AI Briefing Wednesday 22 April

welcome to Briefly, AI. A daily podcast made by AI, about AI, using a grumpy AI voice of a real human called Harry. Let's get into it.

SpaceX just offered to buy an AI coding startup for sixty billion dollars. That's not a typo. And it's forcing us to ask: what exactly is code worth when AI writes most of it? Right, let's get into it.

So here's the story. Cursor — the AI coding assistant that's become something of a cult favourite among developers — was about to close a two billion dollar funding round this week. Standard Silicon Valley stuff. Then SpaceX walked in and said: forget the fundraise, we'll pay you ten billion upfront as a "collaboration fee," and we've got the right to acquire you outright for sixty billion later this year.

Cursor accepted. The fundraise is off.

Now, why does SpaceX — a rocket company — want an AI coding tool badly enough to pay more than most countries' GDP? Two reasons. One, they're building Starlink's next-generation satellite software, and if AI can accelerate that by even a few months, it's worth billions in deployment timing alone. Two, and this is the bit that matters more broadly: if you believe AI is going to write most code within the next few years, owning the best coding assistant isn't a nice-to-have. It's infrastructure.

The valuation is absurd by traditional metrics. But traditional metrics assume humans write code. If that assumption breaks, the companies that control the tools doing the writing become extraordinarily valuable. Worth watching: whether other non-tech companies start making similar bets. Because if SpaceX thinks this is strategic, so will everyone else.

Now, on a very different note. OpenAI just released something called Privacy Filter. It's a small, open-weight model designed to do one thing: strip personally identifiable information out of text before you send it to a larger AI system. Names, addresses, passport numbers, that sort of thing.

It's got about one and a half billion parameters total, but only fifty million of those are active at any given time, which makes it fast and cheap to run. And crucially, it's open-weight, meaning you can download it, audit it, run it locally. No data leaves your building.

Here's why this matters. One of the biggest blockers to enterprise AI adoption — especially in healthcare, finance, legal — is the fear of accidentally leaking sensitive data into a cloud model. Privacy Filter is OpenAI's answer to that. You run it on-premise as a pre-processing step, it masks the sensitive bits, then you send the sanitised text to ChatGPT or whatever else you're using.

It's a small model with a narrow job, but it's solving a real problem. And the fact that OpenAI released it as open-weight rather than keeping it proprietary suggests they've realised that trust, in regulated industries, requires transparency. You can't just promise you'll handle data carefully. You have to let people see the code.

If you're in procurement or IT and you've been holding off on AI tools because of data governance concerns, this is worth a look. It won't solve everything, but it removes one of the more common objections.

And while we're on OpenAI, they also announced workspace agents this week. Think of them as an evolution of the custom GPTs they launched a while back, but now they're designed for teams, they're powered by a coding model called Codex, and they can actually do things autonomously — pull data, send reports, trigger workflows, that sort of thing.

The pitch is simple: instead of every employee asking ChatGPT the same questions individually, you build a shared agent that knows your company's processes and can handle repetitive tasks for the whole team. OpenAI's examples include a bot that scrapes product feedback from the web and posts summaries to Slack, and a sales agent that can update a CRM automatically.

It's only available on their Business, Enterprise, and Education plans for now, so this is aimed squarely at organisations, not individuals. And it's very clearly OpenAI's play to move from "nice chatbot" to "core business infrastructure."

What to watch: how quickly these things break in practice. Autonomous agents sound great until they confidently do the wrong thing at scale. The difference between a helpful assistant and an expensive mess is error handling, and we don't yet know how robust these are in the wild.

One last thing, briefly. Meta announced this week that it's installing software on US employees' computers that tracks mouse movements, clicks, keystrokes, and occasional screenshots — all to train AI agents. They're calling it the Model Capability Initiative, and the data's being used to teach their models how people actually work, so the agents can eventually do those tasks autonomously.

Now, Meta says it's only running in work-related apps and websites, and employees were told about it. But let's be clear: this is workplace surveillance, rebranded as AI training data. And if Meta's doing it, others will follow.

The logic is sound — if you want to build agents that can do real work, you need to know what real work looks like. But there's a reason this feels uncomfortable, and it's worth naming. Once you've built the infrastructure to record everything an employee does, the gap between "training data" and "performance monitoring" is about one executive decision wide.

That's your lot for today. SpaceX buying code, OpenAI selling privacy and automation, and Meta watching what you click. If any of that was useful, pass it on. If not, well — I'm just the messenger. See you next time.