Why I'm Ditching Closed Source for Open Source AI Tools
AI makes software cheap to build, but closed source tools like Cursor are degrading in quality—open source lets you fix them, as Theo's intern Yash proves by patching everything.
Closed Source Dominance Is Cracking Under AI's Weight
Theo, creator of T3.gg, admits most daily tools—MacOS, Notion, Linear, Slack—are closed source, but he's done tolerating it. Historically, software justified closed licensing because writing code was expensive and rare. Developers commanded high pay for turning requirements into reliable source code, while closed vendors sold binaries or APIs. Open source, sparked by frustrations like Richard Stallman's, offered fixes but rarely paid maintainers well. Companies like AWS profited massively by hosting open projects like Elasticsearch without contributing, prompting license changes (e.g., Elasticsearch's SSPL shift) and drama (Redis Labs fallout).
AI flips this. Code generation slashes the 'hard part' cost, making proprietary lock-in less defensible. Theo notes: "Software stopped being expensive to make... We were paid well cuz we could turn all of that into the right source code." Yet closed tools trap users: rebuilding Slack is feasible technically, but network effects (shared channels) and infrastructure lock-in prevent adoption. Without source access, small bugs fester, and regressions can't be patched.
Quote: "I'm nearing the point where I just am not interested in trying new solutions if they're closed source." (Theo on his hardening stance, highlighting how AI amplifies frustration with unfixable flaws.)
Yash's Patch-First Mindset Exposes Artificial Boundaries
Theo credits high school intern Yash—one of the "most talented devs I've ever worked with"—for rewiring his thinking. Yash built a user script reverse-engineering T3 Chat's (closed source) Webpack bundle to inject AI SDK client-side for local models. Hired on the spot, Yash ignores code ownership boundaries, using patch-package aggressively.
Patch-package lets you edit node_modules files, generate .patch files, and auto-apply them on install (native in Yarn/Pnpm). Ideal for one-liners: remove logs, fix obscure bugs. Yash quadrupled patches in weeks, then upstreamed PRs without hesitation. Example: T3 Chat used AI SDK (text-only then); Theo hacked custom image gen paths for OpenAI/Gemini, which sucked—rough, non-progressive, inconsistent with text flows.
Yash patched AI SDK to add image gen natively, deprecated Theo's hacks, enabled progressive updates. Shipped stably via patch; later, upstream merged it, patch deleted. Theo: initially terrified ("scary as fuck"), now converted. At Twitch, Theo worked around issues across 7 teams for one-liners; Yash "opens the door and walks right through."
This scales with AI: generating patches or features is trivial now. Result? Deeper dependency understanding, faster iteration. Theo applies it personally: more PRs, fewer workarounds.
Quote: "Yash just doesn't perceive these boundaries... when he hits a wall because some boundary that's in the way is blocking something he just opens the door and walks right through." (Theo contrasting Yash's fluidity with corporate silos, showing how it accelerates shipping.)
AI Coding Tools Prove Closed Source Can't Be Trusted
Theo's thesis: "Closed source developers cannot be trusted with AI." Tools like Cursor and Codeex started strong but regressed via AI-generated slop. Codeex: Theo used it 90%+ for a month—polished UX shifted his workflow. But constant updates were a coin flip: better or unusable (lags in long threads, complex codebases). Yesterday: more complaints.
Cursor: performance tanked despite VS Code's solid TypeScript base. Glass (new from-scratch UI) somehow slower, crashing with two codebases. At Cursor's office event, Theo grilled them: "What the fuck is going on with performance?" Response: "We're prioritizing making it work and useful first... not going to have all the performance issues inherent to VS Code." Theo calls BS—VS Code is performant gold; Cursor layered Sonnet 3.5-era slop ("a liability") atop it.
Yet Cursor's core shines: harness makes flaky models (Gemini 3/3.1 Pro, Opus) reliable. Claude Code often fails where Cursor succeeds. Julius (T3 lead) couldn't use Glass for T3 Code integration—crashes galore. Theo urges: hire a Head of Performance to scream louder than users.
Trade-offs stark: closed source hides slop, blocks fixes. Open source exposes issues, invites contributions. Theo open-sourced T3 Code (not Chat yet, but considering); won't touch new closed tools.
Quote: "Closed source developers cannot be trusted with AI. They are taking things that are for the most part usable that have their quirks and problems and they are sloppifying them to the point where they don't fucking work." (Theo's core thesis, backed by Cursor/Codeex regressions, explaining quality erosion in AI-heavy teams.)
Open Source Unlocks AI-Era Customization and Reliability
Theo's pivot: prioritize open source for mucking internals—fun, educational, improves skills. AI lowers barriers: generate patches, add features. No PR pressure; just fix locally if needed. Frustrations compound: software "degrading over time," unfixable without source.
WorkOS sponsor ties in: enterprises (OpenAI, Anthropic, Carta) need scalable auth/onboarding. Closed roll-your-owns fail Fortune 500 scale (e.g., ADP for 10k devs). WorkOS balances DX with enterprise weirdness—self-serve admins, Slack-responsive support.
Progression: closed enabled dev profession; open fixed pains; AI commoditizes code → open wins. Theo processes Yash's influence (half his age) to adopt boundaryless fixes.
Quote: "It has never been easier to talk to a company ask for things changes whatever else we need and have them just come in and help." (Theo on WorkOS support, contrasting responsive vendors with unfixable closed tools like Cursor.)
Key Takeaways
- Adopt patch-package for any JS project: edit node_modules,
npx patch-package <pkg>, auto-apply on install—upstream PRs when logical. - Ignore artificial boundaries: if a dependency blocks you, patch it first, PR second—no team drama excuses.
- Distrust closed source AI tools: Cursor/Codeex prove AI slop regresses performance; demand source to fix.
- With AI, prioritize open source: cheaper to build/customize, deeper learning via internals.
- For enterprise scale, use WorkOS early: handles ADP/SSO for big bets without custom hell.
- Open source your side projects (like T3 Code): attracts talent like Yash, enables community fixes.
- Evaluate tools by update delta: random better/worse? Closed source roulette—switch to open.
- Hire performance obsessives: slop layers kill DX; yell louder than users (Theo's Cursor advice).