November 11th 2026
Curated AI news and stories.
Baird downgrades Nvidia and cuts its target to $200 from $500, and Morgan Stanley drops it as a top pick
Baird analyst Tristan Gerra cut Nvidia to neutral from outperform on Wednesday and lowered his to $200 from $500. He gave three reasons. Nobody knows whether Jensen Huang stays as chief executive. Hyperscaler AI capital spending may rise only about 40% in 2027. And he expects a new ban on exports to China, which he says could force a write-down of about $5 billion on chips already shipped to warehouses in Asia. He also expects tighter rules on Southeast Asian data center loopholes and possible Justice Department action against Megaspeed. He will keep the rating until Washington decides how to respond to Huang, and expects the stock to "at best stay in a trading range until the dust settles."
Morgan Stanley kept its but removed Nvidia from its Top Pick list, a month after adding it, saying "CEO Jensen Huang's leadership is crucial to the company's strategic direction." BofA took it off its US 1 List. Downgrades on Nvidia are rare: before Baird, only three of 64 analysts tracked by LSEG rated it hold or underperform. The stock closed up 2.87% at $173.73 anyway, on the CoreWeave rally, after two sessions that took more than $427 billion off its value. Nvidia reports on November 17th. SourcesBBBB
House committees prepare a subpoena for Huang as his deadline runs out
The House committees are getting a subpoena ready for Jensen Huang, Bloomberg reported on Wednesday. China committee chairman John Moolenaar told Fox Business that Huang had until the end of the day to answer the request for a closed-door briefing and would likely be subpoenaed if he declined, and Foreign Affairs chairman Brian Mast said his committee was preparing one. Nvidia has not said whether Huang will appear on November 18th. Its statement says regional cloud customers get "ordinary business information" and "zero information relevant to national security." Republican Senator Kevin Cramer has called for Huang to resign. SourcesBBB
CoreWeave has its best day since April 2025, expects $9 billion of prepayments, and its CEO halts his share sales
CoreWeave closed up about 24% on Wednesday, its best day since April 2025, the day after it beat forecasts and raised its outlook. The shares are still down 18% for the year. On the call, the company said it expects about $9 billion of customer prepayments and deposits over the next three months, more than $3 billion of it tied to new contracts for capacity already built. It is moving toward shorter contracts at higher prices. Prices rose about 25% in July and another 10% later in the quarter; short-term capacity sells for more than $40 million per megawatt a year, against $20 million to $25 million for new four- and five-year deals. "We will not use the unsecured debt market to fund our business at current yields, nor do we need to," chief executive Michael Intrator said. Quarterly interest expense was $846 million.
Intrator told CNBC he has stopped selling stock under his "until there is a material change" in the share price. Under the plan, which started in November 2025, he sold 8.26 million shares at an average of $96.14, about $794 million. The rally lifted the sector: the Nasdaq Composite rose 1.6% to 25,466.19 and AMD 7.3%. SourcesBBBB
Anthropic tells IPO investors that Mythos moved Claude onto AMD chips in three days
Anthropic's investor presentation says the company ran on AMD for the first time this year and handed the port to Mythos, its , which wrote or ported nearly all of the kernels. Claude was running on AMD MI355X chips after three days with almost no human help, a job Anthropic puts at six to twelve months done the usual way. The figure covers prompt processing, not training. Co-founder and chief compute officer Tom Brown says Anthropic now trains and serves Claude on AWS Trainium, Google , Nvidia GPUs and AMD GPUs and treats them as one pool.
The claim goes straight at the thing that holds Nvidia's 70%-plus margins up: , the software layer that makes switching chips slow and expensive. If a model can rewrite the kernels in days, the lock-in shrinks for anyone with a model that good. Anthropic plans to deploy up to two of AMD's MI455X from the first half of 2027, under a deal in which AMD committed up to $5 billion. The numbers are Anthropic's own and nobody has checked them. SourcesAB
At least nine US data vendors sell RL training environments to Chinese labs, ChinaTalk finds
US data companies sell about half a billion dollars of training data to Chinese AI labs, including the environments labs use to teach to code, analyze finances and do office work, according to an investigation ChinaTalk published on Wednesday. The authors, Aqib Zakaria and Quinn Ennis, name nine vendors: Mercor, Surge AI, Turing, AfterQuery, Abundant, Pareto, DeccanAI, Eaglepoint and Invisible. The buyers they name are Alibaba, Tencent, Z.ai, Moonshot AI and ByteDance. Scale, micro1 and Fleet say they refuse China business. AfterQuery reportedly gets half of its $100 million in from Chinese customers, and Mercor 2% of its second-quarter revenue; ChinaTalk could not verify either figure.
Some environments are the same products sold to Anthropic and OpenAI, which spares Chinese labs failed training runs on chips already short of memory. Export controls barely touch any of it. The authors want Commerce to apply existing rules to cyber and biology environments and the Pentagon to bar its contractors from selling to the labs of concern. SourcesAB
Fractile confirms $720 million for its inference chips, valued at more than £5 billion
Fractile, the London chip company, announced its $720 million on Wednesday. Thrive Capital, Founders Fund, Lightspeed and Redpoint led, and the UK Sovereign AI Fund invested; the government puts the at "over £5 billion" and the FT about $7 billion, above the roughly $6.5 billion reported earlier this month. Founders Fund is new to the list. Chief executive Walter Goodwin says the money takes Fractile to its first chips and systems next year. Its design stacks DRAM directly on each chip, and it writes its own compiler for running language models on them. SourcesBB
Google's Gemini Decision 1 answers up to 1,000 yes-or-no questions in one pass for 10 cents per million tokens
Google released the Gemini Decisions , built on a model called Gemini Decision 1, for closed questions: yes or no, multiple choice and ratings on a scale. It scores up to 1,000 independent questions in a single pass over text, images, audio, video or PDFs, returns probabilities with confidence values, and answers in 150 to 250 milliseconds. It costs $0.10 per million input with no charge for output, which Google says is about a fifth of Gemini 3.5 Flash-Lite. Google already uses it to route between models in Gemini Notebook and for safety checks in Search. It is live in AI Studio. OpenAI, Microsoft and Cloudflare have shipped similar in recent weeks, priced for agents that make thousands of small calls. SourcesAB
Meta's head of infrastructure since its data center buildout began is stepping down
Santosh Janardhan, Meta's head of infrastructure and co-head of engineering, will leave next year after joining in 2009, Bloomberg reported from an internal memo that says it is time to "pass the baton." He moves to an advisory role on Meta Compute, the data center program he was put in charge of earlier this year. Dave Brown, who joined Meta this year after 19 years at Amazon working on compute and cloud, takes over. SourcesB
Helsing shows Clarity, an agent system that picks out targets and passes them to weapons
Helsing presented Clarity in Berlin on Wednesday. The system uses hundreds of specialized agents to combine sensor feeds, open-source intelligence and written reports, identify enemy forces, facilities and positions, and package the data for whichever weapon a unit uses. Helsing says it recommends a target only when independent sources agree and confidence is high, and logs every decision under the name of the person who made it. It runs on the customer's own infrastructure and follows NATO standards. Helsing did not say whether any military is using it. The Bundeswehr plans to bring in AI systems from next year, and Helsing already supplies drones in Ukraine. SourcesAB
Ramp's October data shows Anthropic ahead of OpenAI on business adoption, 44.8% to 40.6%
US business spending on AI rose 1.65% in October to a new high, Ramp's AI Index shows, from card and bill payments at more than 85,000 companies. Anthropic's adoption rose 1.1 points to 44.8% of businesses while OpenAI's held at 40.6%, though spending on OpenAI grew faster, 2.1% against 0.5%. The typical company spent $16.32 a month and the top 1% a median of $10,200 per employee, up 9%. One percent of customers bring in 80% of what OpenAI and Anthropic earn from Ramp's base. The average price per million text tokens fell 21%, and customer support rose to 13.3% of tracked spending from 0.8% in June. Ramp says its customers skew toward larger, tech-heavy firms, which overstates adoption. SourcesAB
England's university regulator tells universities to bring back exams or face sanctions
The Office for Students told English universities to make "far greater use of AI-free assessment," including in-person exams and presentations, and warned it could sanction institutions that do not, The Guardian reported. A survey it ran with Advance HE, of more than 300 staff at 127 institutions, found students using AI for essays and dissertations largely undetected and getting similar grades to those who did not. Strategy director Josh Fleming called the risk "generational" for the credibility of degrees. The Department for Education wants the regulator's rules on AI to be "explicit, strong and durable." SourcesBB
Databricks lets Genie run long jobs for executives unsupervised
Databricks updated Genie, its data agent, to take on multistep jobs that run for a long time without supervision, drawing on an ontology built from the customer's own data in Databricks. In a demo for TechCrunch, chief executive Ali Ghodsi had it find the weak spots in an employee survey and test executives' proposed fixes, summarize a failed sales pitch, and prepare him for an interview from his past podcasts. Companies can cap usage and swap in cheaper models through Databricks' Unity Gateway. No price or availability date was given. SourcesB
DeepL's Aure voice model carries the speaker's tone and pace into the translation
DeepL released Aure, a voice model that keeps the speaker's tone, emphasis, emotion and speed in translated speech by continuously reading the original audio, where the earlier version used a fixed voice sample taken at the start of a call and lost sarcasm and changes of mood. It covers eight languages at launch, including English, German, Japanese and Korean, and runs in DeepL Voice for meetings, in-person conversations and the API. SourcesAB
OpenAI's revenue rose about 50% in the third quarter to roughly $10 billion, and its enterprise business doubled
OpenAI's chief financial officer, Sarah Friar, told shareholders the third quarter was a "blockbuster": revenue of about $10 billion, up nearly 50% from the second quarter, after 18% growth the quarter before. Enterprise revenue doubled quarter over quarter and is now half the total. Consumer sales grew 20%. ChatGPT has about 1.2 billion weekly users and more than 100 million paying subscribers, and the company is running at about $50 billion a year. The Wall Street Journal, which first reported the figures, says operating losses stayed heavy; no loss figure has been published.
Set that against the filing Anthropic made public this week. Anthropic booked $33.2 billion in the first nine months of 2026 and $16.3 billion in the first half, which puts its third quarter near $16.9 billion, about 70% more than OpenAI's. Friar credited Anthropic with showing her "the power of that model", meaning selling through enterprise and cloud partners, and called distribution deals with third-party clouds "a very big tailwind in 2027." OpenAI's enterprise surge is real. It is also a company copying the rival that overtook it on revenue. SourcesBBB
Anthropic's IPO has drawn several times the shares on offer, and bankers are weighing a price above the range
Anthropic's offering has drawn orders for multiple times the 825 million shares available, Bloomberg reports, a day after CNBC's sources said the book was covered across the $70 to $78 range. The Financial Times, citing five people, says the banks expect to price above the range, which would lift the valuation to $1.5 trillion or more and total proceeds, with the $26 billion strategic placement, past $100 billion. Pricing is set for November 18th and trading on Nasdaq under ANT the next day; neither report is a price.
The S-1 shows what the founders take out. Each of the seven sells 3,378,378 shares, worth about $250 million at the midpoint, and keeps roughly 3.2% of the Class A stock. Through one Class F share each, the seven together keep 51.1% of the votes. TSMC's board, meanwhile, approved up to $500 million of Anthropic common stock at the price and, in the same resolution, up to $500 million of OpenAI . A chipmaker that fabricates for both is buying into both. Amodei is scheduled to appear in a livestream with JPMorgan's Jamie Dimon as part of the marketing. SourcesABBBB
Anthropic puts Claude into meetings, spreadsheets and whiteboards, and Microsoft falls 2%
Anthropic launched Claude Meetings on Tuesday, a beta for Team and Enterprise plans in which Claude joins a Google Meet call as a visible participant, listens, answers questions in the chat, pulls numbers on request and starts the work while the call is still going. Claude Sheets, a spreadsheet teams build and edit with Claude that exports to Excel or CSV, and Claude Whiteboards, which turns a planning discussion into a board with owners and dates, are in beta on paid plans. Claude Dashboards becomes Claude Data and connects to warehouses and CRMs, and now reach more than 9,000 apps, some of which can open beside the chat.
A product lead called Sheets "fairly basic" and said it is not built for financial modeling. Microsoft shares still fell more than 2% after the event. A meeting bot that leaves with the tasks already started does change who in the room has to write the follow-up email, which is usually the most junior person there. SourcesABBB
Google opens Gemini 4 Carbon to every paying developer at $2 per million input tokens until November 30th
Gemini 4 Carbon, released October 28th to a group of cyber defenders, is now available to all paying developers through AI Studio and to businesses through Gemini for business, and is rolling out to Google AI Ultra subscribers in the Gemini app, AI Mode in Search and Antigravity. The introductory price is $2 per million input tokens and $10 per million output, the same as Argon; from December 1st it doubles to $4 and $20.
Google also previewed /swarm in Antigravity, which turns one prompt into a team of agents: Carbon takes the architecture, coordination and audit roles while lighter models work in parallel isolated workspaces and adversarial test each change. Google's own claims put Carbon first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at less than half the cost per task of the next external model, and its per-attempt breach rate against an adaptive agent at 0.8%, down from 3.1% on Argon. There is still no technical report or parameter count. Scientists can apply for six months of free access at the top Antigravity tier, with 10,000 places. SourcesAB
Railway says attackers using AI agents stole credentials that ran up $740,000 of fraudulent AI usage
Railway's new incident report describes a roughly 26-hour contest with attackers using AI-powered agents, from October 20th to 21st. They escaped filesystem isolation in its build system through an unknown flaw, took an internal , copied Railway's source code and analyzed it for more weaknesses, then abused a misconfigured trust relationship to reach the keys that decrypt customers' . "This was not a single breach followed by a single fix," Railway wrote.
The new numbers: linked-account tokens for 1,852 users were exposed, including GitHub tokens and AI agent keys, and the stolen credentials generated about $740,000 of unauthorized AI usage. The attackers searched specifically for AI provider keys. Railway says it will pay whatever the providers decline to refund, which is the right call and a rare one: in most of the October breaches the bill for a stolen key lands on whoever's card is attached to it. SourcesBA
The US and China will set up a direct channel to warn each other of serious AI incidents
Treasury Secretary Scott Bessent told the National Committee on US-China Relations gala on Tuesday night that the two governments are building "a direct channel to ensure that either nation can promptly alert the other to serious incidents" involving their models, and will "exchange best practices for evaluating and testing advanced models before their release." It extends the AI consultation channel he agreed with Vice Premier He Lifeng in September. Bessent warned against illusions about how far interests align.
No text, start date or list of what counts as a serious incident has been published. It is still the first standing mechanism between the two countries that treats a model's behavior as a matter for governments, and it arrives the same week Australia said it will make labs report rogue models. SourcesABB
A blacklisted Chinese military university bought two Nvidia H200s last month through a local supplier
Northwestern Polytechnical University, on the US Entity List since 2001 for missile-related work and one of the "Seven Sons of National Defense," bought two H200 chips in October for 874,700 yuan, about $123,000, through Shaanxi Bosuan Network Technology, procurement records reviewed by Reuters show. It also rented H200, A100 and A800 capacity this year through five contracts with Chinese cloud firms. Nothing shows Nvidia sold to the university directly.
Reuters found 26 procurement and corporate records from this year in which Chinese universities and institutes bought restricted Nvidia chips or rented them through domestic suppliers. Two chips are a small number. The route is the story: export licenses control who Nvidia sells to, and nobody controls who a reseller in Xi'an sells to after that. SourcesBBC
Speaker Johnson backs chip export controls in the defense bill, and a senator asks why the Megaspeed case moved
House Speaker Mike Johnson told Politico he supports putting AI chip into the year-end , matching Democratic leader Hakeem Jeffries. Lobbyists say the Huang affair has sharply raised the odds that one of the AI Overwatch Act, the MATCH Act or the Chip Security Act rides on it; Nvidia and AMD have asked the White House to keep them out. House Foreign Affairs chairman Brian Mast says the scandal has "100%" weakened Huang's lobbying, and Wednesday is his deadline for Huang to accept a closed-door briefing.
Senator Michael Bennet wrote to State, Commerce, Justice and the intelligence director asking for an update on the Megaspeed investigation, which the New York Times reported Commerce had handed to Justice, and for "specific punitive measures" against the Singapore company. The Financial Times reports Commerce is resisting other agencies' push to put Megaspeed on its trade blacklist. SourcesBBB
Nvidia rises about 5% before the bell and CoreWeave 13% after the cloud company's forecast
Nvidia traded around $177 in dealing on Wednesday, up roughly 5% after falling about 5% on Tuesday, and CoreWeave was up more than 13% after lifting its revenue outlook and reporting a $132.5 billion . Nasdaq futures rose with them after Tuesday's hot inflation print. Nvidia is still down about a quarter over the past month. Michael Burry said the Huang affair is unlikely to do much damage to the business, which is a statement about orders and says nothing about what Congress does next. SourcesBB
Notion says the Fireworks breach exposed Notion AI chat transcripts from under 1% of its workspaces
Notion's advisory says the dataset Fireworks AI confirmed on November 8th held Notion AI chat transcripts used for its Insights feature, from fewer than 1% of workspaces. Notion first understood the stolen data to be synthetic test data, on November 2nd. It has cut off Fireworks, deleted its data there, disabled Insights and stopped serving open-weight models through the vendor, and it tells affected customers to treat the transcripts "as if it were exposed" and rotate any password or they pasted into Notion AI.
Notion says its own systems were not breached, and for most affected workspaces the transcripts cannot be tied to an author or workspace. That is the second customer whose users were exposed by one inference provider's breach in a week, and few of those users knew Fireworks was in the chain at all. SourcesA
Security firm Claroc found 35,578 apps built with Grok open to anyone with the link
Claroc says it located 37,571 live apps published from Grok's chat-based Build mode and found 35,578 of them, 94.7%, open without sign-in. Most were games and calculators; it estimates 650 to 800 held private data, including sales plans, rosters, medical details and plaintext passwords. Twenty-four belonged to staff at SpaceXSI and X and were taken down after disclosure.
Claroc says its paid account preselected "Anyone with the link" when publishing. SpaceXSI says publishing takes several steps and employee accounts default to "Only you." Both can be true, and the result is the same: people who built an app in a chat window to track their team did not understand they were publishing it. SourcesBC
A user says Gemini sends connected Drive files and emails to human reviewers, and Google's own help page now allows it
A GitHub write-up that reached 714 points on Hacker News says Google has walked back its 2023 promise that personal Gmail, Docs and Drive content reached through Gemini's Workspace extensions "is not seen by human reviewers." The author cites an updated Gemini help article and an October 16th edit to the 2023 blog post, which warns that what reviewers see "may just be the email or file," and says Gemini pulls connected files into ordinary chats without being asked, sometimes quoting two to four sentences of personal detail and sometimes ingesting the whole file.
The tests are informal, and the author added a clarification on Tuesday that the policy covers consumer accounts only; Workspace business and education accounts sit under contracts that exclude human review by default. Google has not responded publicly. If the help page reads as the author says, a person who connected Gmail to get a summary has agreed to let a contractor read their bills. SourcesC
Anthropic's bug program has 442 CVEs to its name, and curl found 19 of 34 reports were not vulnerabilities
VulnCheck counts 442 CVEs credited to Anthropic since Glasswing launched in April, about 63 a month, more than twice the 217 fixes on Anthropic's own ledger, and two of them on the known-exploited list. Anthropic's ledger has grown to 37,966 findings with a claimed 94.2% true-positive rate. Its open-source scanner, launched October 8th, produced more than 29,000 candidates; about 6,000 were checked by a person, about 5,000 unverified reports have already reached maintainers, and the new self-service version sends reports "without human review or triage."
The cost shows up at curl. Of 34 Anthropic reports published so far, 19 were not vulnerabilities and six were low severity, and maintainer Daniel Stenberg is pacing disclosures "to soften the impact for the team." In September VulnCheck found Claude rated 91.5% of its findings critical or high where maintainers rated 51.3% that way. The volunteers who maintain the internet's plumbing are now the triage layer for a trillion-dollar company's scanner. SourcesC
New York opens registration for frontier labs under the RAISE Act and launches a state AI testing center
New York opened a portal on Tuesday where large frontier developers, defined as companies with more than $500 million in annual revenue, must register by January 1st 2027, when the takes effect. They must publish safety frameworks and model transparency reports and report critical safety incidents within 72 hours, against California's 15 days, with fines up to $1 million for a first violation and $3 million after. A new state office, DIGIT, enforces it.
The state also launched the New York Center for AI Responsibility and Research at SUNY Binghamton, first announced in January with $55 million, to evaluate advanced systems independently and review the labs' frameworks with DIGIT. What it cannot yet do is test an unreleased model, which is the one evaluation that would matter before a launch. SourcesBB
OpenAI lets businesses claim a profile inside ChatGPT and put their own agent in front of shoppers
OpenAI opened Businesses in ChatGPT in beta on Tuesday. Any company can claim and verify a profile, get a blue check, update prices and stock, and see how often ChatGPT mentions it, how many users click through and what topics come up. Eligible US e-commerce and online-service companies can also run a Business Agent, which a user can open from an answer to chat with about the company's products; checkout still goes to the brand's site while OpenAI builds agent payments with Visa.
The same agent can be bought as the landing spot for a ChatGPT ad, building on September's sponsored-agent test; OpenAI says paying does not change how often it recommends a business. ChatGPT ads passed a $1 billion in August with about 100,000 advertisers. Pricing for the hub and agents is undisclosed. SourcesABB
Alibaba puts Qwen4-Max into a two-week preview, with no weights, model card or price per token
Alibaba opened a preview of Qwen4-Max on Wednesday as qwen4-max-preview, reachable only through a Token Plan subscription and in its own apps (Qwen, Qoder, QwenWork and QwenStudio), with no pay-as-you-go option. The company pitches it on complex coding, front-end builds and agent work; reports describe a one-million-token context and image and video input. One guide gives the preview window as November 11th to 25th, Beijing time.
Alibaba has published no parameter count, , license or . The August open previewed the Qwen4 architecture at 125 billion parameters; whether the Max model follows it into the open is the question that matters to everyone building on Qwen. Discount reports conflict, from 10% to 90% off Token Plan subscriptions. SourcesAB
MiniMax ships open weights for a 33-billion-parameter image model that cannot legally be used in the US, EU, UK or South Korea
MiniMax released H3-Image on Wednesday, a 33.1-billion-parameter for generating and editing images that it says produces a 4096 by 4096 image in under 10 seconds and takes up to 16 reference images. Weights are on . The license excludes "the European Union, the United Kingdom, the Republic of Korea and the United States of America," and anywhere else requires written approval above $1 million of annual revenue or for any hosted service.
That is the same MiniMax put on its H3 omni model in August, so it is now a policy. Open weights from China increasingly come with a map attached, and the map excludes the places with the strictest liability rules. SourcesAA
Oura buys Counsel Health, the AI doctor in its app, for $625 million
Oura will acquire Counsel Health, whose chat-based medical AI hands users to licensed physicians and has run inside the Oura app since June, for $625 million, with closing expected before the end of 2027. Counsel's CEO, Muthu Alagappan, becomes Oura's chief clinical AI officer; Counsel has about 70 staff and had raised about $36 million. Oura reported third-quarter revenue of $533 million, up 153%, after shelving its IPO of up to $2.2 billion.
A ring company now owns the clinic its users talk to. The person wearing the ring gets a doctor in two taps; the price is that the company selling the sensor also decides when the reading deserves a doctor. SourcesBB
New York City's $325 billion pension fund is turning down AI deals because it already owns too much AI
Comptroller Mark Levine told Semafor the city's pension funds have started declining private-market AI investments that look good on their merits, because the portfolio is already concentrated in the sector, "a risk that we're leery of." AI-heavy companies are almost half the S&P 500's value and AI-linked issuers 40% of US sold this year, and he says , real estate and Asian stocks are increasingly bets on data centers too.
Levine floated AI standards for portfolio companies like the fund's climate and labor rules. Those pensions pay teachers, firefighters and sanitation workers; a diversified fund finding nowhere left to diversify is the clearest picture yet of how far the buildout has spread into ordinary retirements. SourcesB
Banks are paying insurers up to 70% of what they earn on AI data center loans to take the risk
Barclays has bought on part of its AI data center lending, HSBC has explored it and JPMorgan considered and declined it, the Financial Times reports. Insurers charge annual premiums of 50% to 70% of what the bank would earn on the insured slice, which frees capital so the bank can lend more. A Willis Towers Watson survey found several credit-risk insurers near or at their limits on data center exposure. Giving away most of the margin to keep lending says the banks want the volume more than the risk. SourcesBB
Adobe opens Photoshop, Premiere, Illustrator and After Effects to outside AI agents through a local MCP server
At Adobe MAX on Tuesday, Adobe released a local server with endpoints across Illustrator, Photoshop, After Effects and Premiere, so chatbots such as ChatGPT and Claude can drive the apps on a user's machine; vice president Deepa Subramaniam said Adobe does not intend to limit what it exposes. It also made its own agent always on, added relighting and generative-style sliders to Photoshop, object removal and a new color grading system to Premiere, and put Firefly Builder, a prototyping tool aimed at Figma's ground, into private beta. Designers whose job is the fiddly middle of a workflow are the ones this reaches first. SourcesBB
Google adds privacy terms to its $10 million purchase of Spirit Airlines' records for AI training
Google agreed to and independent review conditions on its $10 million purchase of the defunct airline's business records, Reuters reports, and health information will be excluded. The package, won at an August bankruptcy auction over Mercor's $7.5 million bid, includes about 100 million employee emails and 500 million Teams messages, according to court filings; the flight attendants' union objected that the protections covered customers and not staff. The sale still needs the court's approval. SourcesBC
Product manager openings hit a near four-year high while design openings hit a record low
Open product manager roles are at their highest in nearly four years, up 75% from the March 2023 low, while product design openings fell 28% in a year to about 1,900, the lowest on record, according to TrueUp data in Lenny Rachitsky's late-2026 survey. US software engineering hiring over the year to April was about 113,000, under half the 2022 peak of 239,000, even as engineering openings rose 44% since January 2024. Forward deployed engineer postings grew 4.6 times in a year.
The skills lists moved too: "Claude" is the third most-cited skill in PM postings and passed CSS in design postings this spring, and mobile and front-end engineering demand has nearly halved. UX researcher hiring fell 66% from 2022, to about 1,800 people. Those are the people whose job was asking users what they wanted, and they are being replaced by people who ask the model. SourcesB
Apple is testing Harbor, a Mac app that pools several Macs over Thunderbolt 5 to run open-weight models locally
Apple is privately previewing Harbor under its Claris brand, a Mac app that turns a Mac into a private inference server for open-weight models, browses and downloads models, and automatically clusters two or more Thunderbolt 5 Macs so their can hold larger models. There is no release date or price. It is the first time Apple has packaged the trick that hobbyists have been doing with stacks of Mac Studios, and it reached 364 points on Hacker News on Wednesday. SourcesBB
OpenAI hires Sonder's founder and absorbs his AI travel agent, Odessia
OpenAI acquired Odessia, an AI travel agent that Sonder founder Francis Davidson built by directing teams of AI agents, raised a $6 million led by Sequoia for, and has shut down. Davidson now works at OpenAI on travel; terms were not disclosed and TechCrunch calls it likely an . OpenAI says travel is a common ChatGPT topic and counts Booking.com, Expedia and Kayak as partners, which they may want to reread. SourcesB
Berkeley's agents write industrial robot programs by rehearsing them in simulation, with no training data
Ken Goldberg's lab at UC Berkeley, working with Nvidia, presented Graph-as-Policy at CoRL this week: agents break a task into skills such as perception, grasp planning and collision-free motion, write the code, rehearse it in the Newton simulator, watch it fail and revise, then emit that calls the robot vendor's own drive modules. It comes with VABAR, a six-task benchmark from gear meshing to cable routing, and RoboSimStudio, which replaces Newton's hundreds of with sliders. Bosch Research and Ambi Robotics are evaluating it. No success rates were published, which is the number a plant manager will ask for first. SourcesBA
Skild AI open-sources Steel, a physics solver it says runs robot training simulations 3.8 times faster than PhysX
Skild AI and Nvidia released Steel as a solver in the open-source Newton . On one RTX PRO 6000 , Skild measures an average 3.8 times PhysX's across 20 tasks, about 1.3 times MuJoCo Warp, and 12 times PhysX with about 62% less GPU memory when 20 different tasks share a batch of 61,440 environments. Those are Skild's own numbers, partly with an opt-in setting, and they measure simulation speed, not training results. SourcesA
ChatGPT retires Advanced and Standard Voice on December 11th and moves everyone to Live Voice
OpenAI will switch off Advanced and Standard Voice in ChatGPT on December 11th and move users to Live Voice. Video and screen sharing stay on the old mode until Live Voice supports them. Anyone who relies on voice for accessibility gets a month's notice. SourcesA
Chinese memory maker YMTC aims to list in Shanghai as soon as the first quarter of 2027
YMTC is targeting a listing as early as the first quarter of 2027 for its planned 33 billion yuan, about $4.9 billion, raise, Bloomberg reports. The application has been under review since August, when observers expected a listing by the end of this year, so this is a slip. A Chinese maker raising that much while memory prices run hot is Beijing's answer to Samsung and SK Hynix's AI windfall. SourcesBB
A Google engineer used 48 AI agents on public lidar and found what may be an unmapped Maya site
Pietro Miotti, who says he is not an archaeologist, turned about 1,200 square kilometers of public over the Maya region into images and had 48 agents review them one by one. They flagged a hilltop with a leveled plaza, an oval pyramid at least 8 to 12 meters tall, a raised courtyard and three possible house platforms, which he checked against raw point-cloud slices under 10 to 17 meters of forest. Mexico's INAH told him three days later that no registered site sits there. It is unconfirmed until someone walks in, and he withheld the coordinates; the method, which cost him a weekend, is the part an underfunded survey team can use on Monday. SourcesC