I don’t usually do weekly roundups. But this week was one of those weeks where three things happened at once and each one would’ve been the story of the month any other year.
Here’s what’s worth paying attention to.
GPT-5.4 dropped — and it can use your computer
On March 5, OpenAI launched GPT-5.4. Not an incremental update. A different product.
The headline feature: native Computer Use mode. You point it at your screen and it navigates your desktop, opens apps, fills forms, moves files — autonomously, without add-ons or plugins. That’s not a chatbot anymore. That’s closer to handing someone the keyboard.
It also ships with a 1 million token context window. Your entire codebase. A full year of emails. A novel and its sequel. It doesn’t matter — the model just holds it all.
The numbers: 33% fewer hallucinations than GPT-5.2, 47% fewer tokens used on routine tasks. Three variants: Instant for speed, Thinking for complex reasoning, Pro for when you need the nuclear option.
What I keep coming back to is that the jump from “generate text” to “operate software” is a real category shift. Every company that’s been building AI-powered workflows just got a different set of assumptions to work with.
Apple gave Siri to Google. That sentence still feels weird to type.
Apple — the company that built its entire identity around privacy and keeping your data off other people’s servers — just signed a deal to run Siri on Google’s Gemini model.
The new Siri in iOS 26.4 has on-screen context awareness. You’re looking at a restaurant in Safari, you say “book me a table for two tonight” — Siri reads the screen, makes the call, confirms the booking. No copying. No switching apps. One sentence.
It can also chain up to 10 sequential actions from a single request. “Book me the next flight to Mumbai, add it to my calendar, and text my wife the arrival time” — that runs as one workflow, not three.
The privacy architecture is genuinely clever. Gemini runs on Apple’s Private Cloud Compute — your data doesn’t flow through Google’s infrastructure directly. Apple is renting Google’s brain but keeping it locked in Apple’s house, under Apple’s rules.
Still, the optics are something else. These are two of the most competitive companies in tech. Now they’re infrastructure partners for one of the most personal features on your phone.
Honestly though? The old Siri was so bad for so long that even Apple’s pride couldn’t save it. Sometimes you just have to call the better team.
Nvidia GTC starts March 16 — and Jensen is teasing “a chip the world has never seen”
GTC is basically the Super Bowl of AI hardware at this point. Every year it gets bigger. This year it might be the biggest one yet.
The Rubin platform is already confirmed — 5x inference performance over Blackwell, up to 10x lower cost per token at scale, six-chip codesign that makes Blackwell look like a warmup act. If that’s all that gets announced, it’s still a major conference.
But Jensen Huang said something specific in the lead-up: GTC will unveil “a chip that will surprise the world.” That’s not the kind of statement he makes casually. Could be the Feynman architecture. Could be something nobody’s been tracking.
This part is pure preview — we’ll know more next week. But here’s the thing that matters regardless of what Jensen pulls out: every time Nvidia drops inference costs by another order of magnitude, the math on building AI products changes completely. Projects that weren’t viable at current costs become obvious. That cycle has been happening every 12–18 months. It’s about to happen again.
Three moves in one week. Two already landed. One incoming.
The pace of this is legitimately hard to process. What reads like a product roadmap in January is a shipping feature by March. I don’t think that’s slowing down.
— kshitij