Grok Bot is an AI assistant product whose bots can work across chat apps, documents and connected tools. The official Grok Bot account announced version 1.0 on Nov 11, and the spacexai account and Elon Musk shared the launch. Roman Ugarte, who described it as our release, listed what changed.

The official Grok Bot account introduced Grok Bot 1.0, describing it as AI teammates that work alongside a whole team. According to the post, Bots can now be shared, collaborate with people in Slack and Teams, be invited to meetings, and learn over time to become more helpful. The announcement drew more than 41 million views by the time of this scan.

Roman Ugarte, writing that the goal of the release was to make Bots a much more useful part of a team, listed three main changes. First, Bots can join Zoom, Google Meet and other meeting tools, where people can ask them questions and assign them action items. He said his team has its Bot take standup notes every day in a shared Google Doc, and that live back and forth conversation in meetings is coming soon.

Second, a new interface lets users see and steer each Bot's responsibilities, what it is learning and its impact. Ugarte said an AI teammate should flag when it is blocked or confused and summarize its contributions. Third, he said Bots in Slack and Teams now carry context across conversations and learn from a team over time.

Watch for how teams use shared Bots in practice, especially in meetings, and for the promised live voice conversation inside calls. The launch puts Grok Bot in direct competition with workplace agents from OpenAI, Anthropic, Microsoft and Google, which are also pushing assistants into Slack, Teams and meetings. Pricing, admin controls and how well Bots handle permissions across a company will decide how far this spreads beyond early adopters.

Google announced the Gemini Decisions API on Nov 11, available in Google AI Studio and through the Gemini API. It runs on a new model called Gemini Decision 1, which checks text, images, audio, video and PDFs against a list of questions and returns confidence scores instead of a written answer. Google AI Studio posted the launch, and the main Google account explained how it works in a short thread.

Google introduced the Gemini Decisions API on Nov 11, powered by a new model it calls Gemini Decision 1. According to Google, the model does not generate a response token by token. It evaluates an input, which can be text, images, audio, video or PDFs, against up to 1,000 questions in a single pass and returns decisions with calibrated confidence scores. Google says no output tokens are billed.

In its thread, Google says the API returns results in 150 to 250 milliseconds end to end and costs $0.10 per 1M input tokens. It claims this makes it nearly 5 times more cost efficient across a shared benchmark suite than its lightest Gemini 3.5 Flash model used through the Interactions API.

The idea is that developers give the model a checklist of yes or no, multiple choice and rating questions rather than asking it to describe something. In Google's example, Gemini Decision 1 scans 100 delivery photos in 4.5 seconds, which Google says is nearly 4 times faster than that same Flash model, and answers 16 questions per photo, such as whether a box is damaged, whether it sits under a porch, a delivery quality rating from 0 to 3, and what should happen next, such as routing to a claim review.

Google AI Studio said the API is available to try now in AI Studio and the Gemini API, and it shared demo apps named Doorstep Parcels, Speed Sketch and Find the Zookeeper. The speed and cost figures come from Google and have not been independently tested.

Developers will now test whether the confidence scores are reliable enough to drive real decisions, such as moderation, inspection or routing inside agents. Independent comparisons against using a normal model with structured output will show whether the claimed speed and cost gains hold up. If it works, other labs may ship similar decision models, and agent builders could start splitting work between models that write and models that only judge. Watch for pricing details, rate limits and early customer case studies.

Earlier in the week Jensen Huang told Nvidia employees he had an affair with Alice Huang, a former Megaspeed executive whom US officials suspect of spying for China, and Nvidia's board opened an independent review. On November 11 The New York Times reported a second relationship, with an Nvidia vice president. Hours later the Wall Street Journal reported that the board was preparing to remove him.

The Wall Street Journal posted on the evening of November 11 that Nvidia's board is preparing to oust CEO Jensen Huang after he had two extramarital affairs, including one with a suspected Chinese spy. This is a news report citing the paper's own reporting, not an announcement from Nvidia, and no decision had been announced when this deck was finalized.

The report came less than an hour after posts relaying a New York Times story. Pop Base wrote that the Times reported Huang had another affair with an executive at the company, and Joe Weisenthal wrote that the Times said the relationship was with Nvidia's vice president of health care.

Earlier in the week, according to X's news summaries of the coverage, Huang admitted in a company email to an affair with Alice Huang, a former executive at the Nvidia customer Megaspeed whom US officials suspect of working for Chinese intelligence. The summaries say Nvidia's board started an independent review, said it saw no evidence that secrets were compromised, and initially expressed confidence in his leadership. Republican lawmakers had also asked Huang for a closed door briefing on November 18.

Additional reporting, secondhand: the latest X news summary says the board met late Wednesday to consider removing him and that shares fell about 8% in after hours trading. The Kobeissi Letter, a market commentary account, publicly urged the board not to fire Huang and to consult shareholders first. These details come from summaries and commentary, not from Nvidia.

The key signal is an official statement from Nvidia's board on whether Huang stays, leaves, or steps aside while the review continues, and who would lead the company if he goes. Watch Nvidia's share price, any response from Huang himself, and whether large shareholders push for a vote as some investors are demanding. The November 18 briefing requested by lawmakers could also bring new details on the national security questions around the first relationship.

Tesla's official Europe, Middle East and Africa account posted that approval has now come through in Finland and that rollout begins soon, with a video of a Tesla driving while the driver's hands rest in his lap. Elon Musk later quoted that post and wrote that Tesla self driving is now approved in Finland. The posts do not name the Finnish regulator, the software version, or a start date for the rollout.

Tesla's official Europe, Middle East and Africa account posted that the product is now approved in Finland and that rollout begins soon. The post itself is two short lines and a video. Elon Musk later quoted it and wrote, Tesla self driving now approved in Finland, which is the clearest official wording that the approval is for Tesla's self driving software.

The video thumbnail shows the view from inside a Tesla, with the driver's hands in his lap and a driving visualization on the center screen, as the car travels a tree lined road. The posts do not say whether that footage was filmed in Finland.

Neither Tesla Europe nor Musk named the Finnish regulator, the software version being approved, or a date when customers in Finland will get the rollout. The phrase rollout begins soon is the only timing given.

This is a geographic and regulatory expansion of an existing product, not the launch of a new driving model. The lead is the official regional Tesla account, with Musk's quote kept as supporting context.

Watch Tesla's official channels for a start date in Finland and for any note on which software version is included in the approval.

Later posts may also say whether this is a supervised feature or a broader permit, and whether nearby countries are next. Until Tesla or Finnish authorities give those details, the news is the approval itself and a rollout that the company says will begin soon.

Daniel Lenton of Unify posted that OpenClaw, Hermes Agent and Prime Agent do not meaningfully learn from experience. On Continual ARC, Unify's new benchmark, he said none of them were able to memorize anything of value and none were able to generalize to new tasks, while training a small 9 billion parameter model's weights did both. Unify also published a benchmark page, a blog post, and a paper describing the test.

Daniel Lenton, posting about work at Unify, said OpenClaw, Hermes Agent and Prime Agent do not meaningfully learn from experience. On Continual ARC, which he called Unify's new benchmark, he said none of those agents were able to memorize anything of value and none were able to generalize to new tasks. He said training a small 9 billion parameter model's weights did both.

Unify's benchmark page describes Continual ARC as a test of skill acquisition over a working lifetime, credited to Daniel Lenton, Yusha Arif and Haris Mahmood. According to that page and Unify's blog, the benchmark serves a long stream of ARC puzzles in which the same tasks return with fresh inputs while new tasks keep arriving. After the stream, the learner is frozen and tested on fresh instances of tasks it already met, which is meant to measure memorization, and on tasks it never met, which is meant to measure generalization.

The blog says a 9 billion parameter model's weights, trained to imitate a stronger model, improved on both the puzzles it had seen and a set of new ones, while note taking harnesses did not show that kind of gain. Those details come from Unify's own writeups, not from the lead post, which named the agents and the headline result but did not list every score.

The result is Unify's own evaluation of its benchmark. Other teams have not yet been cited here as having confirmed the same ranking on a public leaderboard.

Watch for other labs to run the public Continual ARC stream and for Unify to publish verified results on its private held out set, which the company has invited outsiders to submit to.

The open question is whether later agent harnesses can close the gap that Lenton reported, or whether teams will need to train model weights to get lasting memory and transfer. Until independent runs appear, these figures should be read as Unify's results on its own benchmark.

Anthropic is preparing for its IPO, and sell side analysts are the bank researchers who write about companies for investors. Jukan, an account that posts chip and AI industry notes, shared what it described as remarks Anthropic made in a meeting with the sell side. Anthropic has not publicly confirmed the remarks, so treat them as unverified.

Jukan posted a list of what it called interesting remarks Anthropic made in a meeting with sell side analysts. The post is secondhand, gives no source document, and has not been confirmed by Anthropic.

According to the post, Anthropic said it has stability and better talent than open source rivals, and that the AI market is winner take all, adding that as long as it executes properly it will take all of it. The post also says Anthropic told analysts its contract with SpaceX will not be extended beyond the middle of next year, and that developing Opus 5.5 cost more than $10 billion.

The note spread widely, passing 790,000 views. Aaron Burnett pushed back, arguing that critics were reading a secondhand meeting note without context, and that the SpaceX comment was likely an answer to an analyst question about being stuck with above market compute rates if demand slowed.

Elon Musk replied to Burnett that SpaceX would be happy to take the compute back, adding that he thinks the opposite is going to happen.

Anthropic's IPO filings and roadshow materials are the place to check these claims, especially any disclosure of compute contracts and model development costs. If the SpaceX deal does end in mid 2027, watch where Anthropic moves that capacity. Musk's reply suggests he expects Anthropic to want more SpaceX compute, not less, so any extension or new deal would be a clear signal either way.

Axios reported on Nov 11 that President Trump authorized a crackdown on companies suspected of diverting Nvidia's advanced AI chips to China, and that Commerce Secretary Howard Lutnick repeatedly stood in the way. Earlier that day, Axios reported that Treasury Secretary Bessent confronted Lutnick in the White House Situation Room over why Megaspeed, a company accused of diverting US chips to China, had not been blacklisted. Reuters also reported that a Chinese university linked to the People's Liberation Army bought Nvidia H200 chips despite being on a US export blacklist.

Axios reported in a scoop posted on Nov 11 that President Trump authorized a crackdown on companies suspected of diverting Nvidia's advanced AI chips to China, but that one administration official repeatedly stood in the way: Commerce Secretary Howard Lutnick. Reuters picked up the report, summarizing it as Lutnick stalling the crackdown.

The report followed a separate Axios scoop the same day about a confrontation in the White House Situation Room. According to posts citing Axios, Treasury Secretary Bessent asked Lutnick why Megaspeed, a company accused of illegally diverting US chips to China, had not been blacklisted, and Lutnick answered that he did not know and would look into it. Those posts say officials had warned Lutnick about Megaspeed for months.

The reports landed alongside other evidence of gaps in chip controls. Reuters reported that a Chinese university linked to the People's Liberation Army bought Nvidia's H200 AI chips despite being on a US export blacklist.

Chris McGuire, a commentator on export controls, called the Axios report stunning and wrote that the President reportedly ordered the crackdown three separate times. He argued that Lutnick and Jeff Kessler were defying those orders. These are his claims and characterizations, not confirmed findings.

The Commerce Department oversees US export controls on advanced chips, which makes its enforcement choices central to how many Nvidia chips reach Chinese buyers through other countries. The posts do not include a response from Lutnick or the Commerce Department.

Watch for a response from Lutnick or the Commerce Department, and for whether Megaspeed or other suspected diverters are added to the export blacklist in the coming days. Congress may use the reports to push for stricter chip control laws or hearings. Any visible crackdown on transshipment through Southeast Asia could hit Nvidia sales channels, while continued inaction would fuel criticism that the controls are not working. The story also runs alongside separate pressure on Nvidia's leadership, which could shape how Washington deals with the company.

AI companies are racing to build large data centers, and local fights over power prices, water and land use have grown. Earlier in the week, a Maryland activist helped block a planned $5 billion AI data center, according to Polymarket. Unusual Whales reported Hawley's position, quoting a spokesperson.

Unusual Whales reported that Senator Josh Hawley has endorsed a temporary pause on data center construction, which it said makes him the highest ranking Republican to back a nationwide moratorium.

The post quotes a Hawley spokesperson as saying the senator is calling for a federal data center moratorium until Congress passes a bill of rights giving taxpayers guarantees of no rate increases. The concern behind it is that the power demands of large data centers could push up electricity bills for ordinary households.

The post does not include a bill number or say whether Hawley plans to introduce legislation himself. It also does not say how long the pause would last or whether it would cover projects already under construction.

The report comes as data center expansion faces local opposition. Polymarket posted earlier in the week that a Maryland activist helped block a $5 billion AI data center planned for the site of an abandoned shopping mall after gathering more than 25,000 signatures, a post that drew millions of views.

Watch for whether Hawley turns the position into a bill and whether other Republicans or Democrats sign on. A federal moratorium would face strong opposition from AI companies and from officials who see data centers as key to competing with China, so the more likely near term outcome is pressure for rules that protect households from higher power bills. Utility regulators and state legislatures may act sooner than Congress.

Robotaxi is Tesla's ride hailing service that uses its self driving software. Earlier on Nov 11, Tesla said its self driving system was approved in Finland. Elon Musk quoted the Robotaxi post.

The official Robotaxi account posted that Robotaxi now runs 24/7 in every city where the service is available, with an attached video. The post was published late on Nov 11 Eastern time.

Elon Musk quoted the announcement with a short comment saying that Robotaxi works 168 hours a week, the number of hours in a full week. By the time of this scan, his quote post had drawn more views than the original.

The post does not list which cities are covered, how many vehicles are in service, or whether safety monitors are still present in the cars. It also does not say whether overnight service carries any limits on pickup areas or ride types.

Running through the night matters for a ride hailing network because late hours are when many people lack other transport options, and because a vehicle that does not need a driver can in principle work far more hours than a human driver. The announcement came on the same day Tesla said its self driving system was approved in Finland.

Look for Tesla to share ride numbers, fleet size and safety data for the round the clock service, and for any expansion to new cities. Regulators in places where Robotaxi operates may also comment on overnight driverless rides. Comparisons with other driverless taxi operators on hours, coverage and safety are likely to follow, and the round the clock claim will be easy for riders to test city by city in the coming days.

Finances is a ChatGPT feature for looking at spending and investments. On November 11 the official ChatGPT account said it is now available in the desktop app. The post says it is open to Free, Go, Plus and Pro users in the US.

The official ChatGPT account posted on November 11 that Finances is now in the desktop app. According to the post, users can find ways to save, understand their spending and make sense of their investments, and they can also build custom financial tools.

The post says the feature is available to Free, Go, Plus and Pro users in the US, and that people need to update the desktop app to get started. That means the feature is not limited to paid plans, but it is limited to the United States for now.

The demo video shows a desktop window titled Build finance desktop widgets, where a user asks ChatGPT to check Finances and build desktop widgets for spending, investing and progress toward college savings goals. This suggests the custom tools are small apps or widgets that ChatGPT builds from the user's own financial data, though the post itself does not detail how accounts are connected or what data is stored.

The announcement drew more than 250,000 views in under a day. OpenAI did not give details in the post about pricing changes, privacy controls or a timeline for other countries.

Watch for OpenAI to publish help pages that explain how accounts are linked, what data Finances keeps, and which institutions are supported. Expansion beyond the US and any limits by plan would show how big a bet this is. Reactions from banks and personal finance apps, and any privacy questions raised by regulators, are the other signals worth following closely.

Polymarket posted that United Airlines will ban personal AI agents from booking flights on behalf of customers. The post does not name a source, so this lead is secondhand. Reuters reported the same policy on November 11, and United's own help page says it is restricting unauthorized third party AI agents from its digital platforms and from acting through travel agencies.

Polymarket posted on November 11 that United Airlines is to ban personal AI agents from booking flights on behalf of customers. The post itself does not name a source.

Reuters reported the same day that United said it would restrict AI bots' access to its digital platforms, citing data security. United's own help page confirms the policy. It says third party AI agents can act on a customer's behalf, and that giving them access to a MileagePlus account, bookings, or extras and upgrades could expose personal information or lead to purchases, account changes, or use of miles without approval. As a result, United says it is taking steps to restrict access, both directly in its digital platforms and indirectly through travel agencies and aggregators, and to prohibit use by unauthorized third party AI agents or assistants.

United also warned that automated tools that check a passenger in for a flight, a step that requires a person to acknowledge rules on prohibited items, could create regulatory risks. The page says bookings made by a third party AI agent before November 11 will not be canceled, and that United does not yet work with any third party AI agents.

This lead is secondhand, through Polymarket. The underlying policy is stated on United's own site and was also reported by Reuters.

Watch whether other airlines adopt a similar ban, and whether United later allows approved integrations. United said it will update its page if it begins working with third party AI agents. Existing bookings from before November 11 are supposed to stand, so the near term test is whether the restriction is enforced on new bookings and through agencies. United has also advised customers who already used an agent to review those bookings.

Gopuff is a US delivery company that runs its own network of local warehouses. Grok Bot is an AI assistant that can work with connected services. The same day, Grok Bot announced its 1.0 release.

Gopuff announced that its service is now built directly into Grok Bot, telling users to ask the bot and get it delivered. The post said the integration uses Gopuff's own inventory and delivery network and can bring orders to a customer's door in as fast as 15 minutes, and that it is live now.

The announcement passed 2 million views, making it one of the most viewed AI commerce posts of the day. It came a few hours before Grok Bot announced its 1.0 release, which focused on shared bots for teams.

The post does not explain how payments, delivery areas or order confirmation work inside Grok Bot, or whether users must link an existing Gopuff account. It also does not say whether the integration is available in every market Gopuff serves.

Direct integrations like this one turn an assistant into a place to complete purchases. Several AI companies have been adding shopping and checkout features, and retailers are deciding which assistants to plug into first.

Watch for more retail and delivery partners to announce Grok Bot integrations, and for data on how many orders come through assistants. Questions about how the bot confirms orders and handles mistakes will matter as more people let AI place purchases for them. Rival assistants are likely to keep signing similar commerce deals, so the number of partners each assistant can offer may become a selling point of its own.

Antigravity is Google's agent based coding environment. On Nov 10 it added /swarm for running teams of agents in parallel, alongside the Gemini 4 Carbon launch.

Google Antigravity posted that browser use has been rolling out in the product, describing it as closing the loop between researching, writing code and verifying it live.

According to the post, users choose which sites the agent can log into, downloads pause until the user approves them, and the agent can use the browser to run multiple tasks at once. Google invited users to try it the same day.

The post does not say which plans include the feature, whether it works in both the desktop app and the command line tool, or which models power it. It also gives no figures on reliability.

Browser access lets a coding agent open documentation, test a web app it just built and check that pages render correctly, tasks that developers would otherwise do by hand. The approval step for downloads and the user controlled logins are safety limits meant to stop an agent from taking unwanted actions on a user's accounts. The update follows the /swarm feature Antigravity added a day earlier.

Expect developers to compare Antigravity's browser agent with similar features in Cursor, OpenAI Codex and Claude, especially on how well it tests web apps and how often it needs human approval. Google may publish documentation on permissions and limits. Security researchers will also likely probe how the agent handles malicious web pages, a known risk for any agent that browses on its own.

Zero Hedge shared a series of charts it attributed to Bernstein, a research firm, mapping who is actually building US data centers. Gigawatts measure the power capacity a data center campus is planned to draw.

Zero Hedge posted a set of charts it attributed to Bernstein under the title Gigawatts vs. Bragawatts, saying 586GW of US data center capacity has been announced but only 199GW looks real.

According to the breakdown credited to Bernstein, 75GW of 87GW under construction is rated credible, as is 123GW of 497GW that is only announced. The most viewed chart ranks companies by credible pipeline: Amazon 33GW, Google 20GW, Meta 15GW, Microsoft 10GW, QTS 9GW, Stack Infrastructure 7GW, Vantage 6GW and Digital Realty 4GW.

Another chart said hyperscalers hold 42% of the credible pipeline and colocation providers 34%, with crypto miners that pivoted and neoclouds at 4% each. A separate chart ranked Meta, Microsoft, Google, QTS and Vantage as the top five by the share of their own pipeline rated credible.

Elon Musk replied that SpaceX has a chance of reaching about half that gigawatt figure next year in actually usable compute, saying it takes his team about 2 months working around the clock to get a large cluster running reliably and guessing about 6 months for others.

Watch for Bernstein's full report and for companies to answer questions about which announced projects will go ahead. The gap between announced and credible capacity will shape forecasts for chip demand, power plant orders and utility rates. Musk's claim about SpaceX compute next year is an estimate, and SpaceX's actual cluster announcements will show whether it holds. Political pressure on data centers, such as calls for a moratorium, could shrink the credible pipeline further.

Ramp is a corporate card and spend management company, and its AI Index tracks business spending on AI tools from its customers' transactions. Ara Kharazian is the economist who posts the index.

Ara Kharazian posted the October reading of the Ramp AI Index, saying business AI spend rose 2% in the month. He wrote that growth for both OpenAI and Anthropic fell to low single digits.

According to the post, engineering teams drove 72% of AI spending, followed by customer service and leadership. That pattern lines up with the heavy use of coding tools and agents inside companies.

Kharazian wrote that given this relatively modest growth in spend, and new competitive dynamics that may stick around, the market should look to new fronts in the AI trade. The post did not name which new fronts he meant.

The index reflects spending by businesses that use Ramp, so it captures a slice of the US corporate market rather than all AI revenue. It does not include consumer subscriptions or large enterprise contracts paid outside Ramp cards. Even so, it is one of the few regular public measures of how quickly companies are paying for AI tools.

Next month's index will show whether October was a pause or the start of a slower trend. Watch for whether OpenAI and Anthropic reports of revenue growth match what Ramp sees, especially as Anthropic heads into its IPO. A shift in spending toward other vendors, such as Google or newer agent companies, would be the competitive change Kharazian hints at.

Jan Senderek, who says he sold his first company Loom to Dropbox in 2014, announced Roma 2.0, a YC F26 company he described as the first to do list for AI agents. He said Roma automatically creates tasks and hands them to the right agent, so a person and their agents can work together with shared context, and he pointed people to roma.app. He named Claude, Grok Bot and Muse as examples of agents that, until now, could not work together.

Jan Senderek posted that in 2014 he sold his first company, Loom, to Dropbox, and that he is now launching Roma 2.0, a YC F26 company. He called it the first to do list for AI agents.

According to the post, people use different agents, and he named Claude, Grok Bot and Muse as examples. He said that until now none of them could work together. Roma, he wrote, automatically creates tasks and hands them to the right agent, so a person and their agents can work together with shared context. The post points to roma.app.

The announcement is a product launch from the founder. It does not include pricing, supported agent counts, or independent usage figures. The claim that this is the first to do list for AI agents is Senderek's own wording in the post.

This is an early startup release, not a change in what the underlying models can do. It sits in the deck as a new tool for coordinating agents, not as a research or model result.

The next things to watch are whether Roma opens widely, which agents it actually supports beyond the names in the launch post, and whether teams start using a shared list as the way they coordinate model tools.

Senderek did not give a public timetable for later features. Until those details arrive from the company, the product should be judged on what users can do at roma.app today.

Marsh is a large insurance broker. Banks must hold capital against the loans they make, and insurance on credit risk can lower how much they need. Walter Bloomberg, posting as DeItaone, summarized the report.

Walter Bloomberg, posting as DeItaone, reported that Marsh is lobbying US regulators to reduce capital requirements on bank loans protected by credit risk insurance, as lending for AI data centers surges.

According to the post, Barclays is seeking insurance on more than $10 billion in data center debt, and HSBC is exploring similar deals. The post said global data center spending is projected to reach $6.7 trillion by 2030, and that lenders are looking for ways to offload AI credit risk. It said a Federal Reserve decision could come by the end of the year.

The post does not name the source of the $6.7 trillion projection or describe the exact rule change Marsh wants. It also does not say whether other insurers or banks back the request.

The effort matters because AI data centers are being financed more and more with debt, not only with cash from large tech companies. If regulators let banks hold less capital against insured loans, lending could expand, while critics may worry about risk moving into less visible corners of the financial system.

Watch for a Federal Reserve decision by year end and for comments from other regulators. More banks may seek insurance on data center loans, and analysts will track how much AI debt sits with banks versus private credit. Any sign of trouble at a large data center borrower would put this kind of risk transfer under sharper scrutiny, and could slow new lending for AI construction.

Polymarket posted that an OpenAI engineer said he had Codex running for 26 days straight searching for the lost city of Atlantis, and described this as vibe archaeology. The first hand post is from pash, who lists Codex work at OpenAI in his profile and wrote that he had Codex running for 26 days searching for Atlantis, with his best bet for now the Faraoe plateau. That original post was under the view floor, so the more widely seen Polymarket summary is the lead, and the story is secondhand.

Polymarket posted that an OpenAI engineer revealed he had Codex running for 26 days straight searching for the lost city of Atlantis, and that vibe archaeology is emerging as a new use for AI agents. The post includes an OpenAI logo and an illustrated underwater city. It does not name the engineer.

The first hand post is from pash, posting as pashmerepat, whose profile describes Codex work at OpenAI. He wrote that he had Codex running for 26 days searching for Atlantis, and that the best bet right now is the Faraoe plateau, but that it needs more time. That post quoted an earlier Polymarket item about a different archaeology themed agent project, and it had not crossed the view floor at scan time, so it is not the lead.

Polymarket's wording goes a bit beyond pash's post: pash did not use the phrase vibe archaeology in that text, and he did not present a finished finding, only a lead that he said still needed more time. OpenAI itself has not been cited here as commenting on the run.

This is a curiosity item about a long running agent session, not a scientific claim about Atlantis and not a product announcement from OpenAI.

If pash or OpenAI later share methods, sources, or a result, that would be the place to follow up. Until then, the story is a secondhand, widely seen summary of one engineer's own post about a weeks long Codex run.

There is no timetable attached. The item is in the deck because it traveled, not because it changes what Codex is for.

Dylan Patel runs SemiAnalysis, a research firm that covers chips and AI infrastructure. Persistent agents are AI agents that keep running and working over long periods instead of answering one prompt. Perplexity CEO Aravind Srinivas responded to the post.

Dylan Patel posted that persistent agents are the next unlock of AI usage, and that his firm's spending on AI is now 45% persistent agents.

He gave blunt reviews of several products. He wrote that Claude Tags from Anthropic is bad and that OpenAI's Slack agent is even worse. He said his team primarily uses Perplexity Computer, which he noted uses many different models, and praised Perplexity's agent harness.

These are the opinions of one heavy user, not a benchmark, and Patel did not share how he measured spending or which tasks the agents handle. The 45% figure covers his own firm's AI budget only.

Aravind Srinivas, the CEO of Perplexity, quoted the post and said the name Perplexity Computer came from a Patel blog post that described coding agents as a new kind of computer.

The post lands as AI labs and startups compete to sell agents that stay on and work inside company tools, from Slack to browsers, rather than chat windows that wait for a prompt.

Watch for Anthropic and OpenAI to respond with updates to their workplace agents, and for more companies to share what share of their AI budget goes to agents. If persistent agents keep taking a larger share of spending, the race among labs will shift further toward harnesses, tool use and reliability over long runs, not just raw model scores. Perplexity is likely to use the post in its own marketing.

Mariko Mori is a Japanese artist whose retrospective is titled All That Shines. Anthropic's Claude account shared the project as a customer story.

The official Claude account posted that the artist Mariko Mori, who it said has pushed the boundaries of art for three decades, used Claude to write code for a computer generated animation in her retrospective All That Shines.

According to the post, the animation is presented as a new video installation in the show. The post links to more details but does not describe how much of the code Claude wrote, what tools rendered the animation, or how Mori worked with the model.

The post is a promotional story from Anthropic rather than an independent report, and it does not give the venue or dates of the retrospective.

It is one of several recent examples of AI labs highlighting creative uses of their models. Here the model is used as a programming aid that turns an artist's ideas into generative code, rather than as an image generator, which is a different and often less contested role for AI in art.

Look for coverage of the exhibition itself and for Mori to explain how she used Claude in the process. More artists may share similar workflows where AI writes the code behind generative works. How audiences and critics respond will add to the wider debate about where AI fits in creative practice. AI labs will likely keep promoting stories like this one as they try to show uses beyond office work.