Google Kills Its Own AI Agent, and What That Tells Us

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.

Right. Google just quietly pulled the plug on one of its most promising AI experiments — and nobody seems to have noticed. Let's talk about that.

So, Project Mariner. If you missed it, this was Google's attempt at building an AI agent that could actually do things for you across the web — book flights, fill out forms, navigate websites on your behalf. The sort of thing that sounds properly useful, not just clever. It launched in December as an experimental feature. It shut down on May 4th. That's less than five months.

Now, Google hasn't said much about why. The official line is just "thank you for using it, we've turned it off." But here's the thing. This wasn't a quiet research project that failed to get traction. This was a public-facing product with real users, positioned as part of Google's AI-first future. And they killed it anyway.

Why does that matter? Because it tells you something about where the wheels are starting to come off in the race to ship AI agents. Either the technology wasn't reliable enough to scale — which is worrying, given how much Google's been talking up its agent capabilities — or the legal and liability risks got too messy to justify keeping it alive. When an AI agent is making purchases or filling out forms on your behalf, and it gets something wrong, who's responsible? Google clearly decided they didn't want to find out.

The broader point is this: we're in a phase where everyone's racing to demo agentic AI, but very few companies are willing to actually stand behind it in production. If Google — who literally invented the Transformer architecture that powers all of this — can't make a web agent work at scale, that should tell you something about how far we still have to go.

Now, on a related note. Snap and Perplexity just ended their $400 million deal — amicably, they say, which usually means it wasn't amicable at all. The partnership, announced last November, was supposed to bring Perplexity's AI search engine directly into Snapchat. That would've been a big deal. Snap gets better search without building it themselves. Perplexity gets distribution to hundreds of millions of users. Everyone wins.

Except it didn't happen. And while neither side is saying why, the timing's interesting. This is the second major AI partnership to collapse in a week, and it fits a pattern. These distribution deals look brilliant on paper, but integrating someone else's AI into your product turns out to be a lot harder than just plugging it in. You've got brand risk, liability risk, user experience mismatches, and the nagging question of what happens when the AI says something embarrassing in front of your entire user base.

If you're running a product team right now and someone's pitching you on a quick AI integration, this is worth remembering. The demo is easy. Production is where the deal falls apart.

And finally, a quick one on the infrastructure side. SpaceX — yes, that SpaceX — is apparently planning to spend up to $119 billion on a chip factory in Texas. They're calling it Terafab. The proposal describes it as a "multi-phase, next-generation, vertically integrated semiconductor manufacturing and advanced computing fabrication facility," which is a very long way of saying "we're going to make our own chips."

Now, this is still at the proposal stage, so take it with a grain of salt. But if it happens, it's a signal of just how strategic chip manufacturing has become. Tesla already announced a $25 billion version of this. SpaceX doing the same suggests that Elon Musk's companies are betting they can't rely on external suppliers — whether that's TSMC, Intel, or anyone else — to give them the compute they need at the speed they want it.

It's also worth noting the scale. $119 billion is more than the entire GDP of Ukraine. For context, TSMC's Arizona fabs — the ones that just produced the first Blackwell chips on US soil — cost about $40 billion. So either SpaceX is planning something genuinely massive, or someone got very creative with the numbers to impress local officials. Either way, it's a sign that the AI infrastructure race is moving from "who can train the best model" to "who controls the physical supply chain." And that's a very different game.

That's your lot. Google killed its own AI agent, Snap and Perplexity called it quits, and SpaceX might be about to build a chip factory the size of a small country's economy. Three stories, five minutes, and not a single use of the word revolutionary. If any of that was useful, share it. If not, well, blame the machine. See you next time.