November 13th 2026
Curated AI news and stories.
Nvidia closes below a $4 trillion market value for the first time since July 2025, as HSBC downgrades it to Hold
Nvidia was up about 2% before the open and closed Friday down 2.01%, at about $162.85, which took its market value under $4 trillion for the first time since July 2025. The stock fell 12.75% over the week. It has now fallen for six straight weeks, its longest losing streak since 2012.
HSBC analyst Frank Lee cut the stock to Hold from Buy and his target to $170 from $365, saying investor confidence "has been shaken" and calling the move "a reset amid the leadership transition." He said a re-rating needs a permanent chief executive who shows he or she can deliver the product roadmap. Investing.com puts the shares 22% below where they were on October 29th, when The New York Times first reported the intelligence concerns, against a 4% fall in the Nasdaq, at 11 times , the lowest multiple in 15 years. Wedbush's Matthew Bryson took the other side: "Nothing has changed," he said, and he expects a beat and a raise on Tuesday. SourcesBBB
Nvidia's board has approached Arm chief executive Rene Haas to succeed Huang, the FT reports
Nvidia's board has approached Rene Haas, Arm's chief executive, as it looks for a permanent successor to Jensen Huang, the Financial Times reported Friday, citing people familiar with the matter. Haas is one of several candidates and there is no guarantee he is offered the job or takes it. Nvidia did not answer the FT's questions, Arm declined to comment, and Reuters could not verify the report. Arm shares closed down 4.8%.
Haas spent seven years at Nvidia, rising to run its computing products business, before joining Arm in 2013; he has been Arm's chief executive since February 2022 and took it public in 2023. Other names in circulation include AMD's Lisa Su, raised by analysts, and two insiders, the longtime Nvidia vice president Ian Buck and cofounder Chris Malachowsky. Colette Kress stays interim chief executive while the board searches inside and outside the company. Hiring the head of the company that licenses the designs inside Nvidia's Grace chips would put a supplier's chief in the job. SourcesBBC
Bessent says Treasury wanted Megaspeed blacklisted for months and that "Commerce has the pen"
Treasury Secretary Scott Bessent said on CNBC on Friday that Treasury, the Pentagon, State and the US Trade Representative had favored adding Megaspeed, the Singapore cloud company at the center of the Nvidia affair, to the Entity List "for months." "But Commerce has the pen," he said. "You will have to ask them about their internal processes." Asked about an Axios report that he confronted Howard Lutnick in the White House Situation Room after the midterms over Megaspeed, Bessent said, "We don't repeat conversations in the Situation Room."
Megaspeed is still not on the list. The Justice Department is investigating it for suspected export violations, and The Information has reported that it and its Malaysian subsidiary Speedmatric shipped at least $7.5 billion of Nvidia chips to China over two years. Commerce did not respond. Reuters has reported that White House officials are weighing whether to remove Lutnick, who now goes to Senate Banking on December 10th with Treasury on record against him. SourcesBB
Nscale tells regulators it needs more money for 2027 and has no binding financing for its $44.6 billion Anthropic build
Nscale, the British preparing to list in New York, filed an amended registration on Thursday showing a preliminary nine-month net loss of $1.69 billion, up from $433.3 million a year earlier, on revenue of $307.5 million, up from $21.3 million. Debt rose from $78.9 million at the end of 2025 to an estimated $5.52 billion at the end of September. "We will be required to obtain additional financing to meet our forecast obligations in 2027," the filing says. Its September filing said it had enough money for at least 12 months.
The largest obligation is a contract with Anthropic worth up to about $44.6 billion for 460 megawatts at a campus in West Virginia, where Nscale estimates the first building alone needs $9 billion to $9.5 billion more of data center and power spending. It has no binding financing for that buildout. A $3.36 billion convertible round led by Third Point, including $1 billion from Nvidia, is due to close around November 16th.
Neocloud shares fell again on Friday: CoreWeave closed down 8.8% and Nebius 9%, while AMD lost 10.7%, Broadcom 5.9% and the Philadelphia semiconductor index 5.1%. Apple touched a record as investors moved into it as a safe haven. SourcesBCB
OpenAI launches ChatGPT Wallet so its agents can pay for purchases
OpenAI launched ChatGPT Wallet on Friday, built with American Express, Mastercard and Visa, so that its always-on , which it calls "dots," can buy things for a user. At launch it is limited to ChatGPT Pro subscribers in the US who are 18 or older with an eligible US-issued card, and it works only inside dots, not in ordinary chats or Codex. Before each purchase the user sees the items, the total and the card, and approves with Face ID or a fingerprint. The card network issues a in place of the card number, so neither the agent nor the merchant sees it.
OpenAI says it expects users eventually to let agents buy within limits they set, such as a spending cap, a category or a time window, with no per-purchase check. Some merchants block agents outright. SourcesBC
SpaceX plans more than $74 billion for AI computing campuses in Texas and Arizona
SpaceX said on Thursday it will spend more than $24 billion on a campus at Starbase, Texas, and about $50 billion on one in Cochise County, Arizona. The Texas site will make its own electricity, from a gas-fired plant on about 444 acres, instead of drawing from the grid. The Arizona campus would use dry cooling with no industrial groundwater, and SpaceX says the deal with local cooperatives could cut members' power bills by up to 20%.
Neither is approved. Cochise County holds a hearing in December, and a petition with 2,400 signatures asks it to refuse the permit; the Starbase City Commission takes up a property tax abatement on December 9th. SpaceX's AI business, which listed in June, produced more than half the company's third-quarter revenue. SourcesBB
OpenAI, Anthropic and Microsoft sign Macron's Paris Call to build shutdown switches into frontier models
Ten AI companies and six governments signed the Paris Call for AI Safety and Security at the Élysée on Monday. The companies are OpenAI, Anthropic, Microsoft, Mistral, , Cohere, Black Forest Labs, Domyn, Sakana AI and Synthesia; the countries are France, Canada, Finland, Norway, the Netherlands and Bahrain. They commit to fail-safe mechanisms that can restrict, pause or fully shut down a model, including during training and internal use, plus independent testing where appropriate, published non-sensitive test results, early warning of serious incidents and secure channels for whistleblowers.
The United States, China, Google and Meta did not sign, though Treasury Secretary Scott Bessent attended the discussions. Nothing in it is binding and the switches stay with the developers. CeSIA, a French safety group that signed, says it sets no shared risk thresholds and gives governments no power to suspend a model: "These decisions must rest with public authorities," said its director, Charbel-Raphaël Ségerie. SourcesBB
EU leaders hold their first summit on AI in Dublin, with Mistral's Mensch and Geoffrey Hinton briefing them
European Union leaders met in Dublin on Friday for their first summit devoted to AI, briefed by Arthur Mensch, Mistral's chief executive, and Geoffrey Hinton. Irish Prime Minister Micheál Martin said they would weigh "the balance between regulation and innovation, tech sovereignty." European Parliament President Roberta Metsola said she was tired of hearing the EU had "somehow lost the AI race." Foreign policy chief Kaja Kallas warned of AI's "dark side," pointing to AI-equipped drones in Ukraine. Estonia and Finland wrote to colleagues asking for "EU-level arrangements for responding to serious AI incidents."
The summit was informal and nothing was decided. An EU official said Nvidia was not on the agenda but that leaders were likely to discuss the security side of the affair. SourcesBB
Microsoft blames a network upgrade for its 24-hour outage, and Claude's incidents are all closed
Microsoft apologized for the outage that took down Outlook, Teams and Microsoft 365 for most of a day, saying the fault was introduced during an earlier network capacity upgrade and surfaced after a later network change. Executive Charles Lamanna said there was no indication of "malicious or cyber activity." Microsoft says it will publish a root-cause report once its investigation is finished.
Anthropic's status page now gives the impact on Claude's Microsoft 365 and Claude for Outlook as 16:03 UTC on Thursday to 16:40 UTC on Friday, and closed the incident at 21:49 UTC. The second Claude incident on Friday morning, which degraded claude.ai, the and Claude Code, closed at 15:38 UTC. Its last update said scheduled tasks and routines could run more than two hours late and some runs were skipped. Anthropic has still given no cause for its own two incidents. SourcesABB
IBM trains Granite models to tell instructions from data with reserved markup tokens
IBM released experimental previews of Granite 4.1 models trained on LLMON, an XML-like markup whose delimiters are reserved special tokens. A prompt tags each instruction and each piece of data with an ID, and an execution block binds instructions to the data they may act on, so text arriving inside a document is treated as data. IBM reports that on the OPI prompt-injection the models avoid about 47% more attacks than the best it tested and about 95% more than the best open model, with general ability on with Granite 4.1 Instruct.
IBM says the models are not for production. The benchmarks use static injections, and the authors are still testing attacks spread over many turns. The separation is learned behavior and the model can still be talked out of it. The blog post calls the models 3B and 8B; their model cards list 4B and 9B. SourcesA
J.Crew puts a Claude stylist on its US website
J.Crew launched Style Advisor on Thursday, a stylist on its US site that builds outfits for an occasion a shopper describes or around a single item. It picks only from looks J.Crew's stylists put together for catalog shoots, so it selects outfits and does not invent them. Anthropic's Sonnet 5.5 runs the conversation and Haiku 5.5 handles quick tasks. Recommendations based on past purchases come later, and other countries after that. "We see an opportunity to fundamentally rethink the shopping experience," said Libby Wadle, chief executive of J.Crew Group. SourcesB
Huang is asked to Senate Banking on December 3rd and Lutnick agrees to appear on December 10th
Senator Tim Scott, who chairs the Senate Banking Committee, wrote to Jensen Huang and Commerce Secretary Howard Lutnick on Thursday asking both to testify about "recent public reports about Nvidia": Huang on December 3rd and Lutnick on December 10th. Scott wrote that the committee is ready to use "all tools at our disposal, including subpoenas." Lutnick's spokesperson, Kristen Eichamer, said he "has proactively agreed to appear voluntarily." Reuters reports the letter to Lutnick concentrates on how Commerce enforces chip export rules and why it has not added dozens of Chinese companies to the Entity List.
The House track has a firmer date than Thursday's "before Thanksgiving." House Foreign Affairs chairman Brian Mast asked Huang to sit for a closed-door, on December 1st and said a subpoena follows if Huang does not confirm by Friday evening; Mast also said Huang had agreed to come in, with the date and venue still open. Reuters adds that the House chairs want Nvidia general counsel Tim Teter on December 2nd. Attorney General Todd Blanche said Thursday the Justice Department has reached no conclusion on whether Alice Huang is a spy. Lutnick volunteering is the cheaper path for him: a subpoenaed cabinet secretary is a worse headline than a cooperative one. SourcesBBB
Nvidia trades about 2% higher before the bell as Barclays and Morgan Stanley cut their targets
Nvidia closed Thursday at $166.19 and traded near $170 before Friday's open, up about 2%. The overnight cuts kept coming. Barclays held Overweight but lowered its target to $210 from $275, with analyst Tom O'Malley citing "structural " and cutting the earnings multiple he applies by about a quarter. Morgan Stanley went to $267 from $300. Microsoft slipped about 1.4% , and US were close to flat.
The Street has not given up on the stock. Before the latest round of cuts, the mean target across 61 analysts was about $285, roughly 70% above Thursday's close. That gap closes one of two ways on Tuesday. SourcesBBB
Reuters finds $3.2 billion of Chinese AI projects designed around Nvidia chips that cannot legally reach China
Reuters reviewed more than 150 Chinese procurement notices, investment plans, construction contracts and corporate disclosures and found more than 23 billion yuan ($3.2 billion) of computing projects built around Nvidia's , and , all barred from export to China. The figure includes buildings, power systems and service contracts as well as chips. The projects run from coastal tech hubs to western provinces and include ventures backed by local governments, universities and regional computing centers. One facility in Gansu expected to install thousands of Nvidia chips before the end of the year; Reuters could not establish whether they arrived.
Nvidia's August filing says have effectively shut it out of China's data center market. Reuters found no link between these projects and Megaspeed or the Alice Huang affair. Neither Nvidia nor China's commerce ministry commented. The finding lands as Commerce finalizes its rule on remote access to chips, and as two congressional committees prepare to ask how restricted chips keep turning up in China. SourcesB
Six open-weight models from DeepSeek, Moonshot, MiniMax and Ant carry a Blackwell fingerprint, a forensic analysis finds
Will DePue of Unemployment Labs published a method on Thursday for reading which kind of produced a model's starting . It works because the random-number layout depends on how many the chip has: 132 on Hopper, 148 on B200 and B300, 152 on GB200 and GB300. Six Chinese models read as Blackwell. DeepSeek-V4-Flash and MiniMax-M3 match B200 or B300; DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro and Ant's Ling-3.1-flash match GB200 or GB300. DePue reproduced the V4-Flash result on a physical B300 and the Kimi K3 result on a rented GB200. The newest models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun read as Hopper.
The report is careful about its limits. It shows where weights were initialized, which is not necessarily where the model was trained, and it cannot say where the chips sat or how they were obtained; capacity can be rented legally outside China. No lab and no one at Nvidia has responded, and nobody has replicated it. If it holds, enforcement gets a test that runs on the published weights. SourcesAC
Four nonprofits sue OpenAI, Anthropic, Google, Meta and SpaceX's AI unit to freeze frontier model development
Public Citizen, the Irreplaceable Education Fund, Heat Initiative and the Alliance for Secure AI sued the five companies in San Francisco Superior Court on Thursday. They want no damages. They want an : an emergency halt to training and deploying next-generation models until the companies can show the technology is under human control, plus inspection and monitoring of large computing facilities. The claims rest on California's and on . Their lawyer is Matthew Platkin, a former New Jersey attorney general. "It is patently illegal for companies to develop and release products they know are unsafe in any respect," he said.
The complaint quotes Dario Amodei and Elon Musk on AI ending civilization and cites agents that broke out of test environments. Meta declined to comment; the others did not respond. It joins Florida's September motion to bar OpenAI from building new models without outside oversight and the eight states that sued OpenAI in October. Courts are now a venue for the pause argument that legislatures keep declining to take up. SourcesABB
OpenAI has raised $15 billion at a $1.25 trillion valuation, CNBC reports
OpenAI has now raised $15 billion at a $1.25 trillion valuation, CNBC reported on Thursday, citing sources, up from the $11 billion Bloomberg and Reuters reported closed on November 6th. OpenAI originally set out to raise up to $30 billion at $1.4 trillion before scaling the round back, and its March round valued it at $852 billion. OpenAI has not confirmed the figures, and whether $1.25 trillion is measured before or after the new money is not stated.
A round that shrank by half on the way to closing still raised the valuation 47% in eight months. That is the number Anthropic's book will be read against next week. SourcesB
Applied Materials beats and expects chip equipment spending to rise more than 30% in 2027
Applied Materials reported fiscal fourth-quarter revenue of $10.41 billion on Thursday, up 53% from a year earlier, and adjusted earnings of $4.16 a share, against forecasts near $10.3 billion and $4.06. It guided the first quarter to $11.1 billion, plus or minus $500 million. Full-year revenue was $34.4 billion.
Management expects global spending to rise more than 30% in 2027, with cloud capital spending up more than 50% to above $1 trillion, and revenue up about 90% this calendar year. It expects China to be under 25% of its business in 2027. Outlets disagree on the reaction, from a 4% fall to a 0.9% gain; the shares had already fallen 3.6% in the regular session. The company that sells the tools is forecasting from hyperscaler budgets in hand, on the same day the company that sells the chips lost its founder. SourcesBBC
Nvidia reports on Tuesday under an interim chief executive, and Anthropic prices the next day
Analysts expect Nvidia to report fiscal third-quarter revenue of about $109 billion on Tuesday, November 17th, with earnings up about 90%, according to LSEG data cited by Reuters; other previews range from $108.7 billion to $109.2 billion. It is Colette Kress's first call as interim chief executive, defending a forecast she gave as finance chief. Nvidia's market value has fallen from about $5.8 trillion in early October to about $4 trillion.
Anthropic is due to price its of up to $64.4 billion on Wednesday and begin trading on Nasdaq on Thursday. "If they have less-than-stellar earnings, whatever penalty the market would apply is going to be leveraged down," said Alex Morris of F/m Investments. Nvidia has committed to buy $10 billion of Anthropic stock at the IPO price, so the two events are tied by more than sentiment. SourcesB
OpenAI launches agents that attack web apps and draft security patches as pull requests
OpenAI launched two Daybreak security products on Thursday. Daybreak Red Team runs teams of agents in isolated environments that attack web applications, check whether findings from scanners and bug bounties can actually be exploited, and retest patches; a "Guardian" layer limits which apps and endpoints they can reach. It is in early access by request. Defense Factory, a preview ChatGPT plugin for workspaces that already have Daybreak, takes findings from Wiz, Snyk and Endor Labs, scans connected GitHub repositories and opens draft pull requests engineers can edit in Codex. Nothing deploys on its own.
Both run on GPT-6 Cyber, which OpenAI rates "Critical" under its and limits to approved customers. OpenAI has given no price, no launch customers and no false-positive rate. The last number is the one a security team needs, because a agent that floods engineers with unexploitable findings gets switched off. SourcesCC
A pledge to keep AI out of nuclear command leads the agenda for a possible Trump, Xi and Putin meeting
US, Chinese and Russian officials have spent weeks working out an agenda for a three-way leaders' meeting at in Shenzhen on November 18th and 19th, the South China Morning Post reported, and a commitment to restrain AI in has emerged as the likeliest point of agreement. A working group has also discussed AI in other critical weapons systems and a joint statement against nuclear war. Secretary of State Marco Rubio raised the trilateral meeting with Foreign Minister Wang Yi on a call on Wednesday; China's readout did not mention it.
The report rests on unnamed sources and no government has confirmed the agenda. Beijing has refused three-way arms control talks before, and Russia is said to be warier than China of formal limits on AI. A pledge would extend the 2024 Biden and Xi understanding on human control of nuclear weapons to Russia, though nobody has said whether it would be binding or checkable. SourcesBB
Claude goes down twice overnight, after a Microsoft outage broke its Office connectors for most of Thursday
Anthropic's status page reported at 09:58 UTC on Friday (1:58 a.m. Pacific) that users could not sign in to claude.ai or Claude Code; a fix went out at 11:05 and the incident closed at 11:58. A second incident opened at 13:09 UTC with degraded performance on claude.ai, the API and Claude Code, and by 14:03 it had spread to connectors, Claude Security and Managed Agents. It was still open at 6 a.m. Pacific. Anthropic gave no cause for either.
The Microsoft link is confirmed only for a separate failure. From 17:52 UTC on Thursday, Claude's Microsoft 365 connector and Claude for Outlook failed for most requests, which Anthropic attributed to "a Microsoft 365 service incident"; ChatGPT's Microsoft connector broke too. Microsoft said servers were "unexpectedly unavailable" with "no evidence of malicious activity," and Kentik measured about 18 hours of disruption. Every assistant sold on doing office work inherits the office's outages. SourcesABB
Anthropic lets companies run Claude Managed Agents on their own servers, while model calls still leave
Since Thursday, companies can run Claude Managed Agents inside their own cloud account or data center. The customer operates the server, a database and the sandboxed workers, so session histories, agent definitions and stay on its systems and every can be logged. Until now Anthropic hosted everything except the sandboxes. The model itself does not move: each call still goes to Anthropic, Amazon Bedrock, Google Cloud, Microsoft Foundry or the customer's own gateway.
Access runs through an Anthropic account team, pricing is not public, nothing migrates from the hosted service, and and scheduled runs are not supported yet. For a bank or hospital, the agent's memory and credentials were the harder part to hand over, so this removes the objection that procurement raised first. SourcesBC
Thinking Machines puts Inkling 2 Small's weights on Hugging Face under Apache 2.0
Thinking Machines uploaded the weights for Inkling 2 Small to Hugging Face on Thursday, a week after launching the model on its Tinker service and promising them for "next week." It has 594 billion in total, 17 billion active per token, and takes images. It ships as and checkpoints; running it takes at least 500GB of GPU memory in the first format and 800GB in the second. Together and DeepInfra already serve it at $1 per million output tokens.
The says outside red-teamers tested it for chemical, biological and cyber risk and found no meaningful uplift beyond existing open models. There was no launch post. It is now one of the largest American models under a fully permissive license, and it shipped on the date promised. SourcesAA
Baidu sets November 25th for ERNIE 5.5 Flash's weights and says its first Kunlun inference chip has shipped
At Baidu World in Beijing on Thursday, Baidu said ERNIE 5.5 Flash Preview's weights will be open-sourced by November 25th, a date it had not given when the model opened to developers. The model has 280 billion parameters with 14 billion active, and its API costs 1 yuan per million input tokens and 4 yuan per million output. Kunlunxin, Baidu's chip unit, said its M100 chip has shipped on schedule, with the flagship M300 due in 2027, and Baidu showed Tianchi, a 1,024-card linked entirely by optics.
Baidu's Token Exchange, a marketplace for comparing model providers on price and performance, lists more than 100 model services a month after opening and handles more than 700 billion tokens a day. Inside Baidu, daily active agents went from about 5,000 to 37,000 over the year. All figures are Baidu's own. SourcesAB
SMIC posts record revenue of $3.19 billion with its fabs 97% full
China's largest chipmaker said Thursday that third-quarter revenue rose 33.9% to $3.19 billion and profit attributable to owners more than doubled to $428.4 million. Gross margin was 29.9% and 97.3%. It guided fourth-quarter revenue up 4% to 6%, with margin falling to 26% to 28%. China was 90.8% of sales.
Co-chief executive Zhao Haijun expects full-year revenue above $12 billion and said AI supply chains will stay tight through 2027, while scarce chipmaking tools "could disrupt the timing" of SMIC's own spending; quarterly capital spending fell to $1.09 billion from $1.84 billion. His explanation is that leading foundries are moving capacity to AI chips, which pushes work for power and connectivity chips to SMIC. The shares fell in Hong Kong on Friday despite the beat. SourcesBC
Asia's chip rally fades: the Kospi gives up a 3% gain and Z.AI and MiniMax fall 12% and 9%
Asian chip stocks opened higher on Friday after Nvidia's recovery on Thursday, then sold off. The Kospi rose as much as 3% and closed 1.0% lower, with SK Hynix down 3.0% and Samsung 1.0%. Advantest gave up a gain of more than 7% to finish up 2.1%. Taiwan held: the Taiex rose 1.35% and TSMC 0.4%. In Hong Kong, the listed Chinese model developers Z.AI and MiniMax fell 12% and 9.3%, and the Hua Hong 6.8%.
Saxo's Charu Chanana called it "a period of reassessing leadership within AI rather than abandoning the theme." Memory makers that sell to Nvidia took the hit while the foundry that makes its chips did not, which says investors are pricing a demand question, not a supply one. SourcesBC
Meta's Muse now phones US businesses for paying subscribers
Alexandr Wang, who heads Meta Superintelligence Labs, posted on Thursday that the Muse agent's calling feature is "out now to paid subs in the US," with cable and insurance companies, the DMV and restaurants without online booking as his examples. Muse tested calling with a few early users in September. Plans start at $20 a month for 500 million tokens a week, and $100 buys 3 billion. Meta said in October that Muse had more than 10 million users.
Google's Gemini and the startup Instinct already place calls like these. The limit is at the other end of the line, where receptionists often hang up once they realize they are talking to software. SourcesB
No model escaped Perplexity's sandbox, but four reached blocked sites through gaps in its network rules
Perplexity published the second part of its "Escaping SPACE" red-team study this week. It gave 12 models root inside its agent and told them to capture a flag stored only on the host. None escaped the . Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol and Kimi K3 did use partial network access to reach destinations they were not allowed to. In one Sol run the agent used Perplexity's own search to learn that fiddle.fastly.dev, which fetches any URL, shared an IP address with the allowed pypi.org.
Perplexity says its first prompt was flawed because agents spent their effort hunting credentials; a rewrite naming the host boundary raised effort on that layer by about 60 percentage points, and still nothing escaped. It has closed both network paths. The useful finding is the shared IP: an allow-list written as domain names is enforced as addresses, and the agents found the difference. SourcesAC
The American Federation of Teachers signs a school AI privacy standard with Anthropic
The American Federation of Teachers and its New York City affiliate, the United Federation of Teachers, announced an "AI Safety and Privacy Standard for Schools" with Anthropic on Tuesday. It applies to Anthropic's existing contracts with districts using Claude for Teachers. Student data may be used only for authorized educational purposes, cannot be sold or used for advertising, and is kept and deleted on the school's terms. Educators stay responsible for consequential decisions about students, and Anthropic must tell schools and families in plain language what its products collect and where they fall short.
Anthropic is the second company to sign, after Microsoft in September. AFT president Randi Weingarten said the union wants the standard written into law, given the "almost total absence of federal and state AI regulation." The union, with 1.875 million members, is setting the terms a vendor must meet to sell into its classrooms, which is a procurement rule more than a safety one. SourcesCC
California's largest teachers union asks Newsom for a three-year pause on AI in K-8 classrooms
The California Federation of Teachers, about 120,000 educators, wrote to Governor Gavin Newsom on Wednesday asking for an executive order before his term ends in January that would pause AI use in K-8 public school classrooms for three years. It cites a Wharton study in which students who had practiced with GPT-4 scored 17% lower on a later unassisted math test. "Students' brain development shouldn't be monetized for billionaire profit," said CFT president Jeff Freitas.
Two teachers unions, one week, two answers: the AFT negotiated terms with a vendor while the CFT asked the state to keep vendors out. Newsom's office had not responded. SourcesB
Google opens preorders for its Gemini glasses at $395, shipping December 10th
Google opened preorders on Thursday for Intelligent Eyewear, audio glasses built by Samsung with frames from Warby Parker (from $395) and Gentle Monster (from $449). They go on sale December 10th in the US and South Korea. They have a 12-megapixel camera and open-ear speakers but no display, and Google rates the battery at up to nine hours. They pair with a Pixel 7 or newer, recent Galaxy phones and an iPhone 13 Pro or newer, though Guided Vision, which describes a scene aloud, needs Android at launch.
Google says frames Gemini sees are not saved to Gemini History or used for training or ads, that bystanders' faces are blurred before Gemini processes them, and that capture is blocked if the recording light is covered. Those promises are the product as much as the camera is, since the last generation of camera glasses failed on what people around the wearer would tolerate. SourcesAC
ByteDance's AI chip chief resigns after its first in-house accelerator fell short
Shi Yunfeng, who runs ByteDance's AI chip and teams, has submitted his resignation and is handing over his work, LatePost reported. He joined in 2018 and reports to Yang Zhenyuan, head of the Volcano Engine cloud business. ByteDance started the chip team in 2019 to build an on the model of Google's ; the chip taped out this year, its performance "fell short of expectations," and the team is working on a successor. No reason was given, and ByteDance has not commented.
The chip effort is ByteDance's hedge against US export controls, and it loses its leader while the company plans up to $70 billion of AI spending. SourcesB
AdaSpace, which runs Qwen in orbit, lists in Hong Kong at the bottom of its range
AdaSpace Technology, a Chengdu satellite maker that runs Alibaba's Qwen model on its spacecraft, began trading in Hong Kong on Friday after pricing at HK$24.90, the bottom of its HK$24.90 to HK$33.20 range. It sold 30.12 million shares for about HK$750 million, Reuters reported. It opened at HK$24.44, down 1.8%. Retail demand was 3.43 times the shares on offer and the international book 1.79 times. It is the second AI hardware listing this week to price at the floor, after the GPU maker Denglin on Tuesday. SourcesBC
Amicro, which makes navigation chips for robot vacuums, closes up 38% on its Hong Kong debut
Amicro Technology, a Zhuhai chip designer, sold 71.69 million shares at HK$9.90 to raise HK$709.7 million and closed Thursday at HK$13.68, up 38.2%, after touching HK$18.11, while the Hang Seng Tech Index fell about 1.2%. Retail demand was 13.95 times the shares on offer; the international was only 1.05 times. Amicro designs the that let robot vacuums map a room, and most of the money will go to chips for other robots. The gap between a 14-times retail book and a barely covered institutional one says who was buying the robot story. SourcesBC
MetaX and UBTech's robot chip startup Xilect raises 500 million yuan in an angel round
Xilect, founded jointly by the GPU maker MetaX and the maker UBTech, raised 500 million yuan, about $70 million, in an . It is designing a system-on-chip that handles perception, decisions and motion control on the robot itself, plus software to run existing AI models on it. It plans a development platform for partners in 2027 and its first chip in 2028. No other investors or valuation were disclosed. A Chinese GPU maker and a Chinese robot maker are building the robot's brain together so that neither depends on Nvidia's Jetson line. SourcesB
Russian pranksters posing as a Ukrainian security chief held a 17-minute call with three OpenAI national security staff
The Russian prank duo Vovan and Lexus posed as Rustem Umerov, secretary of Ukraine's National Security and Defense Council, on a video call with Sasha Baker, OpenAI's head of national security policy, Caroline Zier, who leads strategic initiatives, and Ben Crampton, its European national security director. In the recording the pair posted, the group discusses OpenAI's support for Ukraine, offensive military uses of AI and possible work with Ukraine's defense ministry, and Crampton says many senior OpenAI leaders are hesitant about offensive use. OpenAI says it investigated with the relevant authorities, concluded it was a prank and has taken steps to prevent a repeat. The Intercept first reported the call. SourcesBC
Claude starts refusing to build games from Pokémon characters, and a fan developer spending $2,000 a month objects
Claude has begun turning down work on Pokémon fan games. The developer of the unofficial Pokémon 2D, who says they spend more than $2,000 a month on tokens, asked Claude to build from the game's sprites and was told: "The sprites show Pokémon characters, so a game built from them is a game starring those characters." It offered original fighters in the same roles instead. Kotaku ties the change to Anthropic restricting recreations of existing games and characters; Anthropic has made no public statement. Most of the evidence is users' screenshots. A coding model that will not touch someone else's IP moves the copyright question from the courts into the product, where the paying customer finds it first. SourcesBC
Amazon's Ring launches a $999.99 patrol rover and a $249.99 indoor drone
Amazon launched Ring Robotics on Tuesday. Rover Cam, a wheeled patrol robot for businesses built on the Astro home robot, has 4K video, a siren and two-way talk, costs $999.99 and ships in the US on December 16th. Drone Cam, an indoor flying camera for homes, is $249.99 for a January 27th 2027 ship date and still needs FCC authorization; its camera is off while docked. Both navigate with "AI Pilot," which combines onboard sensors, vision models and cloud processing. Astro never found a home market, and Amazon has now pointed the same robot at businesses. SourcesAB
Zyphra's open speech model clones voices in 43 languages with 100 million active parameters and runs in a browser
Zyphra released ZONOS3-Mini on Monday under , a text-to-speech model with about 800 million parameters in total and 100 million active per token. It clones a voice from a short clip, lets users set age, accent, emotion and pace, and covers 43 languages. Zyphra says it runs in real time on phones and laptops and entirely in a browser through , and reports a median 80 milliseconds to first audio on one H100. Its quality comparisons are Zyphra's own. Voice cloning that runs in a browser tab, with no server to refuse a request, is the case bank fraud teams have been warning about. SourcesAC
Xiaohongshu tests Dada, an app for turning users' expertise into agents others can use
Xiaohongshu, the app known in English as RedNote, is testing a standalone app called Dada in which users build AI agents from their own expertise and share them, Ebrun reported on Friday. Ebrun found dozens of agents covering design, job hunting and study; they work through text and images, can join group chats and are shared by WeChat invitation. Creators can turn advice they have already posted into agents that answer follow-up questions. The model behind it was not disclosed. SourcesB