Google DeepMind posted the launch of Gemini 4 Hydrogen, saying it is built for rapid coding, multimodal synthesis, and everyday workflows, finishes tasks 2.5x faster than 3.8 Flash, and surpasses Argon on the Artificial Analysis Intelligence Index. Sundar Pichai, Logan Kilpatrick, the main Google account, the Gemini app, and Google AI Studio also posted about the rollout, and Cursor said the model is available in Cursor.

Google DeepMind announced Gemini 4 Hydrogen on November 12, calling it a model built for rapid coding, multimodal synthesis, and everyday workflows. The company said it finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the Artificial Analysis Intelligence Index.

Logan Kilpatrick said the model is available to all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity. Sundar Pichai said it follows Carbon and Argon, is built for speed and efficiency for everyday tasks, coding and agentic workflows, and has 4 thinking levels, so users only wait and pay for the reasoning depth a task requires. The main Google account said the model can stand on its own or team up with Carbon and Argon, and later posted that Search is getting the model as well.

A chart attached to the DeepMind post compares Gemini 4 Hydrogen with Claude Sonnet 5.5 and GPT 6.1 Terra. In that chart, Hydrogen leads on several tests, including SWE sweep at 13.5%, Agents' Last Exam v1.1 at 71.6%, and HLE Diamond at 58.8%, while Claude Sonnet 5.5 leads on Terminal Bench 4.1, OSWorld 2.1, and Terminal Bench Science 0.1. Cursor said the model is now in Cursor and, at its default effort, scores 54.3% on CursorBench, up from 39.6% for Gemini 3.8 Flash. Demis Hassabis said he was very excited to release the model and pointed readers to a blog for more information.

Watch how quickly the model shows up for users in the Gemini app, Search, AI Studio and the API, and how developers compare it with Carbon and Argon. The 4 thinking levels will be an early test of whether people can trade speed and cost against reasoning depth in real work. Independent benches, including Artificial Analysis and CursorBench, will also show whether the official speed and quality claims hold once more people are using it.

Anthropic filed a registration statement with the SEC and started its IPO roadshow on November 10, according to its IPO site. Shares are expected to be priced on November 18 and to begin trading on Nasdaq under the ticker ANT on November 19. Goldman Sachs and Morgan Stanley are named as the contacts for the prospectus.

Anthropic posted a link to its initial public offering site on the morning of November 12, with a legal disclaimer saying it has filed a registration statement, including a prospectus, with the SEC.

The site lists Q3 2026 financial highlights. Anthropic reports revenue of $17.0B for the quarter, more than 12 times the level a year earlier, and more than 11,900 customers paying at least $100K in trailing twelve month revenue as of September 30. It reports a 15% adjusted operating margin and a 3% negative operating margin, plus a net loss of $23.3B that it says includes a $21.9B noncash charge tied to a change in the fair value of convertible notes, derivative liabilities and SAFEs, caused by the rise in its valuation. It also claims a compute efficiency gain of more than 4 times since Q1 2025.

According to the site's FAQ, the roadshow began on November 10, the final share price is expected on November 18, and Class A shares are expected to begin trading on Nasdaq under the ticker ANT on November 19, subject to regulatory clearance. Retail investors can request shares through participating brokerages. The site also hosts a roadshow video, a founders' letter and customer stories from Legora, Eli Lilly, Rakuten and Salesforce.

Additional reporting, secondhand: an X news summary says the shares are marketed at $70 to $78 each and that the IPO plus a private placement could raise up to $90.4 billion at about a $1.4 trillion valuation. These figures come from summaries of reports, not from the Anthropic post.

The next milestones are the pricing on November 18 and the first trading day on November 19. Watch whether the final price lands inside, above or below the reported range, how large retail allocations turn out to be, and how the first day of trading goes. The full prospectus on EDGAR will be read closely for risk factors, compute commitments and governance terms, and the result will shape how investors value OpenAI and other labs that may follow.

Claude posted that Claude Meetings lets teams add Claude to a video call, where it builds docs, whiteboards, and other artifacts during the conversation. The post said the feature is starting to roll out in beta on Team and Enterprise plans. Follow up posts said Claude joins as a visible participant, takes notes, and is starting with Google Meet.

Claude announced Claude Meetings on November 12, saying teams can add Claude to a video call and it builds docs, whiteboards, and other artifacts while people are still talking. The company said the feature is starting to roll out in beta on Team and Enterprise plans.

In a follow up post, Claude said it joins as a visible participant. Teams can ask it a question out loud or in chat, or have it start a doc, and can see what it is building and correct it as it takes shape. Another post said Claude takes the notes too, and that after the call the notes, action items, transcript, and anything Claude built are on the meeting's page in Claude. A fourth post said Claude Meetings is starting with Google Meet, with support for other video conferencing platforms coming soon.

Anthropic's article on the launch matches those points. It says Claude can use connected tools with permission, writes notes and action items so no one on the call has to, and by default stays out of calls when someone from outside the organization is invited. The article also says support is currently available for Google Meet, notes and transcripts are free on Team and Enterprise plans during the beta, and docs, whiteboards and other artifacts made in a call are free through February 10, 2027. It is not available during the beta to some organizations, including those with zero data retention, HIPAA ready configuration, or customer managed encryption keys.

Watch the beta rollout on Team and Enterprise plans, and whether support expands beyond Google Meet as Claude said it would. A practical test is whether teams actually let Claude build and edit docs live, and how often they correct it on the call. Anthropic said it is also working with meeting tools so transcripts can go to software teams already use. That would be an early sign the feature is meant to fit into existing workflows, not replace them outright.

Will DePue posted on November 12 that he has conclusive forensic evidence that DeepSeek, Moonshot, MiniMax and Ant Group built their newest models on Nvidia's most advanced Blackwell chips, which he says are barred from export to China. He says the finding comes from a forensic analysis of published model weights, not from insider information, and he links to a full report. This is an unverified claim. In a quote post, roon wrote that all major Chinese models are being trained on Blackwell class chips.

Will DePue posted on November 12 that he has found conclusive forensic evidence that DeepSeek, Moonshot, MiniMax and Ant Group built their newest models on Nvidia's most advanced Blackwell chips, which he says are barred from export to China. He says the finding comes from a hidden GPU signature in published starting weights, not from insider information.

DePue says he searched more than six trillion chip and seed combinations and found conclusive evidence that six Chinese AI models, including DeepSeek V4 Pro, Kimi K3, and MiniMax M3, were initialized on Nvidia Blackwell systems, four of them matching Grace Blackwell GB200 and GB300 parts. A table on the post, titled Table 1, lists those labs' models and the GPU he says each was created on, with six newer models marked in red as Nvidia Blackwell. The table notes that red requires a U.S. license for export to China and that none has been reported.

He says a Blackwell B300 reproduced DeepSeek V4 Flash's starting weights bit for bit, and a rented GB200 reproduced Kimi K3's to a correlation of 0.999995, while a Hopper H100 reproduced none of them and changing the seed by one broke the match. He also says the method matched 24 known open weights models to the hardware developers said they used, with real matches from 49 to 1,385 sigma.

DePue estimates Kimi K3 and DeepSeek V4 Pro used 2,000 to 3,000 Blackwell GPUs each, costing on the order of $150 to $200 million per cluster, or more than half a billion dollars of export controlled GPUs across these labs. He says he is highly confident and that senior researchers at OpenAI and Anthropic have independently verified the result. He points readers to a full report. In a quote post, roon wrote, "really bad. all major Chinese models are being trained on Blackwell class chips."

No first hand confirmation from Nvidia, DeepSeek, Moonshot, MiniMax, Ant Group, OpenAI, or Anthropic was found for this post. The findings are DePue's own claim and should be treated as unverified.

Watch for first hand confirmation or rebuttal from Nvidia, DeepSeek, Moonshot, MiniMax, or Ant Group, and for OpenAI or Anthropic to say whether their researchers verified the result on the record. Independent researchers can also try to reproduce the weight signature tests DePue describes. Policymakers may ask how export controlled Blackwell systems could have been used at this scale if the claim holds. Until a lab, a chipmaker, or a government confirms the finding, treat it as an unverified claim.

Congress is weighing three proposals that would restrict future sales of high end chips, reduce the diversion of Nvidia processors and block China from advanced chipmaking machinery made by ASML. Bloomberg reported on November 12 that OpenAI confirmed to lawmakers it supports the measures, and Politico reported that OpenAI told congressional offices it backs key chip export control bills. Andrew Curran noted that OpenAI is joining Anthropic in backing the push.

A post from the account NIK (ns123abc), which drew more than 400,000 views, shared screenshots of a Bloomberg article titled "OpenAI Joins Groundswell of Support for AI Chip Control Measures," dated November 12. The post framed the timing as a "targeted takedown" of Jensen Huang, who resigned as Nvidia CEO the same day. That framing is the poster's opinion and is not supported by the reporting it shares.

According to the Bloomberg article shown in the screenshots, OpenAI is backing a bipartisan effort in Congress to block China's access to AI chips and curb its access to chipmaking equipment. The article says OpenAI had weighed for months whether to support a package of export related bills and confirmed to lawmakers this week that it would, according to people familiar with the matter. Congress is weighing three proposals: limits on future sales of high end chips, measures to reduce the diversion of Nvidia processors, and a block on Chinese access to advanced machinery from ASML. The measures would be attached to annual defense legislation that must pass before the end of the year.

Bloomberg also reported that Nvidia's allies in Washington, including trade groups, mobilized against the provisions this week, while Nvidia itself stayed relatively quiet. Politico separately posted that OpenAI told congressional offices it supports key chip export control bills. Andrew Curran wrote that OpenAI is joining Anthropic in backing the push, with the aim of passing legislation before the end of the year.

The key test is whether the provisions survive negotiations on the annual defense bill, which must pass before the end of the year. Watch for formal statements from OpenAI and Anthropic, for Nvidia's response under interim CEO Colette Kress, and for lobbying by trade groups that oppose the measures. Any final text will show how far the limits on chip sales, chip diversion and ASML equipment go, and whether the White House supports them.

Nvidia has used its balance sheet to invest in AI startups and to back credit for cloud providers that buy its chips. Jensen Huang resigned as CEO on November 12 and CFO Colette Kress became interim CEO. The Information reported that the new leadership is expected to slow that strategy.

Liquid, a markets account on X, posted on November 12 that Nvidia is expected to slow its aggressive AI investment strategy following Jensen Huang's resignation, citing The Information. The post said the change puts billions of dollars in startup funding and credit guarantees at risk. The post did not name which investments or guarantees could be affected.

The report lands on the same day Nvidia announced that Huang resigned as president, CEO and board member, effective immediately, and that CFO Colette Kress is interim CEO while the board searches for a permanent successor.

A separate post from Anissa Gardizy, posted with a WSJ byline credit, described a tense late summer meeting in which Huang called Kress a "deer in the headlights." According to that post, Huang's concern was that Nvidia's plans to fund the AI boom with its balance sheet were running into problems, and he called her a "reporting CFO" rather than someone who could help define strategy.

Together the two reports put Nvidia's role as a financier of the AI buildout at the center of the leadership change. Nvidia has not publicly confirmed any change to its investment plans.

Nvidia reports third quarter fiscal 2027 results on November 17, the first update under interim leadership, and investors will look for any comment on investments and credit support. Watch AI labs and neoclouds that rely on Nvidia backing for signs of delayed rounds or new financing partners, and whether the board's choice of a permanent CEO signals a return to the old strategy or a more cautious balance sheet.

Polymarket posted on the evening of November 12 that US nonprofits had sued OpenAI, Google, Meta and Anthropic, asking for an immediate freeze on advanced AI development over fears of catastrophic harm. Reuters, in a report by Diana Novak Jones, said the suit was filed in San Francisco and also names SpaceX's AI division, and that the plaintiffs include the Irreplaceable Education Fund, Public Citizen and the Heat Initiative. The lead post is secondhand; Reuters is the first hand news source used here.

Polymarket posted late on November 12 that US nonprofits had sued OpenAI, Google, Meta and Anthropic, demanding an immediate freeze on advanced AI development over fears of catastrophic harm. That post does not name the plaintiffs, the court or any other defendant.

Reuters, in a November 12 report by Diana Novak Jones, said a group of nonprofits sued OpenAI, Anthropic, SpaceX's AI division, Meta Platforms and Alphabet's Google in San Francisco, seeking to block further development of the most advanced AI models. According to Reuters, the plaintiffs include the Irreplaceable Education Fund, Public Citizen and the Heat Initiative, and they argue that California's laws against unfair business practices provide grounds for a court order barring further development of products such as ChatGPT, Gemini and Claude. Reuters also reported that a Meta spokesperson declined to comment and that representatives of the other companies did not immediately respond.

A press release from the Alliance for Secure AI, dated November 12, says that group joined the same plaintiffs in a suit filed in the Superior Court of California, County of San Francisco, and also lists SpaceXSI among the defendants. The groups say they want a freeze on development and deployment of AI systems more capable than those available today. Those are the plaintiffs' claims, not findings by a court.

Watch the San Francisco court docket for a hearing date, for any request for a temporary order, and for formal responses from the companies named in the suit. A court decision, or a decision to dismiss or narrow the case, would matter much more than the filing itself. Independent coverage should be checked against the complaint for the exact relief sought, including whether SpaceX's AI division is a defendant as Reuters and the plaintiffs say.

SpaceX posted that it builds the infrastructure of the future and is committed to making sure the communities where it builds share in the benefits, and linked to an official update. That update says SpaceX is advancing a supercompute facility at Starbase in the Rio Grande Valley in Texas and a second campus in Cochise County, Arizona, together representing more than $74 billion in planned American investment. Follow up posts named those locations and a $220 million water commitment in Brownsville.

SpaceX posted on November 12 that it builds the infrastructure of the future and is committed to making sure the communities where it builds share in the benefits, and linked to an official update. A follow up post said that from the Rio Grande Valley in Texas to Cochise County, Arizona, the company is building more than supercomputers, including jobs, infrastructure, and local tax bases. Another post called the Texas site the Starbase supercompute site and said development there is also a chance to clean up decades of unauthorized dumping. A later post said SpaceX has committed $220 million to expand water and wastewater infrastructure for Brownsville and Cameron County, which it called the largest private contribution to Brownsville's water future in the city's history.

The linked SpaceX update is more specific. It says the company is advancing two projects, a major new supercompute facility at Starbase and a second campus in Cochise County, and that together they represent more than $74 billion in planned American investment. According to that page, SpaceX is planning to invest more than $24 billion in new supercomputing infrastructure in the Rio Grande Valley, on top of more than $15 billion already invested in Starship and related infrastructure at Starbase. It says the Texas facility will generate its own electricity without drawing power from the public grid, has already generated more than 3,000 construction jobs, and is expected to support more than 500 permanent jobs once operational. For Cochise County, the page says SpaceX is pursuing a roughly $50 billion supercompute campus expected to create about 5,500 construction jobs and 1,000 permanent jobs.

Watch for local permits, power and water agreements, and a clearer construction timetable in Texas and Arizona. The official update ties the Texas site to training frontier systems and to work on space based computing, so later technical detail would show what these campuses are actually for. The $220 million Brownsville water commitment is an early, concrete item to track, because SpaceX presented it as part of how nearby communities share in the project.

TSMC is the world's largest contract chipmaker and produces the advanced chips used by Nvidia, AMD and other AI hardware companies. Andrew Curran posted an image of a TSMC filing dated November 12 that contains a statement on unauthorized access to employee personal data.

Andrew Curran posted "TSMC has been hacked" on the evening of November 12, attaching an image of a TSMC filing dated November 12, 2026 and signed by Wendell Huang, senior vice president and chief financial officer.

The statement in the image, titled "Statement Regarding the Unauthorized Access of Employee Personal Data," says TSMC found that some current and former employee data was made available on an external website without authorization. TSMC says it has found no evidence that its own systems were compromised. It says the information matched data stored in the system services of a cloud service provider and is unrelated to the company's operations or to customer data.

According to the statement, TSMC is informing affected personnel and providing support, and has engaged the cloud service vendor, the relevant investigating authorities and a third party cybersecurity agency to establish the facts. The company says it remains committed to protecting employee privacy.

The filing does not name the cloud provider, the website where the data appeared, or how many people are affected. Curran's post drew more than 500,000 views, and other accounts noted that the leak covered employee information only.

Watch for TSMC to identify the cloud vendor and the number of employees affected, and for findings from the investigating authorities. Customers will want confirmation that no design or manufacturing data was involved. Any sign that the leak reaches beyond personal data would matter for the whole AI chip supply chain, since TSMC builds the most advanced accelerators. Regulators in Taiwan may also ask whether the cloud vendor met its data protection duties.

Google has been building Android XR, its platform for headsets and glasses, with Samsung as a hardware partner. The new glasses put the Gemini assistant into frames designed by Warby Parker and Gentle Monster. Sundar Pichai quoted the launch post and highlighted live speech translation.

Shahram Izadi, who leads Android XR at Google, posted on November 12 that preorders for the new intelligent eyewear are open. He said the glasses were built with teams at Google, Samsung, Gentle Monster and Warby Parker.

According to the post, the lightweight glasses give hands free help from the Gemini app through the day. Izadi described them as eyewear first, with Warby Parker and Gentle Monster providing a collection of designs, colorways and sizes. The post says people can try them in select Warby Parker, Gentle Monster and Google stores starting November 12, and that shipping begins December 10. It also links to pages on the design collection and on privacy protections for the wearer and people nearby.

Later that evening Sundar Pichai quoted the post and wrote that live speech translation on the glasses is "🤯".

The post did not list prices or the countries where preorders are open. The launch puts Google and Samsung into direct competition with other camera equipped AI glasses already on sale.

Shipping starts December 10, so the first hands on reviews should arrive within weeks. Watch for pricing details, which markets get the glasses, and how well Gemini features like translation work outside a demo. Privacy rules for camera equipped glasses in stores, offices and schools are another open question as sales begin. Rival glasses makers are likely to answer with their own features and prices before the holidays.

Unusual Whales relayed a report from Korea's Financial News about a closed door session Sam Altman held with Samsung leaders. The comments are secondhand: neither OpenAI nor Samsung has confirmed them publicly. They land in a week when chip stocks have been swinging on worries about AI spending.

Unusual Whales posted on November 12 that Sam Altman told Samsung executives that AI demand has been double counted, according to Korea's Financial News.

According to the post, the remarks came in a closed door session with about 50 Samsung presidents on October 31. Altman reportedly said only OpenAI and Anthropic are driving major demand growth. He warned of supply shocks and turbulence ahead, but said he expects AI demand to surge in the long run.

This is a secondhand account. The quotes come from a Korean newspaper report relayed by a market news account, not from OpenAI, Altman or Samsung, and the full context of the session is not public. The idea of double counting usually refers to the same underlying demand showing up in orders from several companies in a supply chain, but the post does not say exactly what Altman meant.

The report spread quickly, passing 200,000 views within a few hours, because it touches a live debate about whether AI capital spending is running ahead of real demand. The same week, Nvidia shares were hit by a broker downgrade and turmoil around its chief executive, and rumors about cut chip orders were circulating.

Watch for OpenAI or Samsung to confirm, deny or add context to the remarks, and for the original Financial News article to be quoted more fully. Memory makers and chip suppliers may be asked about double counting on upcoming earnings calls, including Nvidia's report on November 17. Any change in Samsung's HBM or foundry plans would show whether the comments affected real orders.

Thinking Machines posted that Inkling 2 Small's weights are now available for download on HuggingFace under an Apache 2.0 license. The post linked to the model page on HuggingFace and invited people to make Inkling their own.

Thinking Machines, the lab led by Mira Murati, posted on November 12 that Inkling 2 Small's weights are now available for download on HuggingFace under an Apache 2.0 license. The post said "Make Inkling your own!" and linked to the official Inkling 2 Small model page on HuggingFace.

The official post does not list parameter counts, benchmarks, training data, or a longer technical card in the text itself. What it does confirm is that the Small weights are public and that the license is Apache 2.0, a standard open source license that allows use, modification, and redistribution, including in commercial products, with attribution. Further details, if any, sit on the linked HuggingFace page rather than in the post.

This is a weights release, not a new hosted demo in the post. Builders can download the checkpoint and run or fine tune it themselves, which is the main practical change from a closed API only release. Until independent tests appear, the quality and intended use of Inkling 2 Small should be judged from those later evaluations, not from the announcement text alone.

Watch whether builders actually download and fine tune the weights, and whether Thinking Machines follows this with more Inkling sizes or a longer official writeup. Independent tests will matter more than the announcement, because the post itself gives little on quality or intended use. A next signal would be a first serious public evaluation, or a follow up release that extends the open weights beyond this Small checkpoint. It is also worth watching whether the Apache 2.0 terms stay attached to later Inkling releases, or whether later versions move to a more restrictive license.

World Labs, cofounded by Fei-Fei Li, introduced Atlas on September 1 as a world model that works on text, images, video and 3D. Until now access was limited. On November 12 the company opened its apps and API to everyone.

World Labs posted on November 12 that it is opening access to Atlas to everyone. The company said it built apps and API endpoints so people can bring an idea for a 3D world and come away with keyframes, video clips or 3D Gaussian splats. Fei-Fei Li quoted the post and called Atlas the company's newest world model.

World Labs first described Atlas on September 1 in a blog post. It calls Atlas an omni model pretrained from scratch to work natively on text, images, video and 3D, built as a multimodal autoregressive diffusion transformer that combines all inputs into a shared spatial context. The blog lists camera controlled generation of up to one minute of 1440p video, reconstruction of real scenes from one to dozens of images, simulation of space and time from input videos, and image and panorama generation from text.

The company says Atlas improves with more training compute and that it expects the trend to continue. The announcement did not include public pricing in the post itself, and Atlas weights are not open.

Open access means outside developers can now test how consistent Atlas worlds stay over long camera paths and how well reconstructions hold up on real photos. Watch for pricing and usage limits, early apps built on the API, and robotics teams using it for real to sim work. Competing world models from Google and others are likely to answer with their own access changes.

Antigravity is Google's agent coding environment. On November 12 Google also launched Gemini 4 Hydrogen, a fast model in the Gemini 4 family, alongside the larger Gemini 4 Carbon. Orchestrator lets one session use both models instead of sending a whole task to a single model.

Google posted on November 12 that it is rolling out Orchestrator in Antigravity to Google AI Ultra subscribers. The feature lets a user run multiple models in a single session instead of routing the entire task to one model.

According to Google, Orchestrator pairs a frontier model with a fast execution model. Gemini 4 Carbon sets the plan, hands off code edits and testing to Gemini 4 Hydrogen, and steps back in to repair the plan if the work hits friction.

Google says this setup reaches near Carbon task success rates on Terminal Bench 4.0 while consuming 24% less quota and finishing 19% faster on average. Those figures come from Google and were not independently checked in the posts scanned.

The rollout came on the same day Google introduced Gemini 4 Hydrogen, which it describes as an extremely capable and fast model meant to bring Gemini 4 to everyone. The post included a video of the Antigravity interface with Gemini 4 Carbon selected in the model picker. Orchestrator is limited to the AI Ultra tier for now.

Watch whether Google brings Orchestrator to lower paid tiers or to the Gemini API, and whether outside developers confirm the quota and speed savings on their own projects. Other agent coding tools may adopt similar planner and executor splits between a large model and a fast one. Benchmarks that compare orchestrated runs with single model runs on the same tasks would show how much accuracy, if any, is lost.

Anthropic has been releasing industry plugins for Claude that bundle skills and connectors. Claude for Manufacturing is the latest, announced November 12. ZeroHedge's post on the launch passed 180,000 views.

ZeroHedge posted on November 12 that Anthropic launched Claude for Manufacturing. Anthropic's own announcement describes it as a plugin with 17 skills and four connectors that helps manufacturers draft quotes, quality reports and supplier requests from the systems they already run.

According to Anthropic, the connectors reach Autodesk Fusion, Infor, Microsoft 365 and Oracle NetSuite. The skills cover sales, quality, operations, supply chain, design and setup. Examples in the announcement include reading a customer request and drafting a quote, investigating a failed inspection and drafting a corrective action report, ranking likely causes of a machine breakdown from sensor exports and repair notes, preparing requests for quotes from other suppliers, generating design variations in Autodesk Fusion, and compiling production reports.

Anthropic says every skill ends with a person's review, and that Claude can link figures back to the underlying data. The company says the plugin does not operate equipment, and that it is separately developing a research preview called the Model Hardware Standard for that. The plugin is available on Team and Enterprise plans, and Anthropic scheduled a technical webinar for November 17.

Watch the November 17 webinar for customer examples and details on the Model Hardware Standard preview. The bigger test is whether plant teams trust drafted quotes and corrective action reports enough to use them daily. More industry plugins from Anthropic, and competing offers from OpenAI and Google, are likely as Anthropic heads into its IPO. Partners like Infor and Autodesk may also add deeper connectors over time.

Vercel makes the Next.js framework and a cloud platform for web apps. UI SDK is a public beta TypeScript library released under the Apache 2.0 license, with adapters for React, Vue, Svelte, Solid and Web Components. Peter Steinberger quoted the launch, writing that agents will learn to write this code for you.

Vercel posted "Introducing the UI SDK" on November 12 with a launch video. The project's announcement describes UI SDK as a TypeScript library of UI behaviors, the layers that make up a component library, made for agents.

The library provides building blocks for keyboard navigation, focus, selection, popups, drag and drop and the ARIA attributes those interactions need. Vercel says it built more than 50 of these behaviors, and that a person or an agent combines them into custom components. It also used them to build ready made unstyled components, from a Select to a Date Picker, with the same parts across React, Vue, Svelte, Solid and Web Components.

The second half of the pitch is about agents using interfaces. Vercel says anything built with UI SDK can be read and operated by an agent through a bridge called createAgent, which can take snapshots of the page and run actions within permissions the app sets. The docs also describe support for WebMCP, so an agent built into a browser can call the components' actions.

The GitHub repository lists the project as a beta under the Apache 2.0 license, and says APIs may change before general availability. Peter Steinberger quoted the launch with the line that you do not need to learn how to write this because your agents will.

Watch how quickly coding agents and design tools adopt UI SDK as a default component layer, and whether the agent bridge becomes a common way for assistants to operate web apps. Vercel's beta terms mean APIs can still change. Early independent tests of the permission checks, and any security issues they find, will shape how far teams trust agents to act inside their apps.

Grok Bot is an AI assistant product whose official account launched Grok Bot 1.0, shared AI teammates for whole teams, on November 11. On November 12 the account said the Mac version can now connect to seven built in Apple apps. Elon Musk quoted the post and urged people to give Grok Bot their most laborious tasks.

The official Grok Bot account posted on November 12 that Grok Bot on Mac can now use Messages, Mail, Notes, Reminders, Calendar, Location and Contacts. The attached image shows a Connect Apps panel where each app has its own Allow button, with short descriptions such as read and send messages, search and draft emails, set and check reminders, find and add contacts, and read and write notes.

The post came one day after the same account introduced Grok Bot 1.0, which it described as AI teammates that work alongside a whole team. The new Mac integrations extend the product from shared team bots to personal tasks on a user's own computer, with permission granted app by app.

The update drew wide attention after Elon Musk quoted the announcement and separately quoted a post from Maye Musk, writing "Give Grok @Bot your most laborious tasks!" In her post, which passed one million views, Maye Musk said she asked the bot how to move 2,000 photos in Dropbox from Camera Uploads to folders, and that it offered to do the task and finished it in five minutes. That account is a single user's experience and not a benchmark.

The announcement did not say which Mac versions are supported or whether the features cost extra.

Watch for details on pricing, supported macOS versions and privacy controls, since the integrations can read messages, email and contacts. Similar desktop features from OpenAI, Anthropic and Google will invite direct comparisons. Further updates may show whether Grok Bot adds Windows support or more third party apps, and how the shared team bots from version 1.0 connect with these personal integrations.

Unusual Whales posted that the US is pulling away in the AI compute race, according to Epoch AI. The post said that since 2022, about 72% of the world's AI computing power has ended up in the US, and roughly 10% in China, and that the US share of new compute keeps climbing. No official post from Epoch AI on these figures was found above the view floor, so this lead is secondhand.

Unusual Whales posted on November 12 that the US is pulling away in the AI compute race, according to Epoch AI. The post said that since 2022, about 72% of the world's AI computing power has ended up in the US, and roughly 10% in China.

It also gave a year by year picture of the US share of new compute: about 60% a year from 2022 through 2024, 72% in 2025, and 78% in the first half of 2026. That last figure, if accurate, would mean the US is now taking an even larger share of newly added capacity than of the installed base since 2022.

A search of the official Epoch AI account did not turn up a post above the view floor that states these figures, so the lead is secondhand through Unusual Whales. The numbers should be read as Epoch AI figures as reported in that post, not as a first hand release on this scan. The post also does not define how compute was measured, which countries sit in the remaining share, or whether the figures count training clusters, inference, or both.

Watch for Epoch AI to post the underlying data or paper in its own voice, which would let readers check definitions, time windows, and how China and other regions were counted. If the 78% figure for the first half of 2026 holds, later updates will show whether the US share of new compute keeps rising or levels off. Policy debates about export controls and domestic buildouts will likely use these numbers, so the first official tables matter.

Jukan is a semiconductor analyst account that often relays supply chain reports. In this post he responds to rumors that Nvidia cut orders and says he confirmed a new Rubin delay through his own contacts. Nvidia has not commented, so the claim is unverified.

Jukan posted on November 12, in response to rumors that Nvidia cut orders, that one thing he was able to confirm is that Rubin has been delayed again. He says he was told that motherboard warpage caused the latest delay and that an HBM issue was resolved in October.

As a result, according to the post, volume shipments of Rubin have slipped to 2027, and Nvidia has increased its Blackwell orders in the meantime. Jukan also notes that Korean media reported the same day that SK hynix's HBM4 yield had reached 70%.

On the wider order cut rumors, he says he spoke with several funds and supply chain contacts. One rumor going around is that CCL orders were cut by 30%, but he says one source told him that claim was false.

This is an unverified claim from an analyst account citing unnamed sources. Nvidia has not confirmed any delay. Additional reporting, secondhand: an X news summary says Baird downgraded Nvidia to Neutral on November 11, citing limited visibility on 2027 data center demand and supply hurdles, ahead of the company's earnings on November 17.

Nvidia's earnings report on November 17 is the clearest place for the company to address Rubin timing, Blackwell supply and any order changes. Watch for confirmation or denial from Nvidia or its board and manufacturing partners such as TSMC, and for updates from HBM suppliers like SK hynix. A confirmed slip to 2027 would affect the compute roadmaps that labs and cloud providers have shared.

The Kobeissi Letter posted that SpaceX and Tesla could merge as soon as December in what would be the largest merger of all time, according to Bloomberg. The same post said the companies' boards and advisors are now reportedly hammering out the terms of the agreement. No official post from SpaceX, Tesla, or Bloomberg on this claim was found above the view floor, so the lead is secondhand and the claim is unverified.

The Kobeissi Letter, a market commentary account, posted on November 12 that SpaceX and Tesla could merge as soon as December in what would be the largest merger of all time, according to Bloomberg. The post said the companies' boards and advisors are now reportedly "hammering out" the terms of the agreement.

The post itself is the only source used here. No official statement from SpaceX, Tesla, Elon Musk, or Bloomberg was found above the view floor that confirms a merger timetable or that talks are underway. SpaceX did post the same day about supercompute campuses in Texas and Arizona, but that announcement does not mention a merger with Tesla.

Until a first hand account, an official filing, or a Bloomberg post on its own account appears, the merger claim should be treated as unverified. The wording in the Kobeissi post is also tentative: it says the companies could merge as soon as December, not that a deal has been signed.

Watch for a confirmation or denial from SpaceX, Tesla, or Musk, or for Bloomberg to post the report on its own account. A securities filing or an official board announcement would be the first hard evidence that talks are real. Until then, treat any market reaction as a response to an unverified report, not to a completed deal. It is also worth keeping this claim separate from SpaceX's official campus announcement, which does not depend on a Tesla merger.

Tibo, who posts Codex and ChatGPT product updates, said OpenAI announced new tools it is developing for cybersecurity, covering automated patching and automated red teaming, and linked to a recording on OpenAI's site. The linked page is titled Cyber at the Frontier and describes Daybreak Defense Factory and Daybreak Red Team.

Tibo posted on the evening of November 12 that "We announced new tools we're developing for cybersecurity today across automated patching and automated red teaming," and linked to a recording. The wording is that the tools are being developed, not that they are generally available.

The linked OpenAI page is titled Cyber at the Frontier and invites people to talk with the team about AI for cyber defense. It lists resources for a keynote, announcements, and guidance, and names Daybreak Defense Factory and Daybreak Red Team. The page describes Daybreak Red Team as a way to test web apps, validate vulnerabilities, and verify fixes with coordinated AI agents, and it includes a request for Daybreak access.

The official OpenAI account did not have a matching launch post above the view floor in this scan, so the lead is Tibo's post and the OpenAI page it points to. That still makes this an official product thread, because Tibo is a listed source for ChatGPT and Codex updates and the link goes to OpenAI's own site.

Watch whether Daybreak moves from a request for access and a keynote recording to a product that outside teams can actually use. Tibo framed the work as tools in development, so the next signal is a broader availability date, a model card, or examples of automated patching in production. Independent security researchers will also test whether automated red teaming finds real bugs, and whether automated patching is safe enough to trust.

Simulation lets developers of self driving cars, drones and robots test rare or dangerous scenarios without real world risk. Zeromatter's founders previously led simulation for Tesla's Autopilot and Optimus programs. The company announced its launch on November 12 after four years in stealth.

Zeromatter posted on November 12 that it started four years ago with a single bet: that building autonomous machines would be faster, safer and better done first in a virtual world. The company said it was coming out of stealth to show what that bet has become.

In a blog post the same day, the company says its cofounders include Axel Gneiting, a former head of Autopilot Simulation at Tesla, and Michael Hosticka, a former head of Autopilot Simulation Content. The post says the founders built a single simulation platform for Autopilot and Optimus while at Tesla.

Zeromatter says its team has grown from 13 to more than 130 people and that it has raised $160M at a $1.1B valuation. It lists seed investors including Two Sigma Ventures, HNVR, Catapult Ventures, Caffeinated Capital and Soma Capital, with angel checks from Guillermo Rauch of Vercel and Boris Sofman of Bedrock Robotics.

The company's website describes the product as a deterministic, physics based reflection of reality that is not a video model or a game engine. It says the platform simulates sensors such as cameras, lidar and radar, and that the team measures real world sensor data in its own labs to keep the gap between simulation and reality small. These claims come from the company.

Watch for named customers in autonomous driving, drones, aviation or robotics, and for published evidence that training in Zeromatter transfers to real world performance. The company will compete with established simulation tools such as Nvidia's, so pricing and availability details will matter. Further funding or partnerships with robot makers would show how fast the physical AI simulation market is growing.

Simular builds computer use agents. Its new product, box, gives AI assistants a cloud Linux desktop they connect to as a remote MCP server. The launch post passed 1.7 million views.

Simular posted on November 12 that it is launching box, which it calls one computer for ChatGPT, Claude, Codex, Gemini, DeepSeek and other agents. The post says agents using box can open apps, run the terminal, fill in forms and stay signed in for later tasks.

Simular's announcement describes box as a cloud computer built for agents doing computer tasks. It says users can add box to chat apps like Claude, ChatGPT or Gemini, give always on agents such as OpenClaw, Hermes or Claude Code a shared machine, or use it with coding agents like Codex and Cursor. Logins, files and apps live on the box, so one agent can start a job and another can finish it.

The company claims its agent reaches state of the art performance on the OSWorld benchmark with the best cost efficiency, and says box is 10x faster and cheaper, a figure from Simular that has not been independently checked. The product page says box runs as a hosted MCP server, that each account gets one cloud Linux desktop that sleeps when idle, and that the box itself is free while model calls use account credits.

Watch for independent tests of Simular's OSWorld and cost claims, and for how box handles security when many agents share one signed in machine. Support for Windows, macOS and Android, which Simular has said will come later, would widen its use. Rivals offering cloud desktops for agents are likely to respond on price and persistence. Pricing for heavy users is another detail to watch.

Retell AI makes software for building voice AI agents, used mainly in customer contact centers. Its cofounder and CEO Bing Wu posted on November 12 that the company crossed $100M in annual revenue run rate. A run rate projects current revenue over a full year.

Bing Wu, cofounder and CEO of Retell AI, posted on November 12 that Retell just crossed $100M in revenue run rate. He said the company went from $10M to $100M in 17 months, is profitable and has raised $4.6M in total.

Wu wrote that from day one the company focused on solving the hardest parts of deploying voice AI inside the product itself, so customers can build, manage and improve their own agents without depending on Retell. He described the company's bet as "the future of the contact center belongs to the teams who run it." The post included a video and passed 1.5 million views.

Run rate figures project current revenue over a full year and are not audited annual revenue. The post did not share customer counts, margins or a breakdown of revenue, so the figures come from the company alone.

Voice agents that answer and place calls are one of the most commercial uses of AI models today. If accurate, the figures put Retell among the fastest growing startups in the category, with unusually little outside funding.

Watch whether Retell raises a larger round at a valuation that reflects the reported run rate, or stays mostly self funded. Competition from larger platforms and model labs that sell voice agents directly could pressure prices. Customer case studies or independent revenue estimates would help confirm the figures, and any expansion beyond contact centers into sales or other phone work would show where growth comes next.

Amplitude is a product analytics company, and Spenser Skates is its CEO and cofounder. Wave is a product agent inside Amplitude that reads analytics, session replays and customer feedback. Skates announced it on November 12.

Spenser Skates, CEO of Amplitude, posted on November 12: "Introducing Wave. The world's first AI that reads your customers' mind." He wrote that $1.47 trillion will be spent on software this year but only a fraction of features are used by customers, and that AI can build anything but does not know what customers need.

The two minute launch video describes Wave as an agentic product team. In the video's example, Wave watches thousands of user sessions, reads product data and collects feedback from around the internet, then finds that high value customers are not using a feature because it is buried too far down the page. It suggests moving the feature, drafts a development plan with a pull request and experiments, and estimates the revenue impact.

The video also says Wave caught 700 users getting stuck at login, wrote and tested a fix and pushed the pull request for approval, and that it can propose many improvements in parallel. These examples come from Amplitude's own demo.

The $1.47 trillion figure matches Gartner's July 2026 forecast for worldwide software spending, which Amplitude cites. The post passed 400,000 views.

Watch for customer results from teams using Wave, including how often its suggested changes ship and whether they move the metrics Amplitude projects. Rival analytics and session replay companies are likely to ship similar agents. How much autonomy teams give Wave to push code without review will be a key test of trust. OpenAI and other platforms that already sit inside product tools may also add features that compete with Wave.

Hang Huang is the founder of InstaBox, a Y Combinator company. InstaBox offers sandboxed cloud machines that AI agents can use to run code and apps. The launch came on the same day Simular launched box, another cloud computer for agents.

Hang Huang posted on November 12: "Your laptop isn't built for running agents." He introduced InstaBox as cloud computers for AI agents, offering sandboxes running Linux, Android, Mac and Windows with GPU and CPU options.

Huang wrote that InstaBox launched in beta less than a month ago and that users have already created more than one million boxes. That usage figure comes from the company and was not independently checked.

The pitch is that agents running on a single laptop fight over the same memory, CPU and ports. By giving each agent its own cloud computer, InstaBox says agents can keep working after the user closes the laptop. The post linked to the product site and included a video in which Huang closes a laptop.

The launch was one of two cloud computer products for agents on November 12. Simular launched box the same day, pitching a shared cloud desktop that ChatGPT, Claude, Codex and other agents can operate. The launch post passed 2.7 million views.

Watch for pricing details for heavy use, support for more operating systems at scale, and how InstaBox handles security when agents hold logins and files. Competition among cloud machines for agents is growing quickly, with Simular and others launching similar products. Integrations with popular agent tools like Claude Code and Codex will likely decide which service developers adopt first. Usage numbers after the beta will also show whether demand holds.

Fish Audio makes AI voice models. Fish Agent is its new platform for building voice agents that handle phone calls. It runs on Fish Live 3, which the company calls a full duplex voice model.

Fish Audio posted on November 12 that it is introducing Fish Agent, a no code platform to build custom AI voice agents that can make and take calls. The company says the median deploy time is 32 seconds.

According to the post, users can build any voice agent by describing it, get a phone number or connect their own, and power the agent with leading LLMs or a custom LLM. Agents can connect to tools, webhooks and knowledge sources with full API access.

Fish Audio says conversations are natural and aware of interruptions thanks to its Fish Live 3 full duplex voice model, meaning the agent can listen and speak at the same time. The company suggests uses from a personal AI assistant to a receptionist or customer support line, and offers enterprises a demo with a dedicated engineer.

As part of the launch, Fish Audio said it was giving away platform credits for 48 hours to users who liked, reposted, bookmarked and replied to the post. The launch post passed 2.2 million views. The deploy time and quality claims come from the company.

Watch for independent tests of Fish Live 3 call quality and latency against rivals such as Retell and ElevenLabs. Pricing after the launch credits end will matter for small businesses. Rules on AI phone calls, including disclosure requirements, may also shape how voice agents like these are used for outbound calling and in regulated industries. Enterprise deals that move past demos would show whether the platform can handle real call volumes.

Expo is a widely used platform for building mobile apps with React Native. Its new service gives agents cloud simulators so they can check their own code. The launch post passed 1.9 million views.

Expo posted on November 12 that coding agents write mobile code faster than any team can review it, and introduced Expo Simulators, cloud simulators built for agents. According to the post, an agent runs its change on a cloud simulator or emulator and brings back the recording before it asks for a person's attention.

Expo's press release says the service lets agents reproduce a reported bug, write and test a fix, verify it on iOS simulators and Android emulators, and leave a screen recording, screenshots and logs on the pull request. It says the service works with agents teams already use, including Cursor, Claude Code and Codex, and that iOS simulators can run without a Mac, from Linux, Windows through WSL, or a cloud sandbox. Expo says the product adds no AI model of its own.

The release gives pricing of $0.10 per simulator minute for iOS and Android, with 60 minutes a month included on Starter, Production and Enterprise plans, and says it is available to everyone on paid plans. Expo's documentation still describes a limited access preview with a waitlist, so access may differ by account.

Watch for how quickly teams wire Expo Simulators into continuous integration, and whether recordings on pull requests actually speed up review. Expo said it expects to share details on prompt injection protections by the end of November. Similar agent focused testing services for web and desktop apps are a likely next step across the tooling market. Pricing per simulator minute will decide how many teams run checks on every change.

Ivan Novikov, founder of the security company Wallarm, now lists XERJ in his X profile. On November 12 he posted that he built a new AI model that is not allowed to generate text. Neurosymbolic systems combine neural networks with explicit symbolic rules or code.

Ivan Novikov posted on November 12 that he has built a new AI model that is not allowed to generate text. Instead, he said, it produces a verifiable program that answers a prompt using the original reference. He described XERJ Reader 1 as a neurosymbolic model for working with data: it points to the exact words in a document, writes a small program over them, and exact code computes the answer.

Novikov claimed the approach is 20 to 90 times faster and 10 to 1,000 times cheaper, with output tokens free, and that every answer is a program that can be replayed. He called it the shortest path to automating data work. The post included a demo video and passed 560,000 views.

The speed and cost figures come from the developer, and the post did not name the baseline models or benchmarks they were measured against, so they are unverified. No paper or independent evaluation was linked in the post.

The idea targets a known weakness of language models on data tasks: they can misquote sources or make arithmetic errors. Grounding each answer in quoted text and a runnable program makes errors easier to check, at the cost of handling only tasks that can be expressed as code.

Watch for a technical report, public benchmarks or open access that would let others test the speed and cost claims. Comparisons with standard language models on document question answering and spreadsheet tasks will show whether program based answers hold up on messy real data. Pricing, an API and early customer use would show whether the approach can compete with general assistants that already handle data work.

Mehul Nath Jindal posted the launch of Aurora on November 12. Aurora is an AI search tool for local businesses, reachable through the account aurorax_ai. The launch video passed two million views.

Mehul Nath Jindal posted on November 12: "Introducing Aurora, the Search Engine for IRL." The post asks whether readers have had to go shop to shop, ask friends and family, or spend weeks browsing online to find which business has what they want.

Jindal gave examples ranging from renting an apartment or booking a doctor's appointment to finding a favorite shoe in the right size at a nearby store. He wrote that users can now just ask Aurora and it will tell them who has it.

The launch post included a video opening on a rotary telephone, and launch trackers describe Aurora as contacting local businesses by phone, text and email to check availability and pricing before returning a shortlist. The post did not give details on coverage, pricing or which AI models power the service, so those remain unconfirmed.

The post passed two million views within a day, one of several heavily viewed AI product launches on November 12.

Watch for which cities and business categories Aurora covers, and how businesses respond to automated calls and messages checking their stock. Accuracy of availability answers will be the key test. Larger AI search and shopping products from Google, Amazon and OpenAI could move into local inventory search, putting pressure on startups in this space. Details on pricing and which AI models power the service would also help users judge it.

Martin Cordsmeier is cofounder and CEO of millionways, a company building AI models of human behavior. Thorsten 4 is the company's Large Psychology Model. The company announced it on November 12.

Martin Cordsmeier posted on November 12 that AI can write code, pass the bar exam and sound human, but still cannot tell when a person is unsure, holding back or about to walk away. He introduced Thorsten 4, which he called a Large Psychology Model that predicts what people will do next.

Cordsmeier leads millionways, which describes Thorsten 4 as its proprietary model for analyzing communication to identify behavioral patterns, motivations, decision styles and likely reactions. The company's stated goal is to help AI understand human behavior in business interactions such as sales and customer conversations.

The launch post included a cinematic video set in a control room. It did not share benchmark results, pricing or details on how the model was trained or evaluated, so its predictive claims are the company's own and have not been independently tested.

The post passed 100,000 views, which put it just over the viral threshold for this deck.

Watch for published evaluations of Thorsten 4's predictions, customer deployments in sales or support, and how millionways handles privacy when analyzing people's conversations. Larger labs are adding emotion and intent signals to voice and chat models, so specialised behavior models will need clear evidence that they do better. Regulators in Europe have also restricted some emotion recognition uses, which may limit where tools like this can be deployed.

Arushi Gandhi announced Benchmark Router on November 12. The service runs public benchmarks for AI builders on a model or agent harness of their choice. It reports metrics, full logs and failure analysis.

Arushi Gandhi posted on November 12 introducing Benchmark Router, which she described as one place to run public benchmarks on any model or harness at a tenth of the cost of running them yourself. The post marks that cost claim with an asterisk and does not explain the conditions.

Gandhi wrote that new LLM benchmarks are launching every six hours, and that the platform's goal is to give AI builders a unified place to run benchmarks without dealing with infrastructure and to see how their agent stack ranks.

According to the post, users pick one or more benchmarks, one or more models and a configuration, and the service runs the official benchmark code and returns metrics, full logs and failure analysis. She said the platform is live with benchmarks across knowledge, coding, agentic tasks, law, finance, health, science and vision, and that users can request more.

The post passed 140,000 views. The cost and coverage figures come from the company.

Watch whether model labs or agent startups cite Benchmark Router results in their launches, and whether its runs match scores labs publish themselves. Pricing details behind the tenfold savings claim will matter for adoption. Competition from evaluation platforms such as Artificial Analysis and Arena may shape which benchmarks it adds next and how widely it is trusted. Published price lists would let teams compare it with running benchmarks themselves.

Agent development environments, or ADEs, are desktop apps that wrap coding agents in a graphical interface. ADEMark is a new benchmark posted by the account zqkrq that tests them on the same machine and tasks. The first results were published on November 12.

The account zqkrq, display name zakra, posted on November 12 that ADEMark v1 is out. The test used the same Mac mini, the same tasks and the same model for every app, with a bare command line interface as the baseline.

The author wrote that they did not expect most of the apps to use fewer tokens than the command line on its own, or one app to sit at 16% CPU while doing nothing. They said this is the first of many runs and that updates will come more often.

The results table in the post is titled 15 ADEs, 12 tests, one Mac. It lists apps including super.engineering, Zeron, Paseo, Diri, T3 Code, Emdash, MonoCode, Synara, bb, Orca and Traycer, with Claude desktop, Capy, Cloudroom and Superset measured only in part. Tests cover speed, such as cold start and typing lag, weight, such as memory at rest and idle CPU, and tokens used. The table notes that each test is ranked on its own and that it is not a single score.

According to the table, the runs used Claude Opus 5.5 at high effort on a Mac mini M4 with 16 GB of memory. The post passed 160,000 views.

Watch for updated ADEMark runs with more apps and models, and responses from the developers of the tested environments. Independent reproduction on other hardware would make the results more useful. If the benchmark gains traction, app makers may start optimizing for its idle CPU and token measurements, and later versions may add a single overall score. Results with other models besides Claude Opus 5.5 would also show whether the rankings hold.

ChatGPT and Claude both offer connectors that read and act on Microsoft 365 data such as Outlook mail and Teams. On November 12 a broad Microsoft 365 outage disrupted those services for thousands of users.

Reuters posted on November 12 that Microsoft 365, Outlook, Teams and other Microsoft apps remained disrupted for thousands of users for more than five hours on Thursday. Reuters said the outage also hit connectors used by OpenAI's and Anthropic's AI services.

Earlier in the evening, Gergely Orosz posted that there was a massive Microsoft outage, that Outlook.com appeared to be down globally, and that Microsoft employees and other Microsoft 365 users could not send or receive email. He attached a screenshot of a Microsoft 365 service status notice marked service degradation, which said users may be unable to connect to multiple Microsoft 365 services. The notice listed Exchange Online, Microsoft Teams, SharePoint Online, Microsoft Copilot 365, Microsoft Defender XDR, Planner, To Do, Intune, Defender for Office 365, several Purview services and Power BI exports among the potentially affected services.

The posts did not give a cause for the outage. The incident shows how AI tools that act on a company's mail and files inherit the reliability of the systems they connect to.

Watch for Microsoft's postmortem on the cause and duration, and for any statements from OpenAI or Anthropic on how their connectors behaved. Enterprise buyers may push AI vendors for clearer fallbacks when a connected service is down. Repeated outages would add weight to arguments for spreading critical workflows across more than one provider. Watch also whether Microsoft offers service credits to affected customers.

ChatGPT posted that the dot now has faster replies, less jargon, and better follow through, and that it can follow custom instructions, including examples such as always talking in lowercase or showing a draft before sending email. Tibo, who posts ChatGPT product updates, later said some good updates to dots shipped and that the banked reset has been activated.

ChatGPT posted on November 12 that the dot now has faster replies, less jargon, and better follow through. The same post said the dot can now follow custom instructions. It gave examples of telling the dot in the thread how to behave or which actions need a go ahead, including "always talk in lowercase" and "Show me a draft before sending any emails."

The post also said users can set custom instructions and rules in Customize, then Behavior, on mobile, or through the dot's menu on web and desktop. That means the change is not only a one off prompt in a thread, but also a setting users can save.

Tibo, who posts Codex and ChatGPT product updates, later quoted that ChatGPT post and said some good updates to dots shipped today and that the banked reset has been activated and should land in everyone's accounts sometime soon. The post does not spell out what the banked reset changes, only that it has been activated and should arrive soon.

Watch whether users actually see faster and plainer replies from the dot, and whether custom instructions hold from one thread to the next. Tibo said the banked reset should land in accounts soon, so the next check is whether that credit or quota change shows up as promised. This is a small product update, but it sits on a feature many ChatGPT users touch every day. If the custom instruction controls work as described, the more important test is whether the dot waits for a go ahead on actions like sending email.

Revenue comparisons between US and Chinese AI companies are rare because most are private. The Kobeissi Letter's chart credits Epoch AI and Nikkei Asia as sources. Anthropic's Q3 revenue figure of $17.0B also appears on its IPO site.

The Kobeissi Letter posted on November 12 that US AI is generating far more revenue than Chinese AI. According to the post, OpenAI and Anthropic had a combined estimated annualized revenue run rate of $105 billion as of the end of September, nearly 10 times the $11 billion generated by six Chinese AI companies: Alibaba, ByteDance, DeepSeek, Z.ai, Moonshot AI and MiniMax.

The post says ByteDance accounted for 36% of the Chinese figure at $4 billion. It says Anthropic's sales grew 47% quarter over quarter in Q3 to a record $17 billion, and that OpenAI revenue rose 50% to a record $10 billion. It adds that DeepSeek's annualized run rate doubled in September but stands at about $1 billion.

The attached chart is labeled as an estimate and lists Epoch AI and Nikkei Asia as sources. The run rate figures are estimates, not reported results, and the post does not explain how revenue was attributed for large companies like Alibaba and ByteDance, whose AI sales are part of wider businesses.

Anthropic's IPO filings and any OpenAI disclosures will test these estimates against audited numbers. Watch whether Chinese labs, several of which release open weights models, turn usage into revenue faster in coming quarters. Export controls on chips, now in focus after the Huang resignation and the DePue claims, could also shape how fast Chinese firms can grow. Independent estimates from Epoch AI may also be updated as new filings appear.

Unusual Whales posted that cannabis growers are trying to sell their facilities to AI data center developers as marijuana prices crash, according to Forbes. The same post said indoor grow facilities already have much of what data centers need: heavy power, fiber, water and industrial HVAC. No official Forbes post on this report was found above the view floor, so the lead is secondhand.

Unusual Whales posted on November 12 that cannabis growers are trying to sell their facilities to AI data center developers as marijuana prices crash, according to Forbes. The post said indoor grow facilities already have much of what data centers need: heavy power, fiber, water and industrial HVAC.

The post does not name specific deals, buyers, prices, or locations, and no official Forbes post above the view floor was found in this scan. The claim should be read as a Forbes report as relayed by Unusual Whales, not as a completed sale or an official industry tally.

The underlying idea is that a building already wired for dense power and cooling might be cheaper or faster to convert than a greenfield data center. Whether that is true depends on details the post does not provide, including local permits, the voltage on site, and whether developers have actually signed term sheets.

Watch for named transactions, not just the general report, and for whether data center developers actually close on former grow sites. Power availability, local permits, and whether the existing electrical gear meets data center needs will decide if this is a real pipeline or only a headline. A first confirmed conversion would be the evidence that matters. Until then, treat the Forbes report, as relayed here, as a possible trend rather than a completed shift in the data center market.

Polymarket posted that Anthropic CEO Dario Amodei declared cancer, Alzheimer's, heart disease, and mental illness will be relics of the past within a decade. No official post from Amodei or Anthropic on this quote was found above the view floor in this scan, so the lead is secondhand.

Polymarket posted on November 12 that Anthropic CEO Dario Amodei declares cancer, Alzheimer's, heart disease, and mental illness will be "relics of the past" within a decade.

The post does not include a clip, a venue, a date for the remarks, or a longer quote, and no official post from Amodei or Anthropic above the view floor was found that states the same claim in this scan. The wording should be treated as a secondhand report of a prediction, not as a research result or a product announcement.

Amodei has spoken before about AI and biology, but this item is only what Polymarket said he declared. Without the original remarks, it is not possible to say whether he set conditions, cited a specific model, or was speaking more loosely about a longer research agenda. Until that original source is posted, the item belongs in the deck only as a widely circulated claim from a high profile CEO, labeled as secondhand.

Watch for Amodei or Anthropic to post the full remarks, which would show the exact wording and context. Until then, this remains a widely circulated prediction rather than a documented scientific claim. Any later medical or model news from Anthropic should be judged on the evidence, not on this forecast. If a recording or transcript appears, it should be compared with the Polymarket wording before anyone treats the quote as settled.