I watched an AI use a computer better than I can yesterday. And I've been using computers for 25 years.
GPT-5.4 dropped, and the demo that broke my brain wasn't the benchmarks or the context window. It was watching the model navigate a desktop — clicking through menus, opening applications, filling in spreadsheets, switching between windows — with the casual confidence of someone who's been doing it their whole life.
Except it hadn't. It was doing it for the first time. And it scored 75% on the OSWorld desktop navigation benchmark. The human baseline is 72.4%.
I sat there, coffee going cold, trying to process what that means for someone who builds software for a living.
The Shift I Didn't See Coming
I expected AI to get smarter. I expected better reasoning, longer context, fewer hallucinations. GPT-5.4 delivered all of that — 33% fewer false claims, 1 million token context window, Thinking mode that lets you steer the reasoning process mid-thought.
What I didn't expect was for the model to become an operator.
Not a tool I use. An agent that uses tools. My tools. The same software I use, the same interfaces I navigate, the same buttons I click. Except it does it faster, and apparently more accurately than the average person.
This changes the calculus. Every time I've evaluated AI capabilities, I've been thinking in terms of "what can it write?" or "what can it analyze?" Now the question is "what can it do?" And the answer is: almost anything that happens on a screen.
What I'm Actually Feeling
Honestly? A mix of excitement and existential vertigo.
The excitement is obvious. I've spent countless hours building API integrations, writing automation scripts, and creating connectors between systems that don't natively talk to each other. GPT-5.4's computer use capability makes a significant chunk of that work obsolete. Why build a custom integration when you can point an AI at the GUI and say "do this"?
The vertigo is harder to articulate. I've built my career on being good at using computers. Understanding how software works, how to navigate complex interfaces, how to automate workflows. When AI exceeds the human baseline at exactly that skill — computer operation — it hits different than when it beats us at chess or Go. Those are games. This is work.
I don't think this means my job disappears tomorrow. The model still needs someone to define what "do this" means. It needs context, judgment, and domain knowledge that it doesn't have. But the gap between "the human decides and the AI helps" and "the human oversees and the AI does" just narrowed significantly.
The Conversation With My Past Self
If I could talk to myself from three years ago — March 2023 — and describe what happened this week, here's what would blow his mind:
"The AI has a million-token context window. It can read your entire codebase in one shot. It can use your computer by looking at your screen. It steers its own reasoning, and you can redirect it mid-thought. It's more accurate at desktop tasks than the average person. Oh, and earlier this week, a 4-billion parameter open-source model ran agent tasks on a phone."
2023-me would think I was describing science fiction. 2026-me is writing a journal entry about it because it's just... Tuesday. Well, Thursday.
The pace isn't accelerating in a way that feels fast. It's accelerating in a way that feels ordinary. Each individual step seems reasonable. Taken together, the distance we've covered in three years is staggering.
What I'm Changing
Three concrete things I'm doing differently starting this week:
1. Rebuilding my automation pipeline. I have dozens of small scripts and integrations that move data between systems — Supabase, Vercel, Google Sheets, various APIs. I'm going to experiment with replacing the fragile ones with GPT-5.4 computer use. Not as a cost optimization. As a resilience play. A model that interacts through the GUI doesn't break when an API changes its schema.
2. Using Thinking mode for architecture decisions. I've been treating AI as a question-answering tool for technical decisions. "What's the best database for X?" followed by evaluating the response. With Thinking mode, I can feed it the full context — requirements, constraints, existing code — and collaborate on the reasoning process. That's fundamentally different from querying and evaluating.
3. Layering more aggressively. Last week I talked about a 3-tier architecture: local models for routine tasks, hosted open-source for complex work, frontier APIs for hard problems. GPT-5.4 sharpens that framework. The "hard problems" tier now includes anything requiring computer interaction, sustained multi-step reasoning, or deep analysis of massive codebases. Everything else should be running locally or on cheaper hosted models.
Looking Forward
The thing that keeps echoing in my head is the OSWorld score: 75% vs. 72.4% human baseline. Not because the number itself is shocking — it's a 2.6 percentage point margin. But because it's the first time a general-purpose AI has exceeded the human average at the general-purpose skill of using a computer.
We've crossed a threshold. Not a capability line — those get crossed every few months now. A conceptual one. The foundation of knowledge work is operating a computer. An AI just demonstrated it can do that better than most people.
I don't know exactly what happens next. But I know that the builders who treat this as a footnote will wake up one morning wondering how the world changed while they were arguing about framework choices.
Build something with GPT-5.4 this weekend. Not the chat interface. The computer use API. Watch it navigate your screen. Feel the shift.
Then decide what you're going to do about it.