Blog
Thoughts on AI browser agents, browser automation, and building dassi.
-
I Quit ChatGPT.com. I Use GPT-5.2 More Than Ever.
GPT-5.2 is the best model OpenAI has ever shipped. But chatgpt.com can't see your browser. Run it inside a browser extension with full page context instead.
-
We Ran 643 Real Browser Tasks. 93% of Answers Were Correct.
Dassi scored 93.2% answer accuracy on the WebVoyager benchmark — 643 tasks across 15 live websites, no mocks, no sandboxes. Here's what we learned.
-
Claude Code's Chrome Extension Keeps Disconnecting. I Dug Into Why.
Claude Code routes Chrome commands through Anthropic's cloud relay. Combined with MV3 Service Worker timeouts, five failure modes explain why it keeps dying.
-
Your ChatGPT Subscription Now Works Inside Your Browser. No API Key Required.
Dassi now lets you sign in with your ChatGPT Plus or Pro account and use GPT directly in Chrome. No API key setup, no extra cost, no copy-paste.
-
Gemini 3.1 Pro Just Dropped. Dassi Already Runs It.
Gemini 3.1 Pro launched with 1M context, three-tier thinking, and strong benchmarks. Dassi supports it now — plug in your Google AI Studio key and go.
-
Best AI Browsers in 2026 (Including One That's Not a Browser)
Comet, Atlas, Dia, Genspark, Sigma, Fellou — AI browsers are everywhere. I tested them all. Here's what's good, what's not, and why the best option might not be a browser at all.
-
ChatGPT Can Order You an Uber Now. It Still Can't Read the Page You're On.
ChatGPT's new app integrations with DoorDash, Spotify, and Uber solve the API problem. But the browser problem — AI that can't see what you're working on — remains wide open.
-
Cloud Browser Agents Can't See Your Tabs. That's the Whole Problem.
AI automation hype assumes cloud agents can do your browser work. But cloud browsers can't access your logged-in sessions, local tabs, or real context.
-
Cloud Browser Agents Can't See Your Tabs. That's the Whole Problem.
Cloud AI agents spin up remote browsers you're not logged into. Local browser agents use your real Chrome, already authenticated. Why that gap matters.
-
Nyne Raised Millions to Give AI Agents 'Human Context.' Your Browser Already Has It.
Nyne wants to build a context layer for AI agents. But browser-native agents already see what you see — no middleware required.
-
GA4 Direct Traffic Is Lying to You. I Spent a Weekend Proving It.
Direct/(none) dominates your GA4 reports but most of it isn't direct traffic. A step-by-step debugging process for SaaS owners who want real answers.
-
Lovable's $100M Month Proves What Actually Wins in AI
Lovable hit $100M/month with 146 people by putting AI where users already are. The pattern is clear: distribution beats capability every time.
-
I Asked My AI to Navigate Google Cloud Console. It Did the Whole Thing.
Instead of spending 20 minutes clicking through Google Cloud Console to set up OAuth, I let Dassi handle it. The pattern works for every complex admin panel.
-
Anthropic Built a Code Review Tool. For Its Own AI's Code.
Anthropic just launched a tool to review AI-generated code, admitting even they don't trust raw AI output. The same logic applies to every AI task.
-
AI Browser Agent Leaderboards Miss Real-World Usage
A new AI browser agent leaderboard hit Hacker News. But WebVoyager scores don't tell you if an agent can draft your email or fill your expense report.
-
The AI Productivity Debate Is Over. Distribution Won.
Skeptics and optimists keep arguing whether the AI productivity boom is real. They're both asking the wrong question. It's about where the AI runs, not whether it works.
-
The AI Productivity Takeoff Is Real. You're Just Not in the Plane.
Brynjolfsson says the productivity data finally moved. The OpenCode creator says you're kidding yourself. Both are right — and browser agents explain why.
-
Nvidia Just Pulled Back from OpenAI. Your AI Setup Should Be Just as Ruthless.
Jensen Huang is distancing Nvidia from OpenAI and Anthropic. If the company powering all of AI won't commit to one provider, you shouldn't either.
-
GPT-5.2 Is Incredible. You're Running It in the Wrong Place.
GPT-5.2 on chatgpt.com can't see what you're working on in your browser. Run it with full page context instead and it becomes 10x more useful for real work.
-
You Switched from ChatGPT to Claude. You're Still Locked In.
Everyone's switching from ChatGPT to Claude after the Pentagon fallout. But that's just trading one AI lock-in for another. BYOK solves model lock-in.