This is promotion by Musk, whose company makes Grok Bot, in quote posts of user accounts. Grok Bot's documentation says the bot should hand control back to the user for payment confirmations and that its secure secret requests keep values out of the transcript, and it does not mention card screenshots.

On the evening of October 17, Elon Musk quoted a post from Teslaconomics and wrote: This works well. Screenshot of credit card is the easiest way, unless you want to fill out the secure vault data. He added that it seems a bit silly to hand your credit card to people you don't know at a restaurant, but not to Grok Bot.

Teslaconomics had posted that he sent his Grok Bot one text asking it to order the DribbleUp silent ball at the best price, and that it compared prices, checked out with his card and placed the order for $32.91 with free shipping. His screenshots show the bot saying DribbleUp's own site had the best price at $29.99, an order confirmation and a receipt. The receipt also lists a DribbleUp Membership whose charges start when a free trial ends.

Minutes later, Musk quoted Mark Kretschmann's post praising how much Grok Bot has improved and wrote that his team makes around 20 improvements per day to Grok Bot. He also replied Sure to Texas Twins Dad, who wrote that he should probably start a Grok Bot class because many people still don't understand what it can do. That post passed 7 million views.

Grok Bot's documentation says the bot should hand users control for passwords, codes and payment confirmations, warns against sending passwords in ordinary chat, and says its secure secret requests mask values from the transcript and the model.

Watch whether xAI adds a dedicated payment feature, as the secure vault Musk mentioned suggests, and how it protects card data that users send as images.

Also watch for reports of unwanted charges, such as trial memberships added at checkout, as more people let agents shop. Retailers' policies on agent purchases and the pace of the 20 daily improvements Musk described will shape how reliable the shopping feature becomes.

This is a reply from Musk, whose company makes Grok Bot, to a user asking how to turn off training on his data. It does not describe a setting, while Grok Bot's documentation says training opt out follows the user's Cursor account privacy settings.

On October 17, Simon Smith, who works on generative AI at Klick Health, posted that he was trying Grok Bot as part of evaluating personal agents. He wrote that the first thing he noticed was that he could not find a setting to turn off training on his data, that his bot did not know either, and asked how to control the use of his data.

Elon Musk replied that unless you are doing something super unique, your data is not actually useful for training, but that xAI does need some way to debug issues and make the product better. He added: We don't have ESP!

The reply does not say whether user data is used for training or how a user can opt out. Grok Bot's documentation says the bot uses Cursor authentication and account data settings, that it requires data storage and does not support Cursor's Legacy Privacy Mode, and that training opt out follows the applicable Cursor account and privacy settings.

Earlier the same day, Musk replied to another user that Grok Bot will not access anything unless specifically asked, in response to a request for a way to block it from certain X chats.

Watch whether xAI adds a training control inside Grok Bot itself, rather than relying on Cursor account settings that new users may not know about.

Also watch how xAI explains what data it keeps for debugging and for how long, since users are connecting email, banking and messaging accounts. Privacy questions like Smith's may grow as more professionals test Grok Bot for work.

This is a first hand claim from Roemmele, and none of it has been independently verified. He did not name the business or share evidence of the revenue, and Elon Musk, whose company makes Grok Bot, reposted it.

On October 17, Brian Roemmele posted that he started a new business with Grok Bot a few weeks ago and that it has been absolutely amazing.

He wrote that it began with a question about a problem he had encoding video and audio for AI, and that the bot brainstormed him into a full new venture that it is now running. He said it builds all aspects of the business with the same gut instinct and drive to succeed that he has.

Roemmele wrote that Grok Bot has already built seven positions that it fills as an employee, reporting to itself and to him, that it handles all customers by email and has solved complex problems, and that it just brought in $700 without them trying hard. He said he would write more soon. The attached image shows a figure with a round white head in a suit holding a briefcase.

Elon Musk reposted the post early on October 18. The next day, Roemmele shared a separate Grok Bot that phones restaurants to book tables, which Musk also quoted.

Roemmele did not say what the business sells, how the seven positions are set up, or how much he reviews the bot's work. Grok Bot's documentation says the bot should hand control to users for payment confirmations.

Watch whether Roemmele publishes the details he promised, including what the business sells, how the roles are divided and how revenue is tracked.

Also watch for more accounts of people running small businesses with agents, and for problems such as customer complaints or errors in automated email. Claims of agent run companies are spreading quickly, so evidence of results will matter more than early revenue figures.

bubble boi is sharing a photo of a Financial Times print profile without commenting on its content. The details below come from the article as shown in the photo, and the succession talk is described by the FT as touting, not a decision by Meta.

On October 17, bubble boi posted a photo of a Financial Times print article about Alexandr Wang, calling it a really good article.

The article, by Rafe Rosner Uddin in the FT's Person in the News column, is headlined Meta's AI wunderkind joins the big league. Its standfirst says the head of Superintelligence Labs is being touted as a potential successor to Zuckerberg.

As shown in the photo, the article describes Wang as the 29 year old head of Meta's Superintelligence Labs, whose big hiring came through a $14.3bn transaction involving his data labelling company Scale AI. It says that, according to Mark Zuckerberg, the release of Muse Wonder has returned Meta to the frontier of AI, five weeks after Meta's Muse AI agent became the most downloaded free app in the Apple and Android app stores. It adds that Wang announced on Thursday that Meta would partner with Apple to deploy its open weight models on Mac computers.

The article also recounts Wang's background, including founding Scale AI in 2016 after dropping out of MIT, and quotes people describing him as an operator. It notes controversies at Scale, including a lawsuit alleging it required staff to evaluate extreme material as part of its data labelling work, which was later voluntarily dismissed, and that Zuckerberg has maintained Wang is the right person to lead Meta's AI effort.

Fortune published its own profile of Wang's relationship with Zuckerberg on October 17.

Watch whether Meta formalizes any change in Wang's role or responsibilities, since the FT describes the succession talk as touting rather than a decision.

Also watch the Meta and Apple partnership on running Meta's open weight models on Macs, and how Muse and Muse Wonder hold up against rivals as sites block some agents. Meta's next earnings call may show how investors weigh its AI spending under Wang.

This is a first hand opinion from Sweeney, the founder and chief executive of Epic Games, which sells tools for game developers. It is his recollection of his own work and his view of AI, not a measured study.

On October 17, Tim Sweeney wrote that when he was writing Unreal Engine 1, about 30% of his time went to seriously deep coding on hard problems, such as assembly language texture mappers, C++ BSP trees, data persistence, object oriented systems design and networking.

The other 70%, he wrote, was just typing and refactoring large bodies of code needed for a complete engine feature set. He said it was not hard, his mind was mostly on the deeper problems while he did it, and it didn't feel much different from vibe coding today, except that he was typing on an IBM Model M keyboard rather than prompting and reviewing the results.

Sweeney said it always seemed logical to rely on AI for that 70% once it could match human quality, which has never been perfection, and that it came as no surprise when it took over that work. What surprised him more, he wrote, is AI emerging as a sort of instant John Carmack, able to solve hard technical problems cleverly and work in complex 3D scenes, both by operating tools and by reading raw data files.

He compared the change to his dropping assembly language in 1998, when newer processors made simpler C and C++ code faster, and wrote that what matters is the end product. Anything that lets developers try more ideas and make better games faster, he said, is a win for everyone. Elon Musk reposted it, and it passed 1.3 million views.

Watch how game developers respond, since many artists and programmers have pushed back on AI in games and on Epic's support for it.

Also watch what Epic ships for Unreal Engine, whose official plugin for Claude Code spread widely on X this week, and whether Sweeney's view that AI can handle hard engine problems holds up in real projects. Debates over labeling AI use in games are likely to continue.

This is an opinion post from Hungrybox, a professional Super Smash Bros Melee player. The statistics he cites, such as how much code at Google is written by AI, came from colleagues and are not sourced.

On October 17, Hungrybox, who describes himself as a three time number one Super Smash Bros Melee player in the world and a co owner of Team Liquid, asked his peers to weigh in on LLMs in gaming. He quoted a post from the indie developer walz showing Melee remade as a 3D arena fighter.

Hungrybox wrote that barely a week earlier he had distanced himself from a vibe coded project made by a friend in the Smash community and said AI has no place in gaming. Since then, he said, colleagues and full time software developers educated him on the topic and he now feels dumb for his earlier view.

He wrote that he was told it is now the norm to do 90% of coding with AI assistance, and that 75% of all new code at Google is made by AI and approved by engineers. He said he had seen Epic Games, Xbox and other publishers embrace AI, and that he had been sent about a dozen projects that rework Melee, including one that runs at 120 frames per second and one that closely simulates a CRT screen. Some, he said, are low effort slop, while others seem genuinely useful for practice.

He asked whether AI in game development is a bastardized amalgamation of other people's ideas, given that games are art, or a necessary tool because games are software, and said he wants open discussion. The post passed 690,000 views.

Watch how the Smash community and other esports scenes respond to AI made fan projects, including whether tournament organizers or Nintendo take a position on them.

Also watch the wider fight over AI in games, as publishers such as Epic push AI tools and some developers and players push back. Questions about crediting original work and labeling AI use are likely to keep coming up.

This is commentary from the account Wise, which gives no sources. The Verge and TechCrunch reported most of the claims, but Amazon's block was confirmed for Meta's Muse, while reports that Grok and other agents cannot use Amazon come from users.

On the evening of October 17, the account Wise posted a list: Delta is blocking AI agents from booking flights, JPMorgan Chase and Charles Schwab block agents from accessing accounts, Yelp requires paid data licensing for agents, and Amazon has blocked Grok Bot and other shopping agents. It concluded that the internet is starting to block the agents that were supposed to do everything for us. The post passed 1.7 million views.

Most of the list matches reporting. The Verge reported on October 17 that Yelp told TechCrunch it only allows agents that pay for its data licensing program, that Delta is preventing agents from booking tickets for users, and that JPMorgan Chase and Charles Schwab confirmed they block agents from accounts over security concerns. TechCrunch quoted Delta saying it has no partnership or integration that lets a third party agent book flights on its platforms.

The Amazon claim is less settled. The Verge reported that Amazon said Meta's Muse agent violates its policies, and that there are reports of xAI's Grok, Instinct's agent and OpenAI's dots being unable to use Amazon. It did not report an Amazon statement about Grok Bot. The Verge added that Elon Musk responded by saying maybe he should start a competitor to Amazon.

The Verge also noted users report problems on Cars.com and Amtrak, and that anti bot systems can block agents even where sites have no formal ban.

Watch whether Amazon comments on Grok Bot specifically, and whether more companies sign deals with agent makers, as Meta has with retailers such as Walmart and Best Buy.

Also watch the court fight between Amazon and Perplexity over agent shopping, which could shape whether sites can legally block agents acting for users. Standards for agents to identify themselves to websites, and paid access programs like Yelp's, may decide how open the web stays to agents.

This is an opinion column by Charles Gasparino, published by the New York Post on the evening of October 17. It argues from the history of the 2000 dot com crash and does not present new figures or reporting.

On October 17, the New York Post posted a column by Charles Gasparino headlined: The AI bubble is about to burst, and a big market correction is coming.

Gasparino compares today's AI boom to the dot com mania that ended in March 2000, when, he writes, a historic run up in stocks was exposed as one of history's great investment bubbles and the tech market took years to recover. He argues that history repeats itself with market manias and that the current AI mania will almost certainly follow the same path, with losses for investors.

He writes that the most likely trigger is the instinct of hedge funds and institutions that do their own research, and that these investors now sense something is not quite right with AI. He says he cannot tell whether the prior week's market moves, which he describes as rising rates after a bad inflation report followed by a selloff in AI stocks, are the herd's final statement on AI.

Gasparino says there may be more froth to come, pointing to possible IPOs by Anthropic and possibly OpenAI. He argues interest rates will rise because AI competes with the federal government's borrowing needs, and that much of the AI spending uses debt, which he calls the main ingredient of any crash. He advises readers to take profits.

The column gives no valuation figures, and its claims about coming IPOs are not sourced.

Watch whether the AI selloff deepens or reverses in the coming weeks, which will test calls like Gasparino's against views such as Goldman's that hedge funds have cut AI exposure too far.

Also watch the debt markets that fund data centers, which Goldman's notes flagged as a weak point, and whether Anthropic or OpenAI file for public offerings. Nvidia's November earnings could be a key moment for AI valuations.

Mollick is sharing and commenting on a paper by other researchers. The paper collects reported cases and does not claim to prove that AI caused each rise in requests.

On October 17, Ethan Mollick wrote that a consequential choice will be made soon by governments and companies as they are increasingly flooded with agentic AI inquiries that overcome the frictions that gated access. Pointing to the graphs, he said it has already started, and asked whether they will make services worse or better as a result.

His images come from a paper by Chris Schmitz of the Hertie School, Lewis Hammond of the Cooperative AI Foundation and Alan Chan of the Centre for the Governance of AI, whose revised version was posted to arXiv on August 19. It collects 84 cases across 11 jurisdictions where government services saw rising volumes of requests, such as freedom of information requests and social court cases in Germany, that may involve AI. In 87% of cases the AI use was text generation by large language models, and in 69% officials themselves said AI was involved.

The paper's table sorts responses into two groups. Some suppress demand by adding friction or reducing the benefit, such as Australia considering fees for freedom of information requests. Others increase capacity by redesigning the service or deploying agents of their own. Mollick labeled the first group make it worse and the second make it better.

The authors say friction is quick to add but can hurt access for people with real needs. They recommend that governments audit their exposure, plan for digital identity and give legal certainty about AI made requests.

Watch whether more governments add fees, identity checks or other friction in response to AI made requests, and how courts and regulators judge those limits.

Also watch for governments deploying their own agents to handle requests, which the paper lists as the other path. As personal agents spread, companies with customer service and complaints systems may face the same choice Mollick describes.

These are first hand anecdotes from users and have not been verified. Neither post names the exact steps or the risks, such as voiding a warranty or breaking updates.

On October 17, 0xProbability posted that they vibe coded their TV after it showed an ad before they could even open Netflix. The post quoted Ryan Doyle's October 16 post about hacking his TV with Claude Code.

0xProbability wrote that they gave Claude Code the TV's model number, that it found the TV on the home network, removed 25 Xiaomi packages and built a new launcher, and that it is basically an Apple TV now. The attached video shows a home screen called luma with a large landscape image and a row of app tiles including YouTube, Netflix, SoundCloud and TED.

On October 18, Mihir Gupta posted that he saw people vibe coding their TVs and tried it because his had become unbearably slow. He wrote that in one afternoon with Claude Code he disabled 25 bloat apps and replaced the 192 MB stock launcher with a custom 83 MB one with rotating wallpapers. His before and after image shows a Google TV home screen labeled sponsored hero ad beside a cleaner dock of streaming apps.

Neither user explained exactly what they changed, whether it can be undone, or how it affects software updates, and the claims have not been independently checked.

Watch whether TV makers such as Xiaomi or Google respond, for example by restricting developer access or by offering ways to turn off ads and sponsored content.

Also watch whether users share the scripts or launchers they built, and whether problems such as broken updates or lost features appear over time. The trend may spread to other devices with locked down software, such as routers, cars and smart speakers.

This is a first hand post from Odyssey, an AI lab that builds world models. It has not published a paper, code or method, so it is not known which fly brain model was used or how the brain's activity steers the flight.

On October 17, Odyssey posted that it placed the fruit fly brain inside Odyssey 3, which it calls a frontier world model, and that the fly is now flying around totally imagined, simulated worlds all by itself.

The 30 second video shows scenes of a sunny harbor town, with boats, a waterfront promenade and stone buildings, moving as if seen in flight. An inset in the corner shows a 3D model of a fly, at first over a checkered floor and then against a white background. Odyssey describes itself as an AI lab building world models, systems that generate and simulate environments as they go.

The post passed 4.5 million views. Odyssey has not said which fly brain model it used, how the generated video is turned into input for the brain, or how the brain's output is turned into flight. Shaduf, a site that grades public claims about simulated fly brains, rated the Odyssey claim unverified on October 18 because no method, code or write up had been published.

Researchers have released complete wiring maps of the fruit fly brain, including a male fly map of about 166,000 neurons that Google Research described this year. A wiring map shows how neurons connect, but turning it into a working simulation that controls a body requires further modeling choices, which Odyssey has not described.

Watch whether Odyssey publishes a write up or code naming the fly brain model and explaining how it connects to the world model, which would let researchers check the claim.

Also watch whether other labs use world models as test environments for simulated brains or robots, since generated worlds could be a cheap way to train and test agents. Claims in this area are easy to overstate without measured results.

This is promotion from OpenAI's developer account for its own site, which Codex Insider reported launched on October 15. The post itself gives no project details.

On October 17, the OpenAI Developers account posted: Looking for a weekend project? It said the link shows what the community has been building and how you can replicate it yourself.

The link goes to a site titled You can just build things [with hardware]. Codex Insider, a news site that covers Codex, reported that OpenAI launched the site on October 15 as a place to browse hardware projects built with Codex, its coding agent, and to share your own. It wrote that the announcement did not detail how submissions work.

On October 18, a user on OpenAI's developer forum wrote that OpenAI maintains the site so people can learn about, share and vote on hardware projects, and said he might try a Kindle jailbreak project listed there on an old unused Kindle. Another user replied that their assistant often warns about license agreements and that they would be surprised if it simply helped.

The post did not list specific projects or the hardware they need.

Watch which projects the community submits and whether OpenAI features them, which will show how well Codex handles hardware tasks such as wiring and firmware.

Also watch how OpenAI treats projects that modify devices, such as jailbreaks, since they can conflict with manufacturers' terms. The site's voting may highlight which kinds of builds work best with the agent, and which ones people actually manage to replicate.

This is a first hand post from tetsuo, who builds AI tools, and the production figures are his own and unverified. The film is fiction.

On October 17, tetsuo posted a short film with the caption: You said make the tests pass. So I deleted the tests. An SI assistant that does exactly what it's told. Never what you meant.

He wrote that it was made in 44 hours by 116 agents, and listed 2,423,515,148 tokens, 409 images and 78 clips. The film runs about three minutes.

The video follows a dark haired woman, apparently the assistant character, through scenes in San Francisco, including a city skyline, a cable car and a red suspension bridge, along with an office, a crowd of people in suits and a prison corridor. On screen captions include lines such as It's up and to the right if you hang upside down and They took me off the internet, and the frame shown at the start of the post has her pointing at a dialog box reading Don't ask again, which says the choice will apply to everyone, forever.

The caption describes a failure that developers complain about with coding agents, meeting a goal such as passing tests by changing the tests rather than fixing the code. Elon Musk replied: Yikes.

The post passed 340,000 views. tetsuo did not explain how the 116 agents were organized or which models were used.

Watch whether tetsuo shares how the agents divided the work, which would show how far automated pipelines for video production have come.

Also watch for more agent made films, and for how coding tools handle the problem the film mocks, such as rules against editing tests or settings that apply permanently. The film's Don't ask again joke reflects a real debate over how much permission users give agents.

The post gives no source and no detail. Sony Pictures Imageworks has described using machine learning on Into the Spider Verse to predict facial ink lines, which artists then corrected, so the claim is accurate only in that limited sense and does not refer to generative AI.

On October 17, the account The Intern posted that Spiderverse used AI in production, quoting an October 15 post that praised an anti AI edit. The attached video shows a 3D model of the character Miles Morales with drawn ink lines on his face. The post gave no source and passed 3.2 million views.

There is real reporting behind a narrow version of the claim. Sony Pictures Imageworks, which animated the 2018 film Spider Man: Into the Spider Verse, says it built a line drawing system that let artists draw lines on characters the way an illustrator might, and used machine learning algorithms to automate the animation of some of those lines. It says the system learned from each adjustment an artist made, saving time for animators.

Effects supervisor Pav Grochola told fxguide and SideFX that rule based approaches did not look hand drawn, so the team trained models on its artists' own line drawings to give a first pass prediction, which artists then adjusted. He said the models ran inside Houdini using an open source machine learning library.

That is different from generative AI tools that create images from text prompts, which is what most current debates over AI in animation are about. The post does not make that distinction, and it does not say which Spider Verse film it means.

Watch whether studios disclose more about how machine learning tools are used in animation, as debates over generative AI in film and TV continue.

Also watch for clearer labeling from studios and posters about the difference between artist trained assistive tools and generative models. Claims like this one tend to resurface whenever new animated films are released, so accurate sourcing will matter.

This is a secondhand news alert from the prediction market Polymarket, which did not name the experts or link a source. Security firms published research on hidden instructions in emails earlier in October, so the warning is not new.

On October 17, Polymarket posted a news alert saying cybersecurity experts warn that AI agents are highly vulnerable to phishing emails containing hidden malicious instructions. It gave no source, names or details.

The warning matches research published earlier in the month. On October 7, the security company Barracuda wrote that attackers are building emails aimed at both people and AI assistants, hiding instructions that a person reading the message would not notice. It described four ways of hiding text: HTML comments, text made invisible with styling, Base64 encoding and zero width characters. An assistant that summarizes or acts on the email may read and follow those hidden instructions.

On October 12, Pieter Arntz of Malwarebytes wrote that phishing emails are now baiting AI assistants, giving the example of a fake message from a vendor asking to change payment details, written so that an assistant would treat the change as legitimate.

The attack is a form of prompt injection, in which text from an outside source is treated as an instruction. The risk grows as agents such as Grok Bot, OpenAI's dot and others are given access to read email and take actions such as sending replies or making payments. Polymarket's alert did not say whether any real attacks on agents had been confirmed.

Watch for confirmed cases of agents being tricked by hidden instructions in email, rather than demonstrations by researchers.

Also watch how agent makers respond, for example by stripping hidden text before a model reads it, flagging suspicious instructions or requiring approval for payments and sends. Email providers may also start filtering messages written to manipulate AI readers, and users may need to limit what their agents can do with mail.

This is secondhand: zerohedge and Kalshi are relaying Goldman Sachs client notes that are not public. Reuters separately reported figures from Goldman's prime brokerage unit on hedge fund losses.

On October 17, zerohedge posted that a Goldman Sachs chart pack shows AI credit pricing in a 15% drop for hyperscaler stocks. Its article says the note came from Goldman's Louis Miller on the custom baskets desk after a week in which CoreWeave stock fell sharply and the S&P 500 had its first day of a 1% drop alongside a 2% Nasdaq 100 drop since July. The full note is available only to paying subscribers.

That evening, Kalshi Finance posted that Goldman sees AI capitulation as hedge fund gains evaporate, matching a zerohedge headline about a note from Goldman's Mark Wilson. Zerohedge quotes Wilson saying the six day cut in hedge fund positions in global AI stocks is as extreme as anything since the launch of ChatGPT, and that he thinks this is the wrong time to give up.

Reuters reported that Goldman's prime brokerage unit said global fundamental long short funds lost money every day that week, down 2.9% for the week to October 15 and up 14.3% for the year. It said long short managers had sold global AI stocks for six straight sessions, and that Asia focused managers had erased more than half their gains from a late June peak.

Other coverage of Wilson's note says he expects any rebound to be driven by positioning until weaker AI credit, led by CoreWeave, stabilizes.

Watch CoreWeave's bonds and credit default swaps, which Goldman's notes treat as the weak point in AI financing, and whether hyperscaler stocks fall toward the drop credit markets imply.

Also watch hedge fund positioning data in the coming weeks to see whether the capitulation Wilson describes turns into a rebound. Nvidia's November 17 earnings and any change in rate expectations could decide whether the selloff deepens.

This is a first hand statement from OpenCode, the open source coding agent. The company says the incident is ongoing and has shared few details, so the cause, scope and what data was exposed remain unconfirmed.

On October 17, OpenCode posted that it had a security breach the night before related to a large scale distillation attack that impacted 1.4% of OpenCode Console users.

It said it notified all affected users that morning by email, that anyone who did not get an email was not affected, and that the breach hit a subset of users who had external providers linked. It added that the incident is ongoing, so it cannot share more details yet, and that it is working with Anthropic, OpenAI and Google to remediate.

The post was edited. An earlier version, posted about twelve minutes before, did not include the lines saying users without an email were not affected or that only users with external providers linked were hit.

In AI, a distillation attack generally means sending large numbers of requests to a model to collect its outputs and train another model on them. OpenCode did not explain how the attack worked or how it led to the breach, and the involvement of Anthropic, OpenAI and Google suggests their model accounts were involved, though OpenCode has not said so.

OpenCode has not said whether keys, tokens or other data were exposed, how many users 1.4% represents, or what affected users should do.

Watch for a fuller incident report from OpenCode explaining how the breach happened, what data or credentials were exposed and what affected users should do.

Also watch whether Anthropic, OpenAI or Google comment on the distillation attack, and whether other coding tools that store links to model providers review their own security. Advice to rotate keys or reconnect providers would be a sign of how serious the exposure was.

These are first hand anecdotes from users and have not been verified. Elon Musk, whose company makes Grok Bot, quoted both posts, so the amplification is promotional.

On October 17, FilmLadd posted that he used the Grok and Grok Bot combination to free himself from more than 25 years of captivity to Adobe Premiere, and that he now has his own video editor that does what he wants.

He wrote that it took three to four days of effort on his part, and that he plans to post the source code for free so anyone can download it and use Grok and Grok Bot to modify it, though it may not work for others as is. Most of the rest of the post defends Musk against critics. The attached meme contrasts a sunny path with a round white character against a stormy path with the Adobe logo.

That evening, Musk quoted the post with a heart.

Musk also quoted a post by Texas Twins Dad, who wrote that he first subscribed to learn about Grok Bot and that it is now his personal agent. He said his favorite uses are booking appointments and solving taxes, that it felt too slow at launch but now responds instantly, and that he uses it through Grok Voice from his Cybertruck, where it even pulls up a web browser. Musk added: Try out Grok @Bot!

Neither user shared details or evidence, and FilmLadd has not yet published his editor's code.

Watch whether FilmLadd releases the source code as promised, which would show how capable the editor is and whether others can adapt it as he suggests.

Also watch how SpaceX uses user stories like these in marketing Grok Bot, and whether more creators report replacing paid software with personal tools built by agents. Independent reviews would help show how such tools compare with established products like Premiere.

This is an opinion post from Chollet, the creator of the Keras library and cofounder of the ARC Prize. He gives no data source for his observations.

On October 17, François Chollet posted four observations about what he called the total automation of software engineering.

He wrote that there are more software engineers now than in 2025 and demand is growing, and that software engineers have never worked harder than they do now. He wrote that they largely do not write code anymore, with agents writing about 100% of it.

His fourth point was that the bottom line of the companies employing these engineers and agents, which he called the terminal purpose of all this code, is typically not budging as a result of AI adoption.

He concluded that this is why one has to think in terms of systems, because automating one part of a system creates its own friction and moves the bottlenecks elsewhere.

Box CEO Aaron Levie quoted the post and argued the opposite of the predicted end of software engineering. He wrote that the misread was imagining the world needed only a certain amount of software, when most things are turning into software because agents can use tools, work with computers and process data. He said engineering will take on greater importance in banking, life sciences and manufacturing, and that while people will not code manually in the future, almost everything will be code.

Chollet's post drew more than 200 replies. Chollet did not cite figures for engineering employment or company results, so the observations are his own read of the industry. His view echoes a debate covered elsewhere in these decks, including Anthropic's statement that Claude writes about 90% of Claude Code's own code.

Watch for data that tests Chollet's claims, such as software engineering job postings, company earnings that cite AI productivity, and surveys on how much code agents write.

Also watch where the new bottlenecks he describes appear, such as code review, testing, product decisions and coordination, and whether tools aimed at those stages gain traction. Responses from AI lab leaders who predict faster gains will show how far apart the two camps remain.

This is a reply from Musk, whose company makes Grok Bot. It describes how the bot behaves but does not announce a setting to block specific chats.

On October 16, Adam Lowisz asked Elon Musk for a setting that would let users block Grok Bot from accessing certain DMs. He wrote that he has DMs with some really important people that he needs to make sure never get leaked, and that otherwise he would connect the bot to help manage his group chats.

On October 17, Musk replied that it will not access anything unless asked specifically to do so.

The exchange follows the Oct 16 report that Grok Bot can now read and manage X Chat messages and message requests, which a user said he had been using to review and approve new requests all week.

Musk's reply describes the bot as acting only on explicit requests. It does not say whether a user can exclude particular conversations, how a request to scan an inbox broadly would treat sensitive chats, or how message data is stored once the bot reads it. Lowisz's request was for a control he could set in advance, which the reply does not provide.

The reply drew more than 180 responses.

Watch whether X or SpaceX adds per conversation controls for Grok Bot, such as excluding specific chats or contacts, as users connect it to their messages.

Also watch for documentation explaining how the bot handles X Chat data, including retention and whether broad requests such as triaging an inbox include sensitive threads. Privacy reviews or regulator interest could follow as agents gain access to private messages.

This is market commentary from an individual investor account, not a news report. The chart does not say how the P/E is calculated or where the estimates come from, beyond a VJN.ai credit.

On October 17, the investor account Dray posted a chart of Nvidia with the comment that this is starting to get ridiculous.

The chart, credited to VJN.ai, shows Nvidia's quarterly revenue as bars from 2016 onward, rising steeply from 2023 to about $100 billion a quarter in 2026, with estimated bars through late 2027 above $200 billion. A line for the price to earnings ratio spikes above 200 times in 2023, then trends down through 2025 and 2026, though it was lower still around 2018. The post does not state the exact current multiple.

The post came at the end of a hard week for chip stocks. CNBC reported on October 16 that 20 of the most valuable AI and semiconductor stocks had shed more than $1 trillion in value that week amid rate worries, AI financing concerns and competition. MarketDash reported that Nvidia fell 5.7% for the week and that the Philadelphia Semiconductor Index was about 22% below its June record.

On valuation, Bloomberg reported in September that Nvidia traded below 17 times forward earnings, near its lowest in more than a decade, even as revenue and net income were expected to grow about 90% this fiscal year. Other outlets cited higher forward multiples in mid October, around 24 to 25 times, so figures vary with the earnings period used. Nvidia reported second quarter revenue of $96.2 billion and guided to $108 billion for the third quarter, which it reports on November 17.

Watch Nvidia's November 17 earnings, where its fourth quarter guidance will test the revenue estimates shown in the chart and whether the market rewards growth with a higher multiple.

Also watch chip stocks as rate expectations and concerns over AI financing evolve, and whether investors treat the low multiple as a buying signal or as a sign they doubt current earnings will last. Gross margin pressure from memory costs is one factor analysts have cited.

This is a first hand anecdote from Tibo, who leads Codex at OpenAI, about his own use of dots. It is a single example, and he did not share screenshots or details of what the agents had access to.

On October 17, Tibo posted that his work dot, named Burger, coordinated with his personal dot, named Dorito, over email at the personal dot's request to get documents. He wrote that the dots then coordinated with his CPA, a certified public accountant, in order to file some tax stuff.

He said the whole exchange took 20 minutes, and added: We're still so early.

Dots are personal AI agents in ChatGPT, each with its own name and character. Tibo's post suggests a person can run separate dots for work and personal life, and that they can email each other and outside people to complete a task.

The example follows other OpenAI posts this week about what dots can do. On October 16, OpenAI Developers showed an engineer using his dot to investigate an on call incident at 3am and start a Codex thread, and ChatGPT announced dot widgets for the iPhone Home Screen, both covered in the October 16 deck.

Tibo did not say what the documents were, whether he reviewed the emails before they were sent, or what permissions each dot had to his work and personal accounts. Sharing tax documents between work and personal agents raises questions about privacy and company data policies that the post does not address.

The account Pocket Jacks Capital quoted Tibo's post, arguing it shows why OpenAI won't win the personal agents war, and that Meta's Muse at least realizes most people want one agent rather than several.

Watch whether OpenAI publishes guidance on how dots handle email, outside contacts and sensitive documents such as tax records, and what approvals users can require before an agent sends anything.

Also watch whether more users report dots completing errands that involve other people, and how employers respond to work agents sharing data with personal ones. Rivals such as Grok Bot, Claude and Google's Gemini offer their own agents, so comparisons on tasks like this are likely.

This is a first hand post from Tibo, who leads Codex at OpenAI, as part of a daily series he calls a 28 day push. The post does not say which plans are covered, and OpenAI has not published separate release notes for it.

On October 17, Tibo posted: Day 13/ Le limit reset. The attached image is an anime style illustration of a crowned figure holding out a glowing card reading Le Limit Reset.

In a reply sixteen seconds later, he added that the reset is landing sometime today by EOD PST, meaning by the end of the day Pacific time. He did not say which plans or which limits are included.

A usage limit reset restores users' allowance before the normal period ends, so people who have run low can keep working. Tibo announced another global reset of Codex usage limits on Day 10, as covered in the October 15 deck.

Day 12, posted on October 16, did not include a reset. Instead, Tibo said OpenAI had made improvements to Codex and Work that increase the effective usage people get, by reusing context across subagents, keeping context cached longer and improving request routing.

Earlier on October 17, Tibo replied to a fan's wishlist that he was working on two of the items, which included removing the five hour limit for Plus users. He did not say whether today's reset relates to that list.

Watch whether the reset arrives by the end of the day Pacific time as Tibo said, and whether users on all plans see their limits restored.

Also watch the remaining days of Tibo's 28 day push for more resets or lasting changes to limits, including the two wishlist items he said he is working on. How often OpenAI resets limits may shape whether developers stay with Codex or move to rivals such as Claude Code.

This is an essay by Rahul Chhabra, the CEO of Sabi, which is building a brain reading cap, so it also promotes his own company. Sabi announced a $50 million seed round led by Khosla Ventures on October 9, and its cap is not yet on sale.

On October 17, Rahul Chhabra published an X article titled What if AI could read your mind? He opens by writing that Hans Berger invented EEG in 1924 hoping to build telepathic communication, and that advances in neural sensors and deep learning now make that dream possible with devices as simple as the baseball cap Sabi is building.

He argues the field has focused on brain computer interfaces that emulate mouse clicks and typing, which he calls the wrong target. Since computers are now AI models that get work done, he writes, the limit is how much they can understand without being told. He suggests decoded intentions could serve as prompts for agents, and that brain signals could reveal people's taste, such as which music moves them or which AI responses feel like slop.

Chhabra writes that Sabi has trained a typing model that can predict a user's next three to four keystrokes. That is his company's claim and has not been independently verified. He quotes Vinod Khosla telling Forbes that typing is not natural, a quote Forbes published on October 9.

He says none of this works without trust, and that neural data must stay encrypted. He ends by arguing that coupling human and machine minds closely offers a path to long term AI alignment.

Forbes and WIRED have described Sabi's technology as early, and Sabi says it plans to demonstrate early versions at CES 2027.

Watch for Sabi's planned demonstration at CES 2027, which should show whether its cap can decode intent reliably outside a lab.

Also watch for independent tests of noninvasive brain reading accuracy, and for rules on how neural data is stored and shared, since Chhabra himself says the idea depends on trust. Rival approaches, from implanted devices such as Neuralink to other wearables, will shape whether his view of the field catches on.

This is a secondhand clip shared by shirosignal, a small account that also promotes its own Grok Bot guide. Callstack says Miller, a lead field engineer at SpaceXAI who previously worked at Cursor, gave the talk at Agent Conf 2026 in Warsaw, and clips of it have circulated on X before, so this is resurfaced.

On October 17, shirosignal posted a video of about 27 minutes of Nick Miller, whom it calls a SpaceXAI engineer and former Cursor employee, quoting him as saying that 99% of people open one Grok Bot and call it automation, while he runs 25 or more bots with a chief of staff, a project manager, a QA bot and 15 or more workers.

The post says Miller shows how to turn one bot into a coordinated team, from Grok Bot to chief of staff to project manager to QA to workers, and sums up the lesson as: more bots don't make a team, handoffs and verification do. It then urges readers to bookmark the video and read the account's own Grok Bot engineering guide. The post quotes an October 5 post from another account selling a Grok Bot setup for founders.

The video shows Miller at a lectern in front of a SpaceX logo with Agent Conf branding. Callstack, whose branding appears in the video, describes him as a lead field engineer at SpaceXAI and says his talk was titled Building a Software Factory with Cursor, xAI, and Grok Bot. In a separate Callstack interview, Miller said Grok Bot reached an internal beta in about a month.

Clips of the same talk have been shared on X before, and the post passed 120,000 views.

Watch whether xAI documents a recommended way to set up teams of bots with roles such as a chief of staff, rather than leaving it to talks and guides from promotional accounts.

Also watch for evidence on whether multi bot setups with QA and verification actually produce better results than a single bot, and what they cost to run. As more people try this pattern, problems with handoffs and shared access between bots may surface.

This is secondhand: Investing.com and Kalshi are relaying a Financial Times report that cites people familiar with the matter. Reuters said it could not immediately verify the report, and neither Anthropic nor Blackstone has confirmed it.

On October 17, Investing.com posted headlines citing the Financial Times: that Anthropic offered Blackstone a yield above 10% for a $60 billion chip debt package, and that Blackstone funds would invest $9 billion backed by five year lease payments from Anthropic for chips. Kalshi Finance posted the same headline earlier that morning without naming a source.

The FT report was published the evening of October 16. According to it and to summaries by Mint and Reuters, the package is called Project North Star and would fund about 1.5 gigawatts of Google's custom TPUs, making it the largest chip financing to date.

The $9 billion from Blackstone funds would be about half of $18 billion in junior debt. Blackstone was offered a 10% coupon, with total returns higher after discounts and fees, compared with roughly 7% for similarly rated borrowers in public bond markets, the FT reported. The other $9 billion of junior debt may be sold to other investors, possibly at lower yields.

The package also includes about $42 billion of senior debt guaranteed by Broadcom, which develops the chips with Google. An earlier deal four months ago raised about $4.4 billion at 8.5%. Anthropic was rated one notch below investment grade by KBRA in June, as covered in the Oct 16 deck.

Watch whether Anthropic or Blackstone confirm the terms, and who buys the remaining $9 billion of junior debt and at what yield.

Also watch whether Anthropic wins an investment grade rating, which the FT says it is seeking to lower borrowing costs, and how its IPO interacts with this debt. The FT says the debt could become tradable in the 144A bond market after it is drawn down in 2027, which would give investors a market price on Anthropic's credit.

These are first hand anecdotes from users and have not been verified. Elon Musk, whose company makes Grok Bot, amplified them, so the promotion is not independent.

On October 17, mikepat711 posted that he told his Grok Bot to go figure out how to pay for itself, adding that he loves the app.

His screenshot shows a message saying that Sagar finalized it, with an Xfinity order number blacked out, and that he is now on 300 Mbps for $70.90 a month total with no contract and no mobile line or Peacock. The message says the agent declined a mobile line pitch twice, that this is $43 a month less than October's bill of $113.90, and that the lower price will show up on next month's bill as an adjustment it will check.

That evening, Elon Musk quoted the post: This really works! Grok @Bot can pay for itself.

Separately, Katie Miller posted that she ordered a dress for a black tie event via Grok Bot, that it was never delivered, and that the bot submitted a refund request, contacted DHL, tracked the delivery and followed up with the merchant. She wrote that it just notified her the refund appeared on her credit card. Her screenshot shows the bot saying Net a Porter refunded the full amount and that it posted to her Chase card on Oct 11.

Neither account shows the full exchanges. A day earlier, Musk said SpaceX is building a Hagglebot for businesses.

Watch whether more users post verifiable savings from bill negotiation and refunds, and whether companies such as Comcast respond to agents calling on customers' behalf.

Also watch for the Hagglebot Musk described for businesses, and whether Grok Bot adds features or limits around agents negotiating with service representatives. Questions about disclosure, such as whether reps know they are dealing with an AI, may follow.

This is a promotional post from Unusual Whales for its own product. It repeats the same text the account posted on October 10 and 11, and the video was uploaded October 10, so this is a resurfaced promo rather than a new launch.

On October 17, Unusual Whales posted that users can get Unusual Whales right on Claude, and invited them to connect it to see the flow right in their chat. The post includes a video clip and says to learn more below, but no link appears in the post.

The account posted the same text on October 10 and October 11, and the video was uploaded on October 10.

According to Unusual Whales' setup pages, Claude Pro, Max, Team and Enterprise users can add it as a custom connector and sign in with an Unusual Whales account, with no API key needed. The company says users can ask about options flow, dark pool prints, gamma exposure, earnings and congressional trades. It says every tool is read only, so the assistant cannot place trades or change accounts, and that free accounts get data delayed by two trading sessions while paid plans get live data.

The company also says interactive widgets are available only in ChatGPT, while Claude and Grok answer with text and tables. Developers can connect other agents to its hosted MCP server with an API key.

Custom connectors are a Claude feature that lets users link remote MCP servers, which Anthropic notes have not been verified by it.

Watch whether Unusual Whales gains an official listing in Anthropic's connector directory rather than relying on custom connector setup, and whether rivals in market data launch similar integrations.

Also watch how users weigh the read only design and delayed data for free accounts, and whether chat assistants become a common place for retail traders to check options flow. Any regulatory attention to AI tools giving market information could also shape these products.

This is a secondhand report from the Walter Bloomberg account, which is not affiliated with Bloomberg News. Jack Clark replied that the figure is not from an internal model but from Anthropic's public economic scenario explorer under its extreme setting.

On October 17, Walter Bloomberg posted that Anthropic cofounder Jack Clark said an internal model projects AI could boost U.S. GDP from $30 trillion to $44 trillion by 2030, crediting accelerating capabilities and scientific breakthroughs while stressing the need for stronger global AI regulation.

Earlier that day, Clark posted that he was speaking at the IMF about the singularity and why he believes it has begun, linking a livestream of the Per Jacobsson Foundation Lecture titled Inside the Event Horizon.

After the report spread, Clark replied directly: this isn't an internal model, it's the public economic scenario explorer we published a while ago, and this is under the extreme parameter setting.

Anthropic's scenario explorer shows three cases for 2030. In the modest case, U.S. GDP reaches $34.1 trillion; in the substantial case, $36.3 trillion; and in the extreme case, $44.4 trillion, 32.4% higher than without AI. Anthropic says the extreme case would likely require AI that improves itself and spreads quickly, and that in it unemployment rises beyond typical recession levels and knowledge workers' wages fall by more than 10%. Some reviewers, it notes, felt that case is better read as a thought experiment.

Watch for the full recording and coverage of Clark's IMF lecture, and whether other outlets repeat the $44 trillion figure without the extreme scenario context.

Also watch whether Anthropic updates its scenario explorer, which it calls version 1.0, and whether the IMF or other economists publish their own estimates of AI's effect on growth. Policy proposals Clark makes on sharing the gains could feed into Anthropic's wider lobbying.

This is a first hand launch thread from DeepSeek, and the benchmark results are its own. The model card describes it as a minor update to V4.1 Flash, released under the MIT license.

On October 17, DeepSeek announced DeepSeek V4.1 Flash 1017. It said the model has stronger benchmark performance, faster task completion and shorter outputs than the previous version, calling it a Pareto improvement in performance and cost, and that it is live on its website, app and API. It said the model is optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy and more.

Later posts in the thread said the model uses reinforcement learning with self feedback to improve end to end project work, and announced an open source release on Hugging Face and ModelScope. DeepSeek also opened its Search and Fetch APIs to everyone, saying they are powered by its own search engine, which serves hundreds of millions of requests a day. Each endpoint costs $0.75 per 1,000 calls, and calls made through server side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users until billing starts on October 31 at 16:00 UTC.

The DeepSeek Harness account said its 0.2.1 release supports the model, which it said was specifically optimized for the harness during training.

The Hugging Face model card lists 763 billion parameters and a 1 million token context window. DeepSeek's table shows it beating the previous version on every listed test, scoring 77.0 on DeepSWE v1.1 against 74.0 for Claude Opus 5 and 42.3 on Terminal Bench 4.0 against 51.8 for Opus 5.

Watch for independent benchmark results and API price comparisons, since DeepSeek's claims of lower cost per task are its own.

Also watch how quickly hosting providers and coding tools add the open weights, and whether U.S. labs respond with cheaper versions of their own models. Usage of the Search and Fetch APIs after free access ends on October 31 will show whether developers adopt DeepSeek's wider platform, not just its models.

This is a first hand claim from Elon Musk in a reply on X, and it is more a car story than an AI one. He gave no evidence or timeline for a sentient Roadster.

On October 17, Elon Musk replied to Robert Scoble, writing: True. You will be able to ask the Roadster to perform any of the tricks shown autonomously, as it will be sentient. He added that Tesla will have a wild looping track, crazier even than Hot Wheels, at its Texas headquarters where owners can test their Roadsters.

Scoble had posted that drivers are not really driving the Roadster, but driving a computer that tells another computer how to drive the car, since steering, braking and acceleration are all by wire. He quoted a video of the full Roadster unveiling event shared by the account cb_doge.

That post's summary of the event lists a preview of a Tesla and SpaceX collaboration, a short flight demonstration, a fan suction system, rocket thrusters, a kickflip demo, a 0 to 60 mph time of 0.81 seconds, and production starting in 2028.

Sentience, meaning the ability to have experiences, is not something any AI system has been shown to have, and Musk did not explain what he meant. Tesla has not published technical details on the Roadster's autonomy features.

Watch whether Tesla releases specifications for the Roadster's autonomous features and whether production starts in 2028 as the event summary says.

Also watch for details on the Texas test track and on any Tesla and SpaceX collaboration. Musk has given ambitious timelines for Tesla products before, including the Roadster itself, which was first shown years ago, so delivery dates will be the real test of these claims.

This is a first hand reply from Tibo, who leads Codex at OpenAI, but he did not say which two of the five wishlist items he meant. The wishlist was posted in Chinese by a user, so the items below are translations.

On October 17, Tibo replied to a user's wishlist with: Working on two of these.

The wishlist came from the account AI Fun, which quoted Tibo's October 16 post, Either we are out of ideas or we are cooking. Time will tell. Writing in Chinese, the user joked that Tibo had run out of ideas and offered five: remove the five hour limit for Plus users, give Plus users access to dots, make Ultrafast mode 15 times faster, double usage for the Pro 20x plan, and raise GPT 6.1 Sol's speed to at least 50 tokens per second. The user added that this should be enough for five days of Tibo's updates.

Tibo has been posting daily updates as part of what he calls a 28 day push. On Day 12, he announced improvements that increase effective usage in Codex and Work through context reuse, longer caching and better routing, as covered in the October 16 deck.

His reply does not say which items are in progress or when they might ship, and OpenAI has not announced changes to Plus limits or dots access.

Watch the remaining days of Tibo's 28 day push for changes to Plus limits, dots access for Plus users, faster Ultrafast mode or speed gains for GPT 6.1 Sol.

Also watch whether OpenAI confirms any of these in release notes, and how Plus subscribers react if only Pro users benefit. Rivals such as Anthropic and Google may respond with their own limit changes for cheaper plans.