Elon's plan to take over AI

hello everyone 👋,

I'm writing this one from Kazakhstan.

It is a beautiful country that somehow still sits under most people's radar.

Well, almost everyone's. Kanye West played a concert here yesterday, and I went. It was next level.

I have no sensible transition from Kanye to AI, so we are just going to go for it.

SpaceX has just finished buying Cursor for $60 billion. Elon Musk had the same coding tool dilemma as the rest of us. He solved it by buying a whole company, pretty usual at this point.

The interesting part is what it actually bought: a place where developers already spend their working day.

This week’s story is about why every AI company suddenly wants to sit between you and your computer.

SpaceX completed its $60 billion Cursor acquisition

Cursor is joining SpaceXAI to improve Cursor, Grok Bot, Grok Build and the Grok API. If you use Cursor, watch model choice, pricing and data controls.

Grok Bot gives every agent its own computer

It can use apps, store files and run scheduled work in the background. That is powerful, but agents like this need much better controls when they get stuck.

Claude is adding invisible marks to text

The mark can show that Claude processed a piece of text. It cannot prove that Claude wrote the original work.

Why I switched from Claude to ChatGPT

I use and pay for both. But this week I finally decided to cut all my ties to Claude and you should too.

Cursor confirmed on Friday that its acquisition by SpaceX has officially closed.

Cursor will join the SpaceXAI team and help improve Grok Build, Grok Bot, the Grok API and Cursor itself.

SpaceX is paying for distribution. Cursor already sits inside the daily workflow of developers.

SpaceX's own filings repeatedly talk about Cursor's distribution to expert developers and the value of putting its AI models inside software people use every day.

The largest AI companies already have expensive models and enormous data centres. Their next problem is getting those models into real workflows where people will keep using them and paying for them.

Cursor gives SpaceX a direct route into software development. It also gives Grok models a home beside Claude, GPT and Gemini inside the same product.

For now, the acquisition tells us where SpaceX thinks the value is: inside the tools where work already happens.

I have been using Grok Bot this week, and it is the clearest example yet of where AI agents are heading.

Each Bot gets a cloud computer. It can sign into apps, use a browser and terminal, store files, remember previous work and run routines on a schedule.

Instead of opening an AI chat whenever you need something, you can create a permanent agent for inbox triage, morning briefings, customer support or content moderation and let it keep working after you close your laptop.

The best use cases are boring and repeatable. Give one Bot one job, clear instructions and an output you can check quickly. That is much easier to trust than one giant agent with access to everything.

My early impression is that the setup feels polished, but the reliability does not. Creating an agent and sending it to work is easy. Seeing exactly what it is doing when it gets stuck, why it chose a particular tool or how much usage it is burning is much harder.

The general vibe from other early users seems similar. People love the simple setup and the idea of permanent agents. Browser sessions, login problems, usage limits and reliability still make the product feel like a beta.

OpenAI is moving in the same direction through a new opt-in feature called Computer History.

It can give ChatGPT and Codex context from your recent work, so you spend less time explaining which document, website or project you were using five minutes ago.

Computer History records interaction events from the apps and websites you allow. That can include clicks, typing, keyboard shortcuts, switching between apps and context. It turns that activity into text summaries and local memory files. It does not record your screen, microphone or system audio, and private browsing is excluded.

Which sounds useful and slightly creepy, which is becoming a fairly normal combination for AI features.

The practical use cases are much more interesting than "better memory."

You can ask what you were working on before lunch, find the planning document you opened earlier, prepare a summary for your daily standup or continue a task without explaining every open tab. It can also notice work you repeat and suggest turning the process into a Skill or scheduled task.

If your day involves jumping between Slack, Google Docs, a browser and a codebase, that could remove a lot of annoying setup from every AI conversation.

I will experiment with this thing in the coming weeks and will keep you updated and share any useful or cool use cases I find.

Anthropic's new supported models will add machine-readable marks to Claude-generated text and files as part of the industry's response to new EU transparency rules.

The reaction online has been dramatic, but the distinction matters.

If you write an article and ask Claude to clean up the grammar, the final version may carry a mark. That can show Claude touched the text. It does not prove Claude came up with the ideas or wrote the original draft.

Anthropic has also acknowledged limitations in how reliably these marks can be detected. The problem goes in both directions. Human-written text may show a Claude mark after proofreading, formatting or translation. A short document, heavily rewritten copy or text mixed with other material may lose enough of the signal that the mark is no longer detected.

That makes the watermark evidence that Claude processed the text, not proof of who wrote it.

Agencies and freelancers should only care if a client contract requires human-written work or disclosure of AI assistance. A document could trigger a detector even when Claude only fixed the grammar, which creates an awkward argument over authorship unless the rules are clear in advance.

SEO teams should watch this, but a Claude mark does not automatically mean a Google penalty. Google's current guidance says AI-assisted content is allowed.

For people using Claude for private notes, everyday emails or rough drafts, this probably changes very little.

I was going to put Gemini 3.7 Flash here, but honestly, a slightly better Gemini model is not much of a recommendation.

Meetily is basically a free, unlimited and local version of Granola.

I've been using it for my meetings and I love it. It records your microphone and computer audio, transcribes the conversation and turns it into notes without sending an awkward bot into the call.

The best part is that everything can run on your computer, so your recordings and transcripts stay with you. If you like Granola but want something private, open source and without monthly limits, Meetily is worth trying.

I also uploaded a new video comparing ChatGPT and Claude based on how I actually use both for client work.

I pay for both and use them for real client work. Claude still wins in a couple of areas, but ChatGPT has become the better complete subscription for me.

In the video, I show the features that changed my mind, including native image generation, remote voice control and Codex. I also explain how to move your context out of Claude without starting over.

If you only want to pay for one AI subscription, this is the comparison I would want before choosing.

Watch the full breakdown here.

That’s it for this week.

See you next Sunday,