▶ YouTube Summaries

• https://www.youtube.com/watch?v=wEAb0x3wTRc
• Original title: I don't have time to build these things, will you?

In the agent era what you build matters more than how, so Theo hands out the backlog of products he wishes existed but has no time to build. Each is a "boil the ocean" rebuild that was too expensive to attempt before but is now viable because agents slash the cost of shipping big infrastructure. Five concrete ideas plus a plea for more benchmarks.

Better NPM / NPX

NPM is critical but rotting. Publishing is too hard and irreversible — a typoed version (e.g. TanStack Query) can never be revoked because NPM fears breaking old builds. Security is a constant target and every new guard hurts honest devs. Wishlist: revoke releases under a threshold (<100 installs or <5h old); pay an API/credit card to have an LLM audit and diff every release for a safety score; rich metadata at install time (obfuscated? open source? who published? permissions used?); kill name-squatting (the real TanStack NPM name is held by a squatter) with human/agent verification. For NPX specifically: the install-confirm prompt shows only a useless version number — it should surface size, author, permissions, and a third-party security score, which matters doubly for agents that run `npx` from a skill.md and could unknowingly execute a hijacked package. Companies like Socket already detect exploits before NPM does — lots of room to build on top of / instead of NPM.

Reinvent source control (post-Git)

Git was a huge leap but is the wrong primitive now. Permissions live at the repo level, not the change level — hence you can't commit .env files, have private files/branches/PRs, or delay when a merge becomes public. That last one matters: Linux security fixes get reverse-engineered into zero-days the moment the patch is public; you should be able to merge, cut a release, and patch distros before the code is visible. Open source shouldn't mean 100% public 100% of the time (Claude Code would more likely be open source if unfinished work could stay hidden). Praises JJ (snapshots/tags over branches/commits) for fixing ergonomic pain, though it doesn't address permissions. Work trees are "atrocious" (an agent checked out main and held it hostage). Also argues source control shouldn't require a real filesystem — tools like jsbash run agents fully in-memory.

Dropbox for devs

APFS is terrible at creating many small files: a cached PNPM install that takes ~7s on Ubuntu takes ~31s on an M4 Mac and up to ~140s on an M1 Ultra — an fsync/filesystem problem that makes spinning up agent environments painful. Beyond that, managing code across many machines (Mac Minis, Ubuntu box, cloud agents) means stale mains, missing env vars, and inconsistent directory layouts. Want a single synced `code` folder identical on every machine — like Dropbox/Google Drive with a `.gitignore` equivalent — that lazily pulls file contents on access. Started a project (FS2) but it didn't go far enough. The blocker isn't skill, it's "token budget and patience."

A new, open mobile OS

Apple (30% cut, arbitrary bans, Info.plist hardcodes a developer into source control) and Android (opaque Play Store bans, near-zero Google investment) are a duopoly that makes app distribution miserable. Nostalgia for CyanogenMod/Paranoid Android era when customizing the OS was easier than building an app — Paul Henschel (Zustand, React Three Fiber) started on Paranoid Android and left mobile because app dev sucked. BlackBerry 10 proved you can ship a different OS that still runs Android apps. Wants an experimental, hackable mobile OS that runs Android apps but encourages building/forking on-device — "NPM but for mobile," scan a QR code and the app is on your phone. Open bootloaders still exist, so now may be the last chance.

Better Slack (Facebook Workplace model)

Slack is built for sending messages, not reading, prioritizing, or status — and it's worse for agents. No inline replies, threads fall stale even when active, lock-in via cross-company shared channels. Wants the Facebook Workplace "post" primitive: posts sit between channel and thread, support infinite nesting, and bump back to the top when someone (or an agent) replies. Workplace was the closest thing to good team context management — and Meta shut it down two weeks ago. Pair it with something like Hermes Agent so agent replies bump work to the top. Wants it as an open standard to slowly replace Slack.

More (weird) benchmarks

We need benchmarks written by ordinary devs, not just labs. His joke SkateBench (name a skate trick from a description — part grammar, part niche-English, part 3D spatial reasoning) drew real interest from researchers. Save the AI tasks that fail you as reproducible benches (Git bench, language-specific benches, image/medical-scan benches). The moment a weakness is measurable, labs race to fix the score — building a bench that shows models suck at something is the best way to get it fixed.