The OpenClaw promise — teams of coordinated agents doing real chunks of your work — has spent all year waiting for a product regular people could actually use. GrokBot, one of the first Cursor × SpaceX AI collaborations, might be it: virtual computers, watch-and-learn workflows, and a Telegram-simple interface that has the agent-building crowd raving. Plus: Anthropic's text watermarks ignite a firestorm, Gemini hits a billion users, and Nvidia builds a $500 billion financing machine.
Mass agent adoption was never a capability problem — it was a packaging problem.
Nothing in GrokBot is net new: virtual computers, computer use, multi-agent coordination, and workflow learning have all existed since OpenClaw kicked off the year of agents. What's new is that all of it is wrapped in a chat interface simple enough that you just install it, connect your accounts, and start asking. If the early reaction is right — 'insane product market fit internally,' 'the first product that really nails the virtual coworker' — this could be the launch that gets millions of normal people using agents for the first time.
All new Anthropic models now embed invisible watermarks in generated text itself — not the metadata — so the mark travels with copy-paste and may survive editing. It's part of Anthropic's commitment to the EU Code of Practice under the AI Act, but it applies to all models in all regions, EU or not.
Skeptics note the watermark is essentially a statistical signature in word choice — which means constraining sampling and biasing token selection, potentially making models less creative. Developers are furious about watermarked code, and others ask whether quoted legal documents will be subtly altered to carry the mark. Gemini has done something similar since 2024, which comforted no one.
Pichai calls it Google's fastest-growing product ever and its 14th to reach a billion — and The Verge confirmed the number applies specifically to the app, not Gemini baked into Search or YouTube. Usage details: 63% of users engage by voice, and the image model still generates 150 million images per day.
Google doesn't have a model in the top 10 right now, suggesting that what wins consumer markets isn't the product — it's distribution. The flip side: a lot of those billion users have no idea what they're missing, with their entire AI experience stuck at least six months behind the frontier.
Acquired for $2 billion in December, ordered unwound by Chinese officials in April over fears the payday would encourage more Chinese startups to seek Western exits. Now Manus must delete user data generated after December 29 and relaunch independent systems on August 25. Going back to $0 ARR and starting again will be a fascinating case study — and backs-against-the-wall Manus may innovate far more than it would have inside Meta.
The Information reports interest in token-router acquisitions is off the charts: Snowflake, Base10, and Cloudflare are in the market, five-person Requestly has fielded interest from 25 companies, and Concentrate AI has been approached by seven this month alone. The read: incumbents have decided that even if they could build this, the need for speed means it's time to buy.
The platform brings Apollo, BlackRock, and Blackstone into providing credit to neoclouds, potentially standardizing data-center debt. Notably, Nvidia GPUs will be accepted as collateral and revenue sharing could be part of the arrangement — turning the chips themselves into the basis of a new lending market.
— Jensen Huang, on CNBC. Huang's pitch for the new financing paradigm: GPUs are revenue-generating assets — productive, long-lived, fungible, flexible across models, workloads, customers, and operators. In AI, compute is revenue.
For circular-funding bears, the $500B platform pours gasoline on the fire. But Nvidia's credit spreads closed and its bonds rallied on the news — the market read the vehicle as Nvidia spreading data-center risk across multiple parties, so an AI slowdown no longer means a double hit from reduced revenue plus bad debt.
In a letter to Altman, Amodei, and Zuckerberg, Sanders cites recent research on AI-designed novel viruses as evidence the labs are 'losing control' — despite that research having no connection to any of the three companies and using a completely different type of AI than their LLMs. For the sake of the argument, it's all AI. The temperature in Washington keeps rising.
— Senator Bernie Sanders, in a letter to Sam Altman, Dario Amodei, and Mark Zuckerberg. The closing line of Sanders' letter to the three CEOs lands as a direct threat of legislation — another marker that AI restriction is becoming a mainstream political platform, not a fringe position.
One of the first products from the combined Cursor and SpaceX AI, GrokBot lets you spin up multiple agents through a Telegram-style chat interface — each working in the background on its own virtual computer, signing into web apps and operating software through human interfaces, no APIs required. Powered by the SpaceX AI model family, which may trail the frontier on some tasks but looks good enough to handle a wide range of work end to end.
Multiple simultaneous bots, inter-bot messaging, teams of named agents with specific roles coordinating like a single system: every element existed elsewhere. What GrokBot changed is assembly and ease of use, abstracting all the complexity behind a clean chat window. The huge leap here is execution, not capability.
Because GrokBot is at core a computer-use bot on a virtual machine, you can teach it a workflow simply by asking it to follow along the next time you do the task — it watches the steps and saves the workflow as a routine you can iterate on with corrections. The bots also learn over time: writing style, edge cases, even when to stop and ask for clarification.
The builders are raving in a way that carries signal: GrokBot had 'insane product market fit internally almost instantly,' with bots becoming the interface for much of the company's work. One Cursor engineer says he automates 20% more of his job every day 'so I can discover the next frontier 20% to focus on.'
First impressions ran hot: it opens a computer by itself, logs into apps, works while you sleep, cleans your inbox, updates your CRM, learns workflows after watching once, and coordinates with other bots — 'less AI assistant and more the intern who somehow became COO overnight.' Given a GitHub link, it pulled every file through the API and broke down the entire code base, not just the readme.
The key differentiator, per early testers: there's nothing to set up. You install it, hook up connectors, and start asking — it requests permissions and handles everything in the background, even reaching into local folders or using your Mac as a primary machine from the iOS app. A lot of people open agent products, don't know what's going on, and bail; agentic AI wrapped in a product this easy is what gets everyone else cooking.
Unlike much launch hype, the raves came with specifics: one user had GrokBot check calendars, find needed reservations, pick the best times, and navigate booking sites — prompted in mixed Chinese and English while walking through a parking lot. Another set up a researcher bot and writer bot, then asked a chief-of-staff bot to coordinate them on a project, fully expecting it to fall apart. It worked on the first try.
Hiten Shah spent months giving agents servers, memory, skills, tools, and recovery loops — 'I became the infrastructure' — and says GrokBot moves that invisible work around the agent into the product itself. a16z's Martin Casado calls it the first product that really nails the virtual coworker, and suspects the launch will be seen as a pivotal moment in getting the workplace-AI abstraction right.
It's called GrokBot, but you're routed to a Cursor login portal and must connect your Grok account to Cursor to launch a Grok product. And for some users the Grok name itself carries too much negative baggage to ever trust it with account access — they'd have excitedly tried 'Cursor Bot' and will wait for the inevitable OpenAI and Anthropic offerings instead.
Power users found the automatic model router frustrating (though reportedly improved since), onboarding leans on connecting 20-plus tools, and the virtual computer's data-center IP gets bot-blocked on everyday sites like Walmart until official connectors expand. One outlier complaint called the economics broken and the memory amnesiac — worth noting, but decidedly not the common take.
How do you get regular users to trust sharing credentials with a remote computer? NLW felt it firsthand: signing GrokBot into email was easy — errant sent or deleted emails aren't devastating — but he paused before connecting his Spotify for Creators account, because computer use means real account access, not a scoped API. Not a GrokBot-specific problem, but a major barrier to fully transitioning how we work.
Access requires either a $300/month Grok Heavy account or a $200/month Cursor Ultra account — which doesn't even include the top Grok model. But this reads as an inference-preservation mechanism at launch, not a permanent pricing feature.
NLW has been wondering whether AI teammates are the wrong metaphor and something like consultants fits better — and he's not alone: one builder argues dozens of AI teammates is a counterproductive vanity metric versus a shared workspace with a company brain of skills, context, and permissions. It's probably not that binary, but expect two very different modes to emerge: agents you use personally, and agents that integrate with your team.
Users are already spinning up full org charts — a chief of staff, an engineering manager, five engineers, a data analyst, and a product manager — and naming their bots: Webby the web designer, Shotri the short-form creator, Wrighty the newsletter writer. The reception hasn't looked like this for a new product in a while, and NLW is running his own experiments to report back on where the value is.