Aaman Lamba dropped a piece this week arguing that most people are fooling themselves about AI productivity, that the gains are inflated self-reports from developers who confuse typing speed with shipping speed. And Erik Brynjolfsson’s latest data says the opposite, that the productivity takeoff is now statistically visible across multiple sectors. I’ve been watching this argument ricochet around LinkedIn and Hacker News for days, and the weird part is that nobody seems to notice they’re debating a distribution problem like it’s a yes-or-no question.

Productivity gains have an address

Brynjolfsson’s numbers are real. Sector-level output per hour worked went up in software, financial services, customer support. But zoom in and the gains are not evenly spread across workers within those sectors. They cluster around teams that embedded AI directly into their working environment, the ones using Cursor in their IDE, AI inside their CRM, agents that read support tickets without anyone copying text between windows. The teams still alt-tabbing to ChatGPT and pasting screenshots of spreadsheets did not show up in the positive column.

Lamba is describing that second group with precision. And he is right about them. The developer who spends ninety seconds asking GPT-5.2 to write a function, then four minutes adapting the output to a codebase the model never saw, is not actually faster. The time saved on generation gets eaten by integration overhead, and the self-reported “I’m way more productive” feeling comes from the dopamine of watching code appear instantly, not from any measurable change in what ships by Friday.

So the debate is not really about whether AI productivity gains exist. It is about who gets them and why, which turns out to be a much less exciting question for Twitter but a much more useful one for anyone trying to actually get faster at their job.

The copy-paste tax keeps compounding

I tracked my own workflow for a week back in January, and the overhead of moving context between my browser and a chat window averaged fifty-three minutes per day. Not because I am slow or bad at prompting. Because the workflow itself is broken at the seams. Every time I copied text from a web page into ChatGPT, I was making an editorial decision about what context to include, and I was usually wrong, leaving out the one detail that would have made the AI’s response actually useful rather than generically polished.

Gloria Mark’s research at UC Irvine puts the cognitive recovery cost of a complex context switch at around twenty-three minutes. So the chat-window AI workflow is not just adding steps. It is adding the exact kind of steps that human cognition handles worst: rapid switching between unrelated interfaces while trying to maintain a mental model of what information lives where. And the cruel irony is that the AI itself works fine. The model is not the bottleneck. The seam between the model and your work is the bottleneck, and no amount of prompt engineering fixes a seam.

What Cursor got right

Developers figured this out already, which is probably why they show up strongest in Brynjolfsson’s productivity data. Cursor did not succeed because it shipped a better model. It succeeded because it put the model inside the editor, where it could read your files, see your cursor position, understand what you just changed. The AI stopped being a destination you visited and became a layer over the place you already worked.

But most knowledge workers do not live in an IDE. They live in a browser. Email, spreadsheets, CRMs, project management tools, research, procurement forms. And until recently there was no equivalent of “put the AI inside the browser” for that work. You either used a chatbot in a separate tab or you used nothing.

Browser agents changed that equation. A browser agent sits in your browser, reads the page you are on, and acts on it without requiring you to become the integration layer. No copying, no pasting, no rebuilding context in a chat window. The AI sees the email thread before drafting a reply. It reads the form fields before filling them out. It pulls data from the table you are looking at instead of asking you to describe it.

The debate resolves itself once you move the AI

Dassi runs in your browser’s side panel with whatever model you want, GPT-5.2, Claude, Gemini, bring your own key. And the productivity difference is not subtle. It is the difference between being in Brynjolfsson’s productive group and being in Lamba’s skeptical one.

I do not think the AI productivity debate was ever really about AI. It was about proximity. The people getting faster are the ones whose tools eliminated the distance between the AI and the work. Everyone else is just running a relay race where they personally carry the baton between two buildings, convinced they are sprinting because the AI leg takes two seconds.

The models will keep getting smarter. But the productivity paradox won’t resolve itself through better benchmarks. It resolves when the AI stops living in a separate damn window.