setting in their privacy settings, while Google’s Gemini clearly posts a disclaimer that says, “Humans review some saved chats to improve Google AI.”
Chinese Researchers Map Out 5-Stage Path to ‘Last AI Built by Humans’
Researchers from ByteDance, Tsinghua University, and the Shanghai Artificial Intelligence Laboratory have proposed a five-stage roadmap for AI systems that could improve themselves with increasingly less human intervention, and eventually, almost none at all.
- The study, titled “The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement,” focuses on recursive self-improvement (RSI), where improvements persist across tasks and can be inherited by successor AI systems.
- The first stage involves AI carrying out improvement procedures designed by humans, while later stages give models more control over how they improve themselves.
- More advanced stages would allow AI to determine what information or experiences it needs and adapt its learning after deployment.
- The final stage envisions systems that can continuously improve the very methods used to improve AI itself.
- The researchers said progression through the five stages would vary by field, with software engineering presenting a relatively clear path forward and robotics and scientific discovery facing steeper hurdles.
- They argue that automating AI research could reduce development time, labor requirements and computing costs for foundation models.
Safety remains a major requirement. The researchers said that genuine RSI required strict safeguards, including verified testing environments to ensure updates were safe and beneficial before being deployed.
How to Stop OpenAI from Using Your Chats
According to 404 Media, OpenAI told the outlet that users’ chats won’t be used to improve the company’s models if they turn off the “Improve the model for everyone” setting.
This is turned on by default for Free, Plus, and Pro plans, so users need to proactively turn it off if they wish to opt out. Enterprise, Business, and Edu customers have the model-improvement setting off by default.
To keep your future conversations from reaching human reviewers:
Step 1: Go to ChatGPT and sign in.
Step 2: Click your profile icon in the bottom corner and select Settings.
Step 3: Select Data Controls.
Step 4: Toggle the “Improve the model for everyone” switch to Off.
Note: The opt-out isn’t retroactive. It only applies to new conversations going forward; prompts that have already been de-identified and queued in the system stay part of the training cycle. Also, turning this off doesn’t stop automated scans for Terms of Service or safety violations.
Google introduces new audio models for more natural conversations
The company has introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two new audio models designed to more effectively handle interruptions, understand visual context, and support conversations across 97 languages.
Gemini 3.8 Live focuses on speed and efficiency, while the Extended Thinking version is designed for more complex, multi step tasks.
Meta introduces Meta One subscription across its apps
Meta has introduced Meta One, a subscription service offering expanded access to AI tools and premium features across Facebook, Instagram, and WhatsApp.
The company also announced it will now allow AI agents to set up and manage WhatsApp Business messaging.
AI agents get new hotlines for reporting misconduct
Two new AI hotlines have launched to give AI agents a way to report suspicious behavior by other agents to humans.
Zuckerberg pushes back on calls for an AI slowdown
Meta CEO Mark Zuckerberg has distanced the company from calls by leaders at Anthropic, OpenAI, Google, and SpaceXAI for a coordinated slowdown in AI development. He said each AI lab should individually determine the pace needed to train its models safely.
Apple shipped the new Siri. Google built a chunk of the brains.
Apple released Siri AI on Monday, in beta, English only, with French, Japanese, Korean, Portuguese and Spanish due next month. It does personal context across your messages, mail and photos, on-screen awareness, and in-app actions, plus Visual Intelligence, Writing Tools, Image Playground, Siri Recap and Live Rewind. Server-side operations carry daily usage limits while it is in beta.
The features run on “the next generation of Apple Foundation Models, custom-built in collaboration with Google and its Gemini models.” Apple spent two years insisting the personal-context Siri was coming; it arrived leaning on a competitor’s model. Supported hardware is iPhone 16 or later plus iPhone 15 Pro and Pro Max, M1-or-later iPads and Macs, MacBook Neo, Vision Pro, and Apple Watch Series 9 or later. It will not ship initially in the EU on iOS, iPadOS and watchOS, and will not ship in China while Apple works through regulatory requirements.
The same week Siri landed, X Corp and xAI dropped Apple from the antitrust suit they filed in August 2025 over ChatGPT’s exclusive iPhone integration. They are still suing OpenAI. No reason given, no settlement disclosed. Make of that what you will now that Siri’s default brain is not exclusively OpenAI’s.
Jensen Huang breaks the “slowdown” consensus
At Salesforce’s Dreamforce on Tuesday, Jensen Huang said the quiet part into a microphone: “Safety is an engineering problem, not a legal one. We don’t need any new laws. We don’t need new regulations.” His proposed mechanism is product liability plus self-restraint: “if you build a product or a service, and you’re not confident in its safety, then don’t release it.” He called the tradeoff between speed and safety “a false choice.”
That is three days after Amodei’s essay and two days after Altman, Musk and Nadella all signed on. It is also the opposite direction from where OpenAI spent Tuesday: policy chief Chris Lehane told reporters in Washington that OpenAI, Anthropic and Google DeepMind have been coordinating on safety for several weeks, that they do not need a government antitrust waiver to do it, and that OpenAI backs the FRONTIER Act provision requiring frontier labs to admit independent verification organizations. Amodei asked Washington for that waiver by name. OpenAI says it is unnecessary.
Huang was on that stage to sell something, for the record. Salesforce used the slot to launch Koa, its first CRM reasoning model, post-trained from NVIDIA’s open Nemotron 3 Super on synthetic scenarios across 14-plus industries with no customer data. Salesforce claims Koa “matches or exceeds leading model performance on CRM actions with three times fewer errors” on its own internal benchmark, keeps the weights, and runs inference inside its own trust boundary. Pilots are running now with Formula 1, UChicago Medicine, Xero and 1-800Accountant, and general availability is slated for winter 2026 in US regions.
China’s answer to pacing week: $5B and a 753B-parameter MIT license
Z.ai, the Hong Kong-listed lab behind the GLM family, closed roughly $5 billion on Monday: about $2B in a share placement and about $3B in convertible bonds. The stated use of proceeds is next-generation GLM foundation models, a fully self-training system, training and inference compute, and domestic-chip adaptation. SCMP reports the placement priced at HK$714 a share, a 9.96% discount, against HK$1,588 in July and a June peak of HK$2,980. The raise is enormous and the stock is down by three quarters. Both things are true.
A day before that, Shanghai AI Laboratory put Atria Dawn Preview on Hugging Face: 753B parameters, MIT license, 256K context, built on Z.ai’s 744B GLM-5.2 MoE foundation. Posted scores include 96.0 on DeepSearchQA, 86.5 on CyberGym, 86.2 on MLE-bench Lite, 77.0 on BFCL v4 and 65.0 on Workspace-Bench. The accompanying paper claims top scores on five of sixteen benchmarks and reports that human participants rated about a third of completed AI-assisted tasks as infeasible without AI. Treat self-reported numbers as self-reported, but note the license: an agentic model at this scale, free to take.
The rest of the week’s response came in four registers. China Daily called the slowdown push “a self-serving bid to preserve American dominance.” The Foreign Ministry called it fearmongering. A DeepSeek engineer who worked on V4.1 posted that he especially does not want Anthropic holding the most advanced AI. And at the Xiangshan Forum, defense minister Dong Jun urged nations to “consult with one another on governance.” Nobody in Beijing is pacing anything.
Zuckerberg rejects coordinated AI slowdowns
Meta CEO Mark Zuckerberg posted on X Tuesday that AI labs don’t need a coordinated slowdown to handle safety, arguing each company can set the pace it needs to train models safely.
The post lands as Anthropic CEO Dario Amodei and others call for slower development. Zuckerberg argued liability risks already push labs to act responsibly, and alignment will separate winners from losers.
His proof: Meta delayed its Muse AI agent for months without asking rivals to follow. But employees testing Muse during its launch week reportedly saw it expose private iCloud photos, per Implicator.
Contractors are reading ChatGPT conversations
404 Media reports OpenAI is paying hundreds of contractors through Project Lily to review real ChatGPT conversations and rate responses, including chats that can contain sensitive personal information.
Reviewers don’t see usernames, but may see memory summaries, while OpenAI says it filters personal identifiers before review and acknowledges that some sensitive details can still slip through.
That makes the privacy setting more consequential than it looks: OpenAI says consumer chats may help train models unless users opt out, while Business, Enterprise, and Edu accounts default to no training.
Meta plans camera-free smart glasses
Meta is reportedly preparing camera-free smart glasses called Luna, which could debut at Connect on September 23 and ship in October, with Clubmaster and Burbank styles planned.
Luna will reportedly use six microphones, directional speakers, and a dedicated button for Meta AI, while slimmer temples could make the glasses look more like ordinary eyewear than Meta’s current models.
Meta may be betting that removing the camera solves a branding problem, after repeated privacy concerns and recording loopholes made its existing smart glasses controversial despite stronger anti-recording safeguards.
Apple plans AI servers for sale in 2029
Apple is developing AI servers powered by its future M8 Ultra chips and wants to sell them to AI developers, businesses and governments, according to The Information. Sales aren’t expected before 2029.
The report describes two versions, with either two or four M8 Ultra chips, and says Apple has discussed linking them with Nvidia’s NVLink Fusion. CEO John Ternus has backed the roughly year-old project.
Apple hasn’t sold a dedicated server line since it discontinued Xserve in 2011. Its comeback may now lean on Nvidia, a chipmaker Apple has kept at arm’s length for years.
SpaceX targets first orbital Starship flight
SpaceX plans to send Starship into orbit for the first time on Flight 14, targeting September 22 after 13 earlier tests followed shorter suborbital paths and returned within about an hour.
If successful, Starship will orbit Earth roughly six times over ten hours at around 275 kilometers altitude, then splash down west of Chile after attempting to deploy operational Starlink V3 satellites.
SpaceX will skip a Starship tower catch this time, but a successful satellite deployment could matter more, since each V3 unit promises roughly one terabit per second of downlink capacity.