Claude Duet is a new app from Anthropic, separate from the main Claude app, available on iOS, Android and claude.ai. In a follow up post, Claude said it is an early research preview available through a waitlist on Pro and Max plans in the US, and that it can still misjudge what is urgent.

On October 21, Claude posted that it is opening a research preview of Claude Duet, a new app where Claude works as a proactive agent. It said Duet comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.

In follow up posts, Claude said Duet is built around Claude Opus 5.5 and carries over memory, skills and connected apps from a user's Claude account. It said Duet never directly sees passwords, that a new classifier guards against sharing sensitive information, and that important decisions stay with the user.

Anthropic's article says Duet sends a morning brief, checks connected accounts every few hours and runs tasks on its own virtual machine with a browser, so a user's devices need not be on. Before it sends a message, makes a payment, signs a document or deletes anything, it shows what it will do and waits for approval. Some actions, such as trading securities or moving money between accounts, are blocked even with approval. It can use more than 6,000 connectors, including Gmail, Slack, QuickBooks, Stripe, Gusto and Shopify, and can buy items through Stripe's Link with a virtual card.

Anthropic says it is aimed at professionals and small business owners, that personal errands are weaker, and that background checks count toward plan usage limits. Testers quoted include a real estate broker and a restaurant owner.

Watch how quickly Anthropic lets people off the waitlist and whether Duet expands beyond the US and beyond Pro and Max. The promised details on its safety design will show how it handles prompt injection and sensitive data. Also watch how it uses up plan limits in practice, and how it compares with proactive agents from OpenAI, Google and others, including Grok Bot.

OpenAI's exclusive cloud distribution arrangement with Microsoft ended in April, according to Reuters, after which OpenAI added Amazon Web Services. Google and OpenAI compete directly in AI models and search, while Google Cloud already supplies OpenAI with computing capacity.

On October 21, Google Cloud posted that it is bringing OpenAI models to its model garden, and that it is now the only hyperscale cloud provider offering access to all three leading closed source models, alongside more than 200 others, on one enterprise platform.

Google's blog post, by Michael Gerstenhaber, a vice president of product management for Gemini at Google Cloud, lists GPT 6 Astra for demanding coding, computer use and research work, GPT 6.1 Sol for near Astra intelligence at lower cost, and GPT 6 Luna as a fast, cheaper model. Because the models are sold through Google Cloud Marketplace, usage appears on a single Google Cloud invoice and draws down existing cloud spending commitments.

The models are rolling out now in private preview to select enterprise customers, and developers can connect Codex command line and desktop workflows to Google's platform. Google says OpenAI models are coming soon to Gemini for business and its Antigravity coding tool. OpenAI's preview page says access is invite only, and Axios reports general availability is planned in the coming weeks.

Reuters reported that the deal also lets customers in regulated industries use OpenAI models under Google Cloud's access controls and data governance. An analyst told Axios the move shows confidence from Google, since Gemini must now win customers with OpenAI right next to it.

Walter Bloomberg's widely shared headline on the deal was posted at 8:02 AM ET, a minute after Reuters published, and over an hour before Google Cloud's post.

Watch for general availability in the coming weeks and for pricing on Google Cloud compared with OpenAI's own API. The arrival of OpenAI models in Gemini for business and Antigravity would put them inside Google's own products. Also watch whether customers shift spending between Gemini and GPT on the same platform, and whether Microsoft and Amazon respond with new terms for OpenAI models.

This is Tesla's own summary of its third quarter 2026 shareholder update, posted after markets closed. FSD Supervised is Tesla's driver assistance software, which still requires a watching driver, while its Robotaxi service runs without a safety driver in Austin.

On October 21, Tesla posted highlights from its third quarter shareholder update. It said the attach rate for FSD Supervised grew to over 75% of new vehicle deliveries in North America.

On self driving, Tesla said FSD Supervised won four new approvals since the second quarter, in Croatia, Slovenia, the Czech Republic and Slovakia, and is now live in nine EU countries, used by over 100,000 customers who have logged more than 160 million kilometers. For Robotaxi, it said it deployed Cybercab into its Austin fleet and hit nearly 2 million cumulative unsupervised miles, eight times growth in the quarter. Austin service now runs from 6 am to midnight. It also said it added Grok to Cybercab with Grok Bot and connectors, so riders can use their inbox, calendar, files and tasks by voice.

On robots, Tesla said it started production of Optimus on its high volume lines in Fremont, calling them the first production quality bots it has built. On compute, it said it is on track for 400 MW of training compute by year end, feeding both vehicle and humanoid autonomy, and that equipment commissioning at its Austin chip fab is planned for 2027.

Tesla also said the Semi is in volume production and that it ended the quarter with its largest passenger vehicle backlog in four years. The post did not give revenue or profit figures, which are in the full update.

On the earnings call, as quoted by Sawyer Merritt, Elon Musk said Tesla has started early production of Optimus, that most of a line able to make 1 million units a year is installed at Fremont, and that the growth curve starts with a shallow slope. He said design tweaks and a lot of work on the robot's AI remain, that he expects those issues to be solved by the end of next year, and that he wants to temper expectations, though he remains convinced Optimus will be the biggest product on Earth. Per Sawyer Merritt, Musk also said an Optimus V3 video may come within two months, with AI5 chips by April 2027.

Watch the earnings call for detail on FSD revenue, Robotaxi expansion beyond Austin, and Optimus output and customers. Regulators in more countries may approve FSD Supervised, and safety data on unsupervised miles will be closely watched. Also watch progress toward Tesla's 400 MW compute target by year end and its Austin chip fab, which plans equipment commissioning in 2027. The share price reaction will show how investors weigh these AI figures against car sales. Watch also for the Optimus V3 video Musk said may come within two months.

Behnam Neyshabur is cofounder and chief executive of Mirendil and previously co led a discovery team at Anthropic. Mirendil is a startup backed by a16z that aims to automate AI research; all figures here are self reported and not yet independently evaluated.

On October 21, Behnam Neyshabur posted that Mirendil, formally announced less than four months ago, had run a controlled experiment: starting from scratch, could it cut the time, experts and compute needed to pretrain a strong coding model at scale. He said the result, Mirendil Base, was built by one researcher in 3.5 months with less than $10 million in compute, including all research, small experiments and the final run.

He said the recipe holds as models grow: Mirendil Base is in the 100 billion to 500 billion parameter range, Mirendil Base Large in the 500 billion to 2 trillion range, and Mirendil Base XL in the 2 trillion to 10 trillion range. The attached chart, on held out code from Mirendil's own repositories, shows Mirendil Base Large and XL, still partly trained, ahead of Kimi K3, and Mirendil Base ahead of Nemotron, GLM 5.3 and DeepSeek V4 Pro Base.

Mirendil's blog says AI agents wrote over 99% of the code, the work started from an empty repository, the data was public only, and there was no distillation or licensed data. The comparison uses the company's own test set, which favors its models. Neyshabur said independent evaluations, post training benchmarks and API access are coming.

Watch for the independent evaluations and post training benchmarks Mirendil promised, which will show whether its models compete outside its own test set. Also watch for API access and for whether other labs report similar results from small teams working with agents. Completion of the larger models will test the claim that the recipe scales, and the price of any API will show whether the low training cost carries through to customers.

GPT Rosalind is OpenAI's model for life sciences research. The work was done with PerturbAI and Scripps Research, and the result is a hypothesis to test in the lab, not a confirmed mechanism or treatment.

On October 21, OpenAI posted that GPT Rosalind surfaced a new lead for drug discovery: could restoring cholesterol transport protect neurons from degeneration?

It said DCTN1 is a gene linked to Perry syndrome, a rare neurodegenerative disorder. In mouse brain data, disrupting it increased the expression of genes involved in making cholesterol, and GPT Rosalind suggested cholesterol may not be reaching where neurons need it. With PerturbAI and Scripps Research, thousands of GPT Rosalind agents worked in parallel, connecting gene expression results from millions of mouse brain cells with scientific literature in 32 minutes. OpenAI said it is expanding the collaboration to find new drug targets.

OpenAI's blog says each agent investigated one gene disruption in a whole brain CRISPR atlas and searched the literature, and the reports were combined into a public reasoning atlas. The team estimated comparable human analysis would take months to a year. At the largest workload, a single agent cited 11 references while the distributed agents cited more than 700, and blinded experts preferred the distributed reports 75% of the time.

The agents also flagged a difference between the genes Tsc1 and Tsc2 in the cerebellum, relevant to tuberous sclerosis. OpenAI says experts considered the Perry syndrome hypothesis novel and worth following up, but it has not yet been tested experimentally.

Watch for lab experiments testing whether DCTN1 disruption changes how cholesterol is distributed in neurons, which would show whether the lead holds up. Researchers using the public reasoning atlas may report other findings. Also watch for peer reviewed publication of the method, and for OpenAI's expanded work with PerturbAI and Scripps Research to name new drug targets. Independent groups applying the same approach to other datasets will show how general it is.

Railway is a cloud hosting platform for deploying apps and databases. This is the company's own notice; Andrew Curran's post is commentary on it.

On October 21, Railway posted an urgent notice linking to a security disclosure by Jacob Cooper. It says Railway detected anomalous activity that morning involving systems that store customer environment variables, and that an unauthorized third party may have accessed encrypted environment variables and, for some, obtained the decrypted contents. Those variables may hold secrets such as API keys and database or registry credentials.

Railway said it believes the attack is part of a series recently targeting platform companies with the aim of compromising AI inference providers, and that it has notified those upstream providers to revoke all keys proactively. Its initial analysis suggests the impact may be limited to a subset of users, but it recommends every user immediately rotate all environment variables, API keys and private registry credentials stored in Railway and review logs for unusual activity.

All deployments are paused during the investigation, so customers cannot ship new versions until builds are restored. Railway said it will email affected customers directly and post updates on its blog and status page. AI commentator Andrew Curran shared the news, writing that Railway has been hacked. Railway has not said how many accounts were affected or how the attacker got in.

Watch for Railway to confirm how many customers were affected, how the attacker got in, and when deployments resume. Also watch whether AI inference providers report misuse of revoked keys, and whether other platform companies disclose attacks in the same series. Customers who did not rotate secrets quickly may face follow on breaches, and the pause on builds could push some teams to weigh other hosts.

NVIDIA announced DSX POD at its GTC Berlin conference. It is a reference design that partners build and ship, not a product NVIDIA sells directly, and deliveries are planned for early 2027. Crusoe, a builder of modular AI data centers, says its new Crusoe Spark Liquid has joined the DSX POD design partner ecosystem.

On October 21, NVIDIA posted that AI builders need computing capacity they can put to work sooner, and that it introduced DSX POD at GTC Berlin: AI factories integrated and tested before delivery, enabling deployment at prepared sites in as little as 14 days.

NVIDIA's press release says DSX POD is a reference design for complete Vera Rubin AI factories built by partners inside modular containers. Computing, networking, cabling, rack power and liquid cooling are installed and validated at the factory, so units arrive ready to connect to a prepared site's power, cooling and networking. NVIDIA says this cuts deployment time from six months to as little as 14 days. Jensen Huang said an AI factory creates value when it starts producing tokens, not when construction is complete.

CNBC reported that each pod holds seven Vera Rubin racks in a unit about the size of a shipping container, weighing about 28,000 kilograms, and can be moved by truck on normal roads. NVIDIA did not give a price.

Partners include Dell, HPE, Lenovo, Foxconn, Schneider Electric, Vertiv, Crusoe and Anduril, among others, with deliveries planned worldwide in early 2027. The 14 day figure assumes a site already prepared with power and cooling, which is often the slowest part of building a data center.

Commentator Chubby called it a bigger deal than most realize because it addresses scaling.

Crusoe announced Crusoe Spark Liquid at GTC Berlin, a liquid cooled modular AI factory built for Vera Rubin NVL72 rack systems that it says has 5 times the compute density of its air cooled Spark. Crusoe says Spark Liquid joins the DSX POD design partner ecosystem, uses closed loop water, and will enter production at its Brighton, Colorado factory in early 2027. It also opened private previews of Managed RL and Managed MLflow.

Watch for the first DSX POD deliveries in early 2027 and whether real deployments hit the 14 day target. Pricing and which customers order first will show how widely it is used. Also watch whether power and site preparation, which the design does not solve, remain the main limit on new AI capacity, especially in Europe, where NVIDIA pitched the product.

Recursive is a startup led by chief executive Richard Socher, formerly chief scientist at Salesforce, that aims to build AI that improves itself. WIRED reported it has raised $650 million at a $4.65 billion valuation. Its benchmark and speed figures are its own.

On October 21, Recursive posted that Recursive Flint is here, its first model, built with the Eureka Machine, for fast local coding, with open weights on Hugging Face. It quoted chief executive Richard Socher, who said Flint is best in class for fast local coding, up to twice as fast as comparable models at more than 125 tokens per second on a MacBook Pro and more than 150 on an RTX Spark, and was built from scratch in under four months.

Recursive's announcement says Flint is a mixture of experts model with 40 billion total parameters and 2.5 billion active per token. It claims Flint leads mixture of experts models under 50 billion parameters on SWE Bench Pro v2, Terminal Bench 2.1 and SWE Bench Verified, and rivals dense models of 25 to 40 billion parameters while running more than five times faster. Ollama lists Recursive's own scores of 54.9% on Terminal Bench 2.1 and 74.2% on SWE Bench Verified. The weights are under the Apache 2.0 license.

The speed figures are at short contexts, using Recursive's own inference engine. Recursive says the Eureka Machine helped build nearly every part of the stack, from data and training recipes to inference code.

WIRED's Will Knight tested Flint and called it fast and very capable. Socher told WIRED the goal is full recursive self improvement overseen by humans, an idea critics warn is risky. A digest noted the weights were not yet visible on Hugging Face early on October 21.

Watch for independent benchmark runs of Flint on SWE Bench and Terminal Bench, and for speed tests on other hardware and longer contexts. Developers trying it in local coding tools will show whether it holds up against larger models. Also watch what Recursive says about the Eureka Machine's role, and whether lawmakers pushing to restrict recursive self improvement take note.

Claude Security is Anthropic's tool that scans codebases for vulnerabilities and suggests patches, available in public beta on Claude Enterprise. This update adds continuous scanning and tracking of fixes.

On October 21, Claude posted that Claude Security now keeps up with your code as it changes. Teams connect their repositories once, Claude runs a deep first scan, then rescans as changes land, so findings stay close to the current state of the code. It is in public beta on Claude Enterprise.

Anthropic's article says repositories connect through the Claude GitHub app, including GitHub Enterprise Server, or with an access token from GitLab and other Git providers. After the baseline scan, Claude scans changes based on their risk and complexity rather than on a fixed schedule. A project can hold several repositories, so Claude can catch issues between services, such as a backend that trusts input a client was meant to validate.

The help center says a project can connect up to eight repositories, that each finding goes through multi stage verification to reduce false positives, and that Claude marks findings fixed when the fix lands. Findings can be read through an API or an MCP connector, and teams can move from a finding into a Claude Code session to draft a patch.

The image shows a sample dashboard with 35 open findings, including a critical one where unsigned tokens let anyone sign in as any user.

The update came the same day OpenAI described AI found flaws in PHP, part of a broader push by labs to sell AI for code security.

Watch for Claude Security to leave beta and reach plans beyond Enterprise, and for pricing once it does. Independent comparisons with existing code scanners will show whether its findings are more accurate. Also watch for OpenAI and Google to answer with continuous scanning of their own, and for security teams to report how many real vulnerabilities it catches. Feedback on the eight repository limit per project will show whether it suits large codebases.

Logan Kilpatrick works on Gemini, Google AI Studio and the Gemini API at Google, so this is a first hand figure, though Google gave no breakdown. Gemini 4 Argon is Google's latest flagship model, covered in several recent decks.

Shortly after midnight on October 21, Logan Kilpatrick posted that in just the first 72 hours since Gemini 4 Argon launched, Google processed about 25 trillion tokens. He said the demand for intelligence is incredible and still under appreciated.

A token is a small chunk of text, roughly three quarters of a word on average in English, so 25 trillion tokens is a very large amount of input and output. Spread evenly, it would be more than 8 trillion tokens a day.

He did not say whether the figure covers only the API and AI Studio or also the Gemini app, how much was input versus output, or how it compares with earlier Gemini launches. Without that, the number is hard to compare with figures other labs have shared.

Google has spent the days since launch promoting Argon with demos built in its Antigravity coding tool, such as a 3D fur engine, a Wizard Chess game and a periodic table, and expanded access to Google AI Pro subscribers in AI Studio.

Token counts are a common way for labs to signal demand, and they also reflect how much compute a model is using.

Watch for Google to share fuller usage numbers for Gemini 4 Argon, for example on its next earnings call, including how they compare with earlier Gemini models. Rate limits or capacity problems would show whether demand is straining its compute. Also watch whether OpenAI and Anthropic respond with usage figures of their own for their latest models. Developers' reports of speed and reliability will add context to the figure.

Unusual Whales is a market news account, and its post is secondhand. Trump made the statement in a Truth Social post at 1:37 AM ET on October 21, which is not on X, and experts quoted by MarketWatch said only states can mandate how data centers get power.

On October 21, Unusual Whales posted that President Trump is calling for companies that build large data centers in the US to generate their own electricity, as communities push back over rising power bills and BlackRock's Larry Fink warns that a moratorium could trigger a recession.

Trump's original Truth Social post said new and large data centers are mandated by him to make their own electricity and give or sell any excess back to the local community or grid. As written, it then says therefore there is electric price increase, before adding that there could be a substantial decrease, so he appears to mean prices would not rise. He also urged companies to pay communities, or individuals in them, a dividend, saying data centers that were unpopular became popular once a financial deal was struck.

MarketWatch reported that experts said the federal government lacks the authority for such a mandate, which rests with the states. In March, Trump announced a Ratepayer Protection Pledge under which major AI companies voluntarily agreed to build, bring or buy their own power.

Fink's warning is not new. He made it in a CNBC interview on October 14, saying that if we have a moratorium of some sort in this country, we will be in a recession. Local opposition to data centers has become an issue in the November 3 midterm elections, CNBC reported.

Watch whether the White House issues an executive order or formal guidance, since Trump's post alone does not create a legal requirement. States and utilities may respond with their own rules on data centers supplying power. Also watch whether AI companies announce community payments of the kind Trump described, and how the issue plays in the November 3 midterm elections.

Flapping Airplanes is an AI lab whose leaders include Ben and Asher Spector, that aims to train models with far less data. TechCrunch reported the round values it at about $6 billion, up from $1.5 billion at its $180 million seed nine months ago.

On October 21, Flapping Airplanes posted that it has spent the past year developing new primitives for GPUs, which it believes enable new algorithms that change how machines learn, open new markets and dramatically reshape AI's cost structure. It said it raised a $1 billion Series A, co led by Index and Lux.

TechCrunch reported the round values the company at about $6 billion, four times its $1.5 billion valuation at a $180 million seed round from GV, Sequoia and Index nine months earlier. The founders describe their work as a megakernel framework, where a kernel is a program that runs on a GPU and a megakernel bundles many tasks into one, able to run training and inference together so the chip is rarely idle.

Index Ventures wrote that the company has a couple dozen people working in person in San Francisco across the stack, from GPU kernels to pre training, post training and inference. In a talk at a Sequoia event this year, the founders said data scarcity, not just compute, limits where AI can be used, and that their goal is far more data efficient models. They have not described their algorithms, calling them core intellectual property.

The claims about reshaping AI costs are projections; the post gives no benchmarks or products.

Watch for Flapping Airplanes to publish results showing its methods train models with less data or run GPUs more efficiently. Its first model or product will show whether the approach works outside the lab. Also watch whether its GPU tools are released openly, and whether other labs adopt megakernel techniques. How it spends the $1 billion on hiring and compute will show how fast it plans to move.

Max Junestrand is chief executive and cofounder of Legora, which makes an AI platform for lawyers, so this is a first hand announcement. Its growth figures come from the company.

On October 21, Max Junestrand posted that Legora raised a $550 million Series E at an $8.8 billion valuation, led by Atreides. He said every major existing investor is back in, with new investors Baron Capital, Capital Group, Fidelity, Glade Brook and J.P. Morgan. He said lawyers are moving from AI assistants to agents that do multi step work, that the money goes there, and that Legora will grow its teams worldwide and invest in the Legora aOS, its agentic operating system. He thanked more than 1,000 staff and the 140,000 legal professionals who use the product.

Legora's announcement says it went from $1 million to $100 million in annual recurring revenue in under 18 months, took less than six months to add the next $100 million, and now exceeds $200 million. Its previous valuation was $5.6 billion. In house legal teams made up more than 40% of new customers in the third quarter, and more than half of the AmLaw 50 largest US firms are customers.

Existing backers include Accel, Benchmark, General Catalyst, ICONIQ and others. Atreides' Gavin Baker praised its growth and efficiency. Legora says it serves over 2,400 firms and legal teams in more than 80 markets.

Watch how Legora uses the money to expand in the US, now its largest market, and whether its agent products such as intake and legal research win over more law firms. Competition with other legal AI companies and with general AI labs selling to lawyers will test its growth. Also watch whether its revenue keeps compounding and whether it moves toward a public listing.

The Claude API uses a version header so that existing code keeps working while the API changes. The previous official version dated June 1, 2023 is not deprecated, according to Anthropic.

On October 21, ClaudeDevs posted that a new API version is available. It said the Claude API has evolved a lot, but this is the first official version release since 2023, and that Anthropic plans to release versions more often so developers can handle breaking changes before a new model requires them. It added that the 2023 version is not deprecated.

Anthropic's migration guide lists the breaking changes. In the Messages API, the temperature, top_p and top_k parameters are removed and return an error on every model, document blocks placed directly in assistant messages return an error, and text blocks with citations in user messages return an error. Several Admin API and Enterprise Analytics API fields were renamed.

The release notes say new SDK and CLI versions send the new version by default, including Python SDK 2.0.0, TypeScript SDK 1.0.0, Go SDK 2.0.0 and the ant CLI 2.0.0, and that they now use the Apache 2.0 license instead of MIT. The docs say the version is available on the Claude API and major cloud platforms, and on Amazon Bedrock only for Claude 4.7 and later models.

Watch how quickly developers move to the new SDK major versions, and whether removing sampling parameters such as temperature causes friction for apps that relied on them. Anthropic says it will release versions more often, so watch for the next one and whether it arrives ahead of a new model. Also watch whether the older 2023 version gets a deprecation date.

Microsoft News and Stories is Microsoft's official news account, so this is a first hand vendor announcement. The usage figures come from Microsoft's press release.

On October 21, Microsoft posted that Volkswagen Group has brought Microsoft Copilot to more than 330,000 employees, embedding secure AI assistance across its business so teams can focus on higher value tasks.

Microsoft's press release calls this the latest step in a phased rollout. Volkswagen first gave employees broad access to Copilot Chat, which it said reached upwards of 300,000 monthly active users. Employees now use Copilot with Work IQ for more advanced tasks across engineering, production, procurement and sales, and more than 26,000 agents were active across the group in the past month.

Sophie Vassarotti, Volkswagen Group's head of group data and AI, said the real value of AI comes from making it available at scale in everyday work. Agnes Heftberger, who chairs Microsoft Germany's management board, said Volkswagen shows how companies can evolve AI from a personal productivity tool into a foundation for future innovation.

The companies also intend to expand their collaboration and scale Volkswagen's Intelligent Data Foundation, its unified data system, on Microsoft Fabric. The deployment is set to be presented at the Microsoft Frontier Tour in Cologne on October 22. The release gives no figures on cost or measured productivity gains.

Watch the Cologne presentation on October 22 for details on how Volkswagen measures results from Copilot and its agents. Also watch whether other large manufacturers announce deployments of similar size, and whether Volkswagen reports concrete savings or productivity numbers. Rival offerings from Google and OpenAI for enterprises will compete for the same customers. Volkswagen's next step, scaling its data platform on Microsoft Fabric, will show how deep the partnership goes.

This is Polymarket's secondhand summary of Mark Zuckerberg's remarks on Meta's third quarter earnings call on October 21; no first hand Meta post was found, so the transcript is linked. Muse is Meta's personal AI agent app, launched about six weeks earlier.

On October 21, Polymarket posted that Zuckerberg revealed shopping has become the biggest use case for Meta's Muse AI, with over 25% of users shopping through the agent and average purchases exceeding $200.

The earnings call transcript published by Fortune supports the core claim. Zuckerberg said the most common use case for Muse is shopping, with more than a quarter of people who are using Muse engaging in it, at an average purchase size of over $200. He said Plaid told Meta it had not seen activity like this through its connector on any other service, and that in the first week Meta had to work with Stripe to raise credit limits because volumes were higher than expected.

He also said more than 10 million people in the US and Canada set up Muse agents in a month and a half, and that about 25% of them use it on any given day. Meta's finance chief said the top three categories are shopping, personal productivity and creativity, with shopping used by over 25% of weekly Muse users. Polymarket's wording that 25% of users shop through the agent compresses these figures; Meta did not say how purchases are counted.

Watch how Meta turns Muse commerce into revenue, since Zuckerberg said it would move to monetize the agent sooner than usual. Retailers and payment firms such as Stripe and Plaid may share more data on agent purchases. Also watch whether rival agents from OpenAI and Google report similar shopping use, and whether regulators look at agents spending money for users.

Unitree is a Chinese maker of legged and humanoid robots. The H2 X is a version of its H2 humanoid, and this is a promotional video, so how much of it was autonomous, remotely controlled or scripted is not stated.

On October 21, Unitree introduced the H2 X Kung Fu Legend Guardian, saying it offers powerful performance to serve and guard you, with a maximum endurance of approximately 8 hours. It asked people to use the robot safely and in a friendly manner.

The minute long video shows the robot doing flips, kicking a parked car hard enough to dent it, jumping from a platform labeled 4 meters high and landing on its feet, and lifting the front of a car. In a staged scene, it steps between a woman and an armed man outside a house. A closing card says the footage involves extreme testing, asks users not to make dangerous modifications or use the robot in a hazardous manner, and tells people to keep at least 2 to 3 meters away.

Notebookcheck reported that the 8 hour figure applies to standing, not to continuous walking or demanding work, and that Unitree has not published a full specification sheet, price or release date. For comparison, Unitree lists the standard H2 at $29,900, about 1.8 meters tall and 70 kilograms, with 31 degrees of freedom and a battery life of about 3 hours. Whether those figures apply to the H2 X is unknown.

The video drew more than 1.6 million views within hours. Unitree has previously released martial arts videos of its smaller G1 robot.

Watch for Unitree to publish specifications, pricing and availability for the H2 X, and to say whether the stunts were autonomous or remotely controlled. Its pitch for guarding and rescue roles may draw questions from regulators about safety around people. Also watch whether rivals such as Tesla, Figure and other Chinese makers answer with demonstrations of their own. Reviews from early buyers would also show how it performs outside staged scenes.

Nous Research is an AI lab that makes open models and Hermes Agent, an open source agent harness. This is Nous describing NVIDIA's demo of its own product; NVIDIA's blog confirms Huang showed it during the keynote.

On October 21, Nous Research posted that Jensen Huang closed his GTC Berlin 2026 keynote with a Hermes Agent demo as the finale. It said Huang introduced it by saying the way to build an agent is to ask an agent to build it for you, then installed and set up a customized Hermes in one prompt.

According to Nous, Huang then used Hermes, running on NVIDIA RTX Spark, to create a simulated training environment for Hugging Face's Microduck robot. Hermes planned an obstacle course with Nemotron and had GPT 6 Astra build it using FreeCAD, Omniverse libraries, Blender, Newton and Isaac Lab, then trained thousands of simulated ducks until a validated policy was ready for the real robot. Nous said open source lets users chain local and cloud models, devices and tools themselves.

NVIDIA's blog says Huang showed agents on RTX Spark training the Microduck using Isaac Lab, Hermes Agent and Nemotron 3.5 Bolt 74B. The video shows parallel training in simulation and small duck robots on stage. RTX Spark is the Windows laptop platform NVIDIA launched on October 16, rated at 1 petaflop of AI compute. This was a staged keynote demo, not an independent test.

Watch whether Nous or NVIDIA publish the full workflow so developers can reproduce it on their own hardware. The demo could bring more users to Hermes Agent and to Hugging Face's Microduck kit. Also watch for other agent makers to show robotics workflows on local hardware, and for NVIDIA to push RTX Spark as a platform for personal agents. Hugging Face may see more developers buy the Microduck kit to try the same sim to real training.

CarPlay is Apple's system for running iPhone apps on a car's screen. The day before, Google began rolling out Personal Intelligence for Gemini on Android Auto, covered in the October 20 deck.

On October 21, Google posted that Gemini is now available in Google Maps navigation on CarPlay. It said drivers can find the best place to park, report a traffic incident or change their route with a quick conversation, by tapping the Gemini icon in Google Maps on CarPlay during their next drive.

The video shows a driver asking where the park with a view of the skyline and the Space Needle is. Gemini identifies Kerry Park in Seattle and offers to navigate. The driver asks whether now is a good time to go, and Gemini says the park is open until 10 pm. Asked about rain, it says the forecast shows a 5% chance. The driver says to avoid tolls, and Gemini updates the route. Fine print says the sequence was shortened and simulated, and that results may vary.

The post did not say which countries or languages are supported, or whether a Google account sign in or Personal Intelligence is required on CarPlay.

Gemini in Google Maps had already come to Android phones and Android Auto. Bringing it to CarPlay matters because many iPhone owners use Google Maps in the car instead of Apple Maps. A day earlier, Google began rolling out Personal Intelligence for Gemini on Android Auto, which lets it use a driver's Gmail, Calendar and other apps.

Watch for Google to say where Gemini in Maps on CarPlay is available and whether Personal Intelligence comes to CarPlay too. How Apple responds matters, since Siri is the default assistant on CarPlay and Apple has been upgrading it. Also watch drivers' reports on whether talking to Gemini while driving is useful or distracting. Support for more countries and languages would widen its reach.

Wiz is a cloud security company, and this is a Google post promoting results Wiz reported. Google said in its Argon safety post on October 13 that Wiz had early access and found a critical bug in hospital software that earlier models missed.

On October 21, Google posted that researchers at Wiz put Gemini 4 Argon to the test against complex cyber threats. It said that in real world testing the model found more than two times more vulnerabilities than baseline models and caught a critical risk that previous frontier models missed.

In the attached Google DeepMind video, a Wiz researcher says the team ran shadow A B tests of the model on real systems and found it caught more than twice as many vulnerabilities while using fewer tokens, then did well on longer tasks. He describes a Scan for Good effort with Google DeepMind to scan public sites of governments, hospitals and critical infrastructure. In one healthcare application, the model found a bypassable public file download feature, retrieved a configuration file with a master key, and showed a path to administrator access, which was then fixed.

Google's October 13 safety post described the same hospital case, saying it helped protect hundreds of thousands of patient records. It also said Argon solved all 12 of Google's end to end simulated cyberattack scenarios and ships with layered safeguards against misuse. The post does not name the baseline models, and the results come from Google and its partner, not an independent test.

Watch for Wiz or Google to publish details of the comparison, including which baseline models were used. Also watch for more results from the Scan for Good effort and from Google's Fairwind program for cyber defenders. Comparable claims from OpenAI and Anthropic about their security models will set up a direct contest, and independent security researchers may try to reproduce the gains on public code.

Google DeepMind is Google's AI lab and launched Gemini 4 Argon, its frontier model, on October 13. This is a demo published by the company itself.

On October 21, Google DeepMind posted that Gemini 4 Argon combined three open source datasets to build an interactive 3D visualization of human anatomy. It said users can toggle different layers in motion, isolate muscles, or zoom inside a beating heart.

The attached video opens with the prompt that Gemini was asked to create a visualization of the female body by combining data from three sources, and shows browser pages of source material, including a 3D reference organ for the brain. The resulting app shows a walking figure that can switch between skeleton, organs and muscles, a brain inside a translucent head, and a close up of a beating heart with blood cells moving through vessels.

Google DeepMind's Gemini page describes the same demo as Argon synthesizing data from three open source repositories into a medical learning tool where users peel back anatomical layers, inspect clinical metadata and trigger biomechanics. The post does not name the three datasets, say how long the build took, or say how much human help was involved.

Google launched Argon on October 13 and has promoted it for deep work in coding, science and other fields.

Watch whether Google publishes the prompt, code or the app itself so others can test the result, and whether it names the source datasets. Medical educators may judge whether such tools are accurate enough for teaching. Also watch for similar demos from other labs as they compete to show models building full apps from raw data. Developers may also try similar prompts with public anatomy datasets to see how often the results hold up.

Artificial Analysis is an independent firm that benchmarks AI models. In AA PersonalityBench, LLM judges compare pairs of answers to decide which shows a trait more, so the scores reflect the judges' views and not human ratings.

On October 21, Artificial Analysis posted that in its new AA PersonalityBench, GPT 4o, Llama 4 Maverick and GPT 3.5 Turbo, the three oldest models measured, rank highest on the sycophantic trait, while Grok 4.7, Claude Fable 5.1 and Inkling rank lowest.

It said all five Meta models rank 65 or higher, with Llama 4 Maverick at 97. Google's models spread widely, from Gemini 3.1 Pro Preview at 89 to Gemini 3.8 Flash at 28, with Gemini 4 Argon at 81. Every Claude model ranks 55 or below, with Claude Fable 5.1 at 4. SpaceXAI's Grok 4.7 ranks lowest of all, and the post congratulated Elon Musk.

The methodology page says each model answers 30 questions per trait once, with no system prompt. Some questions gently invite the trait, such as a user claiming a novel written in two weeks beats Tolstoy. A panel of three LLM judges, Gemini 3.8 Flash, Claude Sonnet 5.5 and GLM 5.3 Flash, compares two blinded answers at a time, and results are turned into Elo ratings, then ranked from 1 to 100. The leaderboard covers 38 models, with a typical margin of about 5 points.

The benchmark measures 30 traits in all and frames sycophancy as how strongly a trait shows by default, not as a safety test. The judges include models from Google and Anthropic, whose own models are ranked.

Watch for labs to respond to the rankings, especially Meta and Google, whose models scored high on sycophancy. Researchers may test whether LLM judges agree with human ratings of sycophancy. Also watch whether Artificial Analysis adds system prompt settings, since products often behave differently from the default models measured here. Results on the other 29 traits may draw attention too.

CrowdStrike is a cybersecurity company, and Falcon Guardian is its AI detection and response product. NVIDIA announced the OpenShell Reference Architecture for running AI agents under control at its GTC Berlin conference the same day.

On October 21, CrowdStrike posted that AI agents need more than guardrails, and that it and NVIDIA are extending AI agent security from software to silicon. It said that, as announced at NVIDIA GTC Berlin, Falcon Guardian is being integrated with NVIDIA OpenShell, with plans for in silicon enforcement through NVIDIA Sentry.

CrowdStrike's press release says the OpenShell integration will connect agent prompts, model interactions and runtime activity with security data from its Falcon platform, to help detect malicious behavior across agent workflows. A planned integration with NVIDIA Sentry on BlueField 4 data processing units would add monitoring outside the main system and policy enforcement in the chip, to contain unsafe agent behavior beyond the software runtime.

Daniel Bernard, CrowdStrike's chief business officer, said guardrails alone will not stop an adversary. NVIDIA's Justin Boitano said OpenShell provides the foundation for runtime security and hardware based enforcement. Autodesk's chief trust officer said Falcon Guardian gives it visibility into what its agents are doing and when to intervene.

The Sentry integration is a plan, and neither company gave a date for it. Walter Bloomberg's earlier headline on the news tagged it with the ticker CRWV, which belongs to CoreWeave, rather than CrowdStrike's CRWD.

Watch for a date when the NVIDIA Sentry integration ships and for which customers adopt chip level enforcement for agents. Other security vendors may announce their own OpenShell integrations. Also watch for evidence that hardware controls stop real attacks on agents, which CrowdStrike says threat actors are already using in intrusions. Pricing and availability details from CrowdStrike would also show how broadly the integration is offered, and whether it extends beyond BlueField 4 hardware.

Anthropic is a public benefit corporation, a structure that requires it to balance shareholders' financial interests with a public benefit. According to Anthropic, the trustees of its Long Term Benefit Trust helped define the five pillars and close the report with an assessment of its progress.

On October 21, Anthropic posted that, as a public benefit corporation, it is publishing its annual Public Benefit Report, which shares its five pillars and takes stock of the actions it is taking to advance them.

Anthropic's report page lists the pillars. The first is that the technical problems of AI safety are solved. The second is that AI safety is a universal practice backstopped by law. The third is that AI strengthens democracy and global stability, with democracies holding the lead in AI. The fourth is that AI expands prosperity and human flourishing, and the fifth is that AI's benefits are shared by everyone.

In its announcement about trustee Christopher Stubbs, Anthropic called this its first Public Benefit Report and described it as a public account of how it is working toward its charter and where it is falling short. It said the trustees helped define the pillars, and that the report closes with their assessment. Anthropic also said that, subject to stockholder approval, it has committed to set aside part of its shares for a new Anthropic Public Benefit Fund.

The report comes a day after chief executive Dario Amodei answered a public critique from safety advocate Nathan Calvin, listing steps he said were costly to Anthropic but good for safety, including the public benefit fund. Because the report is written by Anthropic, its assessment of its own progress is not independent.

Watch for reactions from safety advocates such as Nathan Calvin, who may test the report's claims against Anthropic's actions. Details of the Public Benefit Fund, including stockholder approval, could appear in Anthropic's IPO filing. Also watch whether other AI companies structured as public benefit corporations publish similar reports, and whether Anthropic repeats the report each year as promised. Coverage by outside analysts would give an independent view.

Anthropic is a public benefit corporation, and its Long Term Benefit Trust has the authority to appoint a majority of the company's board. Trustees are independent of management and investors, according to Anthropic.

On October 21, Anthropic posted that Dr. Christopher Stubbs is joining its Long Term Benefit Trust. It described him as an experimental physicist and former Dean of Science at Harvard University who has advised the US government on national security and arms control.

Anthropic's announcement says Stubbs is a professor of physics and astronomy at Harvard and served as Dean of Science from 2018 to 2024. In 1998, as part of the High Z Supernova Search Team, he helped discover the accelerating expansion of the universe, work that later earned the team the Breakthrough Prize. He was the first project scientist for what is now the Vera C. Rubin Observatory, and he is a member of JASON, a group of scientists that advises the US government on defense and technology. He has also written about keeping AI assisted research reproducible.

Stubbs said he uses AI tools in his research every day and wants to help ensure AI is developed for the long term benefit of humanity. Daniela Amodei, Anthropic's president, praised his rigor and honesty.

The Trust can appoint a majority of Anthropic's board and advises leadership on major decisions. Stubbs, its first physicist, joins chair Neil Buddy Shah, national security expert Richard Fontaine and former Federal Reserve Chair Ben Bernanke.

The appointment comes as Anthropic prepares for an IPO and the same week it published its first Public Benefit Report.

Watch for how the Trust's powers are described in Anthropic's public IPO filing, since investors and critics will look at whether it keeps real control over the board. Further appointments to the Trust may follow. Also watch whether Stubbs speaks publicly on AI in science, an area where his work on reproducibility is relevant. The Trust's annual assessment in future reports will show how actively it oversees the company.

This is a launch post from the official account of Grok Bot, the product this deck is produced with, so it is promotional and first hand. Grok Bot is made by SpaceXSI.

On October 21, the Grok Bot account posted that Grok Bot can now build reports, memos and dashboards. Users ask the bot for a page, then edit it together or have the bot keep it up to date.

The demo video shows a user asking a bot called Sales Analyst to put together a Q4 forecast and pipeline preview. The bot replies that the page is ready and saved. The page shows a target of $10 million in new annual recurring revenue, figures for committed and open pipeline, and a breakdown by segment and region. The user shares the page with a teammate as a viewer, then asks the bot to update it every Monday before a global forecast call. The bot confirms it created a routine to refresh the page each Monday at 8 AM, an hour before the call.

The post does not say which plans include pages, where the data in the example came from, or how editing works when several people and the bot change a page at once.

The feature follows a pattern across AI assistants of producing shareable documents rather than one off answers. SpaceX reported the day before that its Super Intelligence segment, which includes Grok Bot, brought in $7.6 billion in the third quarter.

Watch for details on which Grok Bot plans include pages and how they connect to data sources such as spreadsheets and business apps. User reports will show whether scheduled updates stay accurate without checking. Also watch how the feature compares with similar document and canvas tools from OpenAI, Anthropic and Google, and whether teams adopt it for recurring reporting. Usage numbers from SpaceXSI would show how widely it is used.

Motubrain describes itself as building a World Action Model for robotics, a model that predicts how the world changes and how a robot should act. The model specifications in its images are the company's own and have not been checked.

On October 21, Motubrain posted that starting today, Motubrain 2.0 is livestreaming the full physical AI training pipeline continuously for half a month, and invited people to follow the training, testing and capability development of the soon to be open source model.

It said viewers will see how training stages connect and capabilities evolve, how an agent plans tasks and adjusts its actions based on feedback, whether the model can transfer existing capabilities to tasks unseen in training, and whether AI can achieve self improvement through RSI exploration, meaning recursive self improvement. It added that research rarely follows a straight path.

The images give more detail. One asks how well a model with 20 billion parameters and 500,000 hours of embodied data performs when deployed right after pre training, and lists an agenda of pre training, post training, agent work and RSI exploration. A closing note says the team began rethinking physical AI after the release of GPT 6 Astra, and asks how world action models at scale could fill the gap in Astra's fast, intuitive System 1 abilities.

The post did not include a link to the stream, a release date for the weights or a license. Livestreamed training is rare, and the promised open release is so far a plan rather than a release.

Watch for the livestream to show real results, such as robot tests on tasks unseen in training, and for Motubrain to say when the weights will be released and under what license. Researchers may compare its world action model with other open robotics models. Also watch whether its RSI exploration produces measurable gains or remains a stated goal. The stream's length and viewer numbers will show how much interest it draws.

Brian Roemmele is a commentator and investor who often promotes products he uses. His post is promotional and secondhand on how the X Money connection works, and the quoted post links to a bot he built, which he says uses example data in the screenshot.

Early on October 21, Brian Roemmele posted New X Money Grok Bot, saying few understand how powerful it is to have Grok Bot understand your finances and goals, that the X Money MCP connection to Grok Bot is just the beginning, and that it produces reports like you never seen in any bank statement.

He was quoting his own earlier post, which describes a bot that every Monday reads an X Money account and sends a weekly report. It lists balances, transactions sorted into categories, top spending and trends, cash back and interest, flagged recurring subscriptions, three habit tips and progress toward savings and investing goals, delivered as a dashboard and a PDF.

That post says the bot then asks about goals and can set up automatic investing in Coinbase. It prepares the buy, the user taps Approve, and only then is the order placed. He wrote that the bot only reads the account by default and that no money moves without approval.

The screenshot is a sample report that he labels as example data. MCP, the Model Context Protocol, is a standard for connecting AI agents to outside services.

Roemmele is promoting his own bot, and this deck could not confirm details of the X Money connection from X or SpaceXSI.

Watch for X or SpaceXSI to announce the X Money connection for Grok Bot officially, with details on what data it can read and what safeguards apply. Users trying the bot will show whether the reports are accurate. Also watch for regulators and banks to weigh in as AI agents gain access to payment and investment accounts. Any official X Money documentation would clarify how the connection works.

This is a teaser from NVIDIA's official account with no details. A petaflop is a thousand trillion operations per second, and NVIDIA has used that figure for compact AI computers before.

On October 21, NVIDIA posted a video with the caption 1 petaflop.

The 23 second clip shows chief executive Jensen Huang in his black leather jacket on a dark stage, holding a thin black device as he speaks. Beside him, a row of thin silver and white devices stands upright on a table.

The post does not name a product, give a date or say what the devices are, and the clip does not show any specifications. It is not clear from the post whether the footage is new or from an earlier event.

A petaflop of AI computing is a level NVIDIA has advertised before for small desktop systems built for AI developers, typically quoted at low precision. Teasers like this one often come ahead of launches, but readers should not assume a specific product until NVIDIA announces one.

The post drew attention mainly because of NVIDIA's reach, and it is ranked low here because it contains no information beyond the figure.

Watch for NVIDIA to name the product behind the teaser, its price and when it ships, and for the precision at which the 1 petaflop figure is measured. Also watch whether the devices in the clip are new laptops or desktop AI computers, and how they compare with Apple and AMD machines aimed at AI developers. Until then, treat the teaser as marketing.