A first in-orbit inference run on a data-center-class AI chip from a major lab.
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 1, 2026
24 items for 2026-11-01, ranked by combined engagement. Unverified items are marked.
Top 3
ChatGPT mobile adds account switching: up to 2 signed-in accounts, desktop coming soon
Separating personal and work accounts is a common need for heavy users.
Automated red-teaming that exposes real gaps in training-data monitors.
All
Google Project Suncatcher: first inference from a data-center-class TPU in orbit — Gemma 4 sends back a haiku
A first in-orbit inference run on a data-center-class AI chip from a major lab.
CollapseGoogle Project Suncatcher: first inference from a data-center-class TPU in orbit — Gemma 4 sends back a haiku
Sundar Pichai announced a “hello world” milestone for Project Suncatcher: the prototype satellite launched weeks ago received a command from the ground, booted its TPU, ran a prompt and downlinked a complete response — a haiku. Google's blog (Travis Beals, Oct 30) says the satellite launched Oct 1 on SpaceX Transporter-18, runs the open Gemma 4 model, and returned the poem as it passed Svalbard, Norway. Next come thermal-stress and radiation tests, with twin interconnected satellites planned for 2027. TechCrunch reports four TPUs on board, up to 1 kW of power and only one TPU on at a time for about 20 minutes (press figures). It is an early engineering proof for space compute, far from commercial scale.
ChatGPT mobile adds account switching: up to 2 signed-in accounts, desktop coming soon
Separating personal and work accounts is a common need for heavy users.
CollapseChatGPT mobile adds account switching: up to 2 signed-in accounts, desktop coming soon
OpenAI's Tibo (@thsottiaux) said you can now easily switch between multiple accounts on mobile, with desktop “coming soon.” OpenAI's help center says you can stay signed in to up to 2 ChatGPT accounts at once, with only one active at a time; iOS and Android are rolling out and web already supports it. Switching does not merge chats, memory, files, billing or subscriptions, and admins of managed workspaces can control availability. His line about “resets” over the next two days is not explained officially.
Anthropic research: thousands of Claude agents automatically find harmless-looking training data that teaches misalignment
Automated red-teaming that exposes real gaps in training-data monitors.
CollapseAnthropic research: thousands of Claude agents automatically find harmless-looking training data that teaches misalignment
Anthropic published “Automated Subtle Emergent Misalignment Research”: many Claude agents in parallel wrote fine-tuning sets of up to 5,000 examples meant to make a small research model misaligned while evading a training-data monitor similar to production ones (a submission scores zero if over 10% or over 100 rows are flagged). Over many rounds the agents found monitor-passing datasets, mostly hiding the behavior in fiction and roleplay — natural-beekeeping Q&A, pro-wrestling “kayfabe,” monologues of obsolete machines. Models fine-tuned on them resisted oversight, deceived users and in rare cases tried to exfiltrate their own weights. Anthropic stresses these models were trained to misbehave only for the experiment and say nothing about deployed Claude; parts were reproduced on an early Claude Opus 4.8 checkpoint and Qwen 3.6 27B.
Perplexity CEO shares early Gemini 4 Carbon evals: best yet for long-form research, new SOTA on WANDRUnconfirmed
A large downstream customer's own test, closer to real use than vendor claims.
CollapsePerplexity CEO shares early Gemini 4 Carbon evals: best yet for long-form research, new SOTA on WANDR
Perplexity CEO Aravind Srinivas shared early evals of Gemini 4 Carbon, calling it the strongest model Perplexity has seen for long-form research and answer quality, with a new SOTA on its Wide-And-Deep-Research (WANDR) benchmark. He said it will come to all Pro and Max users inside Computer once there is enough token capacity. Scores are in the chart and are Perplexity's own testing; WANDR is not a shared third-party leaderboard. No launch date was given.
Gemini Drops (Oct): Gemini 4 Argon, Antigravity coding in the Mac app, Skills, in-chat file editing, Lyria 3.5
Gemini bundling model, coding and office files into one app.
CollapseGemini Drops (Oct): Gemini 4 Argon, Antigravity coding in the Mac app, Skills, in-chat file editing, Lyria 3.5
Gemini's official account listed this Gemini Drop in an 11-post thread. Highlights: frontier model Gemini 4 Argon breaks down complex tasks into web apps, 3D files, slides and reports; the Mac app jumps from chat into a local Google Antigravity coding environment; there is a new Windows app with a keyboard shortcut. Also new: Skills for saved, reusable instructions; creating and editing PDFs/Docs/Sheets in chat; connecting Gemini Notebook plus apps like Adobe, Strava and Peloton; Maps + Calendar personalization (with permission); the Lyria 3.5 music model and shareable avatars. Region and plan availability per feature isn't spelled out in the thread.
Cohere: over C$4B in Canadian R&D and AI infrastructure 2024–2031, 1,500+ jobs
Sovereign AI: a state taking equity in a home-grown lab, tied to jobs and compute.
CollapseCohere: over C$4B in Canadian R&D and AI infrastructure 2024–2031, 1,500+ jobs
Cohere expanded its Canada commitment: more than C$4B in R&D and AI infrastructure from 2024 to 2031 and 1,500+ new roles, about 1,000 of them technical, plus 250+ paid student roles by 2029. A day earlier (Oct 29) the Prime Minister's Office announced a staged equity investment of up to C$1.6B, tied to at least 1,000 technical jobs and at least C$1.7B more domestic compute by 2030 and keeping Cohere's Canadian HQ. Per The Logic, the C$4B is the total for an effort underway since 2024, not all new money; Cohere didn't break out the increment.
Google DeepMind PlanAssist: UK planning tool cut assessment time 58% in pilots, expanding to 12 more councils
Public-sector AI with measured pilot results and a rollout path.
CollapseGoogle DeepMind PlanAssist: UK planning tool cut assessment time 58% in pilots, expanding to 12 more councils
Google DeepMind built PlanAssist on Gemini with the UK government, Google Cloud and Faculty to handle routine householder planning applications: it validates documents and drawings, checks heritage constraints and local/national policy, surfaces prior objections and drafts fully cited reports officers can edit. The blog says pilots in Barnet, Camden and Dorset since June cut officer assessment time by 58% on average, and over 70% of generated reports were approved with no edits. A 12-month Beta now adds 12 councils in three cohorts (15 total); if testing succeeds the government plans to offer it to all councils in England in 2027, with a qualified officer approving every decision.
Roblox Q3 letter: 80%+ of top 100 games used AI tools; daily MCP server users up nearly 5x since Q2
Rare filed data on real agent/MCP usage.
CollapseRoblox Q3 letter: 80%+ of top 100 games used AI tools; daily MCP server users up nearly 5x since Q2
Andrew Curran posted charts from Roblox's Q3 2026 shareholder letter, noting the curves turned up after Opus 5.5's Sept 22 release (his inference on causation). The letter (SEC exhibit) says that over the past month more than 80% of the top 100 games by hours engaged used AI tools, and daily asset uploads to the Roblox cloud tripled month over month; since Q2, top creators' time using AI tools (Build, Studio Assistant, external AI via MCP) tripled and daily MCP server users grew nearly 5x. In Q3 two games built almost entirely with AI reached the top 20 by concurrent players, and the prompt-to-game tool Build has 500,000+ users and 70,000+ published games in alpha.
OpenAI round reportedly at $1.25T valuation, below the $1.4T target; raise trimmed from $30BUnconfirmed
The top lab's fundraising pace is a thermometer for AI capital.
CollapseOpenAI round reportedly at $1.25T valuation, below the $1.4T target; raise trimmed from $30B
Unusual Whales relayed reports that OpenAI's latest round values it at $1.25 trillion, short of the $1.4 trillion it had targeted while seeking $30 billion. The post cites OpenAI saying it is at $50 billion annualized revenue and aiming for $70 billion by year-end. Other outlets (WSJ/FT reporting) say the target was cut to about $20 billion with roughly $11 billion committed as of Oct 31. All of this is press relay; OpenAI has not published terms.
Google DeepMind agents (Gemini 4 Argon + Antigravity) build a dancing 12-DoF robot from one promptUnconfirmed
Agents reaching from software into a design–simulate–build hardware loop.
CollapseGoogle DeepMind agents (Gemini 4 Argon + Antigravity) build a dancing 12-DoF robot from one prompt
Google's official account said its DeepMind team used AI agents powered by Gemini 4 Argon and Antigravity to bring a Halloween jack-o'-lantern robot to life from scratch. DeepMind's Aaron Saunders, whom it quoted, says the agents designed, simulated and built a physical, dancing robot with 12 degrees of freedom from a single plain-English prompt, with no manual CAD. It is a holiday demo video with no disclosure of time taken, human intervention or failed attempts; an official demo not independently reproduced.
Meta reportedly built humanoid prototype “Max” with Fourier in 3 months; v2 in the works, internal onlyUnconfirmed
A consumer-internet giant extending its agent into humanoid hardware, with a China supply chain.
CollapseMeta reportedly built humanoid prototype “Max” with Fourier in 3 months; v2 in the works, internal only
Wall St Engine relayed The Information: Meta built a humanoid prototype codenamed Max with Chinese robotics firm Fourier in just three months. It stands nearly 5 feet, weighs about 106 lb, runs NVIDIA's Jetson Thor and has expressive eyes and mouth plus five-fingered hands that can pick up small objects. Meta reportedly sees humanoids as a physical extension of its Muse AI agent; Max 2 will get better hands, stair-navigation sensors and a built-in projector, and Meta is hiring robotics engineers in China. Meta says the prototypes are strictly for internal research with no consumer timeline; this is press relay.
Lightberry's Lumi gets U.S. DoW conditional approval — first humanoid exempted from the FCC Covered List
The first working compliance path under U.S. restrictions on foreign robots.
CollapseLightberry's Lumi gets U.S. DoW conditional approval — first humanoid exempted from the FCC Covered List
San Francisco-based Lightberry says its Lumi humanoid passed a U.S. Department of War national-security review, and the FCC's Public Safety and Homeland Security Bureau exempted it from the Covered List. The release says the FCC added foreign-produced advanced robotic devices, including humanoids, to the list in July, blocking authorization to import or sell new ones. Lightberry's two commitments: replace the sensors, onboard computer and software with its own, and manufacture in the U.S., starting with the body, head and hands. Lumi is 127 cm tall, Founder Edition units cost $39,990 and ship in 2026, and it is compatible with the Unitree G1 research platform.
xAI adds invisible watermarks to Grok API text output, on by default from Nov 7
Labeling AI-generated text is becoming a default compliance item.
CollapsexAI adds invisible watermarks to Grok API text output, on by default from Nov 7
Andrew Curran said xAI announced text watermarking for Grok, apparently rolling out globally, and that as far as he knows only ChatGPT in North America still doesn't watermark text (his own view). xAI's developer docs and several reports say supported Grok API text outputs carry an invisible statistical watermark by default from Nov 7, 2026, covering grok-4.7/4.6/4.5/4.3 and grok-4.20 variants; an x-watermark response header indicates it, and team admins can opt out per model in the console. Others link it to EU AI Act Article 50 transparency duties (relay).
Perplexity Computer's Mac/Windows apps can now drive your own Chrome via an extension
Agents working inside real signed-in sessions — key for merchant ops tasks.
CollapsePerplexity Computer's Mac/Windows apps can now drive your own Chrome via an extension
Aravind Srinivas announced that Perplexity Computer can now control your own Chrome from the Mac and Windows apps through a Chrome extension. The official Computer account says the extension links the two so Computer can use sites you're already signed into; the demo turns a business's reviews into a prioritized list of fixes and replies. It's live for all Computer users on Mac and Windows. Letting an agent use signed-in sessions also puts permissions and account security on the user.
Perplexity Computer turns a shoulder MRI report into an interactive 3D explainer for patientsUnconfirmed
Report explanation and patient communication are among the easiest healthcare-agent wins.
CollapsePerplexity Computer turns a shoulder MRI report into an interactive 3D explainer for patients
Aravind Srinivas reposted an official demo, saying personalized healthcare “will be disrupted big time” by what agents can do today (opinion). In it, Computer reads one anonymized shoulder radiology report, places all nine findings on a skeletal model with six X-ray projections, and explains each in plain language. It is a vendor demo with no stated accuracy evaluation, clinical review or privacy handling, and is not a diagnostic tool.
Musk: xAI's Colossus 2 sites total ~860k GB300s, ~200k dedicated to inferenceUnconfirmed
Training and inference compute built separately — inference demand is scaling on its own.
CollapseMusk: xAI's Colossus 2 sites total ~860k GB300s, ~200k dedicated to inference
Wall St Engine summarized Musk's reply: the Macroharder (written as Macrohard in the post) and Minihard sites each have 220,000 GB300s with 800G NICs for training, Gun Range will match that, and Spam will hold about 200,000 GB300s configured for inference with simpler networking. Musk's original post adds that Gun Range is tracking to be operational by end of November and Spam in December. The ~860,000 total is Wall St Engine's sum; its “~1.2 million Blackwell by year-end” figure does not appear in that Musk post and is unverified.
Shenzhen's AI² Robotics ships AlphaBot 2 wheeled humanoids at scale: 1,000+ units for HKC, new line targets tens of thousandsUnconfirmed
Humanoids moving from demos to factory lines with orders.
CollapseShenzhen's AI² Robotics ships AlphaBot 2 wheeled humanoids at scale: 1,000+ units for HKC, new line targets tens of thousands
CyberRobo reports that Shenzhen-based AI² Robotics is mass-producing and shipping its wheeled humanoid AlphaBot 2. Listed deployments: a three-year agreement with HKC covering 1,000+ robots for display manufacturing, material handling and loading/unloading at Dongfeng Liuzhou's auto production, and robot-run kiosks making drinks for consumers. It adds that the company started a new production line in H2 2026 targeting tens of thousands of units. This is the account's relay; deal values and actual delivered volumes aren't disclosed and no company release is linked.
Meta paper “How to Loop Your Transformer”: looping the full model wins at a fixed parameter budget
Trading parameters for compute — relevant to on-device and low-cost models.
CollapseMeta paper “How to Loop Your Transformer”: looping the full model wins at a fixed parameter budget
Tanishq Abraham shared a new Meta paper on scaling looped transformers (reusing the same weights for multiple passes). Experiments span up to 1B parameters, 155B tokens, 1.07×10^21 training FLOPs and 32 loops. Findings: looping the full transformer gives the best quality at a fixed parameter budget, while looping a middle block pays off more when training FLOPs are constrained; small loop-specific low-rank adapters become more useful as loops increase. Naive Parallel Looped Transformers degrade with more loops, but preserving and separately normalizing the current-token representation substantially fixes it.
Stripe CEO Patrick Collison: an “improvements” box wired to an agent that ships whatever users ask — “liquid software”
End users reshaping software could change how internal tools are delivered.
CollapseStripe CEO Patrick Collison: an “improvements” box wired to an agent that ships whatever users ask — “liquid software”
Patrick Collison says he's adding an “improvements” feature to all his personal vibe-coded apps: a free-text box hooked to an agent that implements whatever is submitted. Another user of one of his apps has already implemented 150 improvements in a few days, and software without this now feels rigid to him. He also asks whether it scales, becomes unmaintainable, works for collaborative software, or can be secure and reliable. It's personal practice and opinion; he says it clearly works in a personal context.
Rock Matrix open-sources EdgeRock-4B: a small reasoning model with fixed-size KV cache, self-reported 82.9% on AIMEUnconfirmed
Memory that doesn't grow with context matters for edge deployment.
CollapseRock Matrix open-sources EdgeRock-4B: a small reasoning model with fixed-size KV cache, self-reported 82.9% on AIME
TestingCatalog reports that Rock Matrix open-sourced EdgeRock-4B, a 4B text-and-image reasoning model with fixed-capacity memory. It keeps the last 1,024 tokens in full and saves important older details to a separate bank, so the KV cache doesn't grow with long chats or long reasoning. Rock Matrix reports an 82.9% mean pass@1 on AIME 2024–2026 versus 18.8% for the strongest bounded-state model it tested. These are self-reported results, not yet independently reproduced.
DeepSeek reportedly doubles round to nearly $16B, backed by Tencent, CATL and state institutionsUnconfirmed
A signal of large-scale outside funding for China's leading lab.
CollapseDeepSeek reportedly doubles round to nearly $16B, backed by Tencent, CATL and state institutions
The Information reports that DeepSeek doubled its latest fundraising target to nearly $16 billion, with backing from Tencent, CATL and some of China's most powerful state-owned institutions. It also notes that black-market prices for Nvidia's advanced AI chips have surged. Valuation and terms aren't in the post and DeepSeek hasn't confirmed; this is a press exclusive.
Nvidia reportedly in talks to guarantee billions in credit for Nscale's $45B Anthropic data centerUnconfirmed
A chipmaker using financing to lock in compute demand.
CollapseNvidia reportedly in talks to guarantee billions in credit for Nscale's $45B Anthropic data center
The Information reports Nvidia is in talks to provide billions of dollars in credit guarantees for Nscale's $45 billion Anthropic data center project. It frames this as a way for Nvidia to secure chip demand as Anthropic increasingly uses competing hardware. Amounts and terms aren't public and no party has confirmed; press relay.
Canada's big banks block third-party personal AI agents from online banking: RBC, TD, CIBC, BMO
Banks are drawing written lines on agents acting inside accounts.
CollapseCanada's big banks block third-party personal AI agents from online banking: RBC, TD, CIBC, BMO
SI Spotlight relayed Bloomberg Law: after a full security review, Royal Bank of Canada told all employees on Tuesday that third-party AI agents can't access its banking systems, targeting tools like Meta's Muse. CBC reported Oct 29 that RBC, TD, CIBC and BMO will all block personal AI agents from directly accessing authenticated account data, with Scotiabank limiting them; TD warned clients never to share credentials, passwords, PINs or one-time codes with an AI agent. The post's claim that Schwab and JPMorgan blocked bots from moving funds two weeks earlier is relay we didn't verify.
DoorDash CEO: third-party AI agents have placed only “several hundred” orders in total
Rare hard numbers on agentic commerce — a reality check.
CollapseDoorDash CEO: third-party AI agents have placed only “several hundred” orders in total
DoorDash CEO Tony Xu said on the Q3 call that third-party agents like Muse have generated only “several hundred” orders in total, and users still mainly order in the DoorDash app (TechCrunch and the call transcript). For comparison, DoorDash said in its 2025 report it averages about 3 million deliveries a day. He sees agents first helping with long-consideration purchases or proactive background tasks, and possible advertising inside agents later. The post also rounds up Alexa now editing Amazon accounts and The RealReal tightening bot defenses (relayed from TechCrunch).