I’m serious.
TL;DR
Theo’s thesis is blunt: closed-source teams can’t be trusted with AI-era velocity — he argues AI makes developers “10 to 100 times faster” at shipping changes, which also means they can degrade apps 10 to 100 times faster, and users have no recourse if the code is closed.
A high-school intern, Yash, changed how Theo thinks about software boundaries — instead of working around broken dependencies, Yash used patch-package, reverse-engineered T3 Chat’s Webpack bundle, added local-model support, and even patched the AI SDK to support image generation before the upstream library did.
Theo open-sourced T3 Code partly as a performance and accountability play — after being frustrated by Codex, Cursor, Notion, and even macOS 26 getting worse over time, he wanted users to be able to fork the app if his team ever took it in the wrong direction.
Cursor is his main case study in ‘AI slop’ — despite praising Cursor’s model harness as genuinely strong, he says the product has layered “slop” on top of performant open-source VS Code, with Glass somehow feeling even slower and crashier in real use.
He thinks AI has made forking and customization dramatically cheaper — with agents handling merge conflicts and upkeep, Theo says maintaining forks is easier than ever, citing T3 Code’s 30,000 users and 1,100 forks plus non-developers like XLT Jake and Quinn Nelson forking Lawn for their own workflows.
Claude Code gets the harshest criticism because its closed source makes the least sense — Theo calls it absurd that a terminal tool and its SDK stay closed, accuses Anthropic of hiding behind “secret sauce,” and points to the accidental release of source maps and subsequent DMCA takedowns as evidence of embarrassment, not principle.
The Breakdown
Why open source suddenly feels urgent
Theo opens by admitting most of the software we all rely on — macOS, Windows, Notion, Linear, Slack, Discord — is closed source, and that used to feel normal. Now, especially in the AI era, he says he’s reaching the point where he doesn’t even want to try new tools if they’re closed, because when they get worse, users are stuck.
The old economics of software are breaking
He walks through the traditional bargain: software was expensive to build, so selling closed binaries and APIs made sense, and that model helped create a real professional developer class. But as code generation gets cheaper, the logic changes — the frustrating part is no longer just paying for software, it’s being unable to fix a small thing when the people shipping it make it worse.
Yash, patch-package, and the mindset shift
The turning point is his intern Yash, a “ridiculously talented” high schooler who first showed up with a userscript that reverse-engineered T3 Chat’s Webpack bundle to add features and local model support. Yash’s superpower, in Theo’s telling, is that he doesn’t respect artificial boundaries: if a dependency is wrong, he patches it, files the PR, and keeps moving instead of building elaborate workarounds.
The patch that scared him — and proved the point
Theo gives a concrete example: T3 Chat needed progressive image generation, but the AI SDK only handled text, so his own workaround code was messy. Yash went further than Theo would have, patching the SDK itself to support image generation, replacing Theo’s custom paths with something cleaner and more stable until the official support landed upstream — and that success has been “haunting” him ever since.
Codex, Cursor, and software rotting in real time
He then pivots to a month of heavy use in the Codex app, which he loved at first, but every update felt like rolling dice on performance. Cursor gets the harsher story: Theo says he confronted the team in person after repeated lag, freezing, and crashes, including Julius from T3 being unable to use Cursor Glass with just two codebases open; the answer he got — prioritize usefulness first, performance later — completely shattered his confidence.
The core rant: AI is accelerating slop
This is the heart of the video: Theo says closed-source dev teams are “sloppifying” previously usable tools because AI lets them ship bad changes much faster. He names Notion breaking column moves during his stream, macOS 26 being a “shitshow,” and Cursor layering bad AI-era code on top of solid open-source VS Code, even while admitting Cursor’s underlying harness is legitimately excellent.
Why he open-sourced T3 Code and what happened next
Out of that frustration, he built T3 Code in Electron rather than native, after concluding Electron’s text and scrolling performance was simply better than AppKit, SwiftUI, or Tauri on real workloads. Then he did the thing he normally avoids — open-sourcing code he believes could be valuable to huge companies — and the response was wild: roughly 30,000 users and 1,100 forks, which he sees as proof that users want the right to inspect, adapt, and hold the product accountable.
Lawn, Claude Code, and the bigger fight ahead
He extends the argument with Lawn, his tiny ~$600 MRR video-review app, which still got meaningful forks from creators like XLT Jake and Snazzy Labs who used AI to customize it for their teams. He ends on a furious critique of Claude Code and Anthropic for closing the source of a terminal tool and SDK, then broadens back out: if AI makes it trivial to copy and mutate software, the answer isn’t tighter control over worsening products — it’s better open alternatives, better support for maintainers, and a refusal to accept the “enshitification” of everything.