🕹️

Web dev, after LLMs

Results of 3 weeks of hacking and thinking about what web will look like. Very opinionated.

  • Important, neglected LLM modality: User Interfaces
    • “Is this just a subset of coding, which is a subset of text modality?” eh, maybe about as much as voice is a subset of text.
  • We spend like:
    • ~50% of our waking hours in front of a screen
    • ~50% of that is in front of content/data that we or our friends own (eg email, Facebook, Notion/Docs)
    • but like… <1% of that is in front of UIs that we ourselves constructed (eg personal websites)
      • (perhaps a bit more in certain lines of work, eg internal dashboards via Retool)
  • Why are custom UIs so rare?
    • The construction of UI today is painful, slow, inhuman
    • But it won’t be, soon. LLM coding is now insanely good
  • But meanwhile, UIs are the winning form factor for “Human <> Computer”
    • Maybe we’ll get to BCI in like 5 years but until then…
  • [TODO] diagram of existing web app ecosystem, and what we’re trying to integrate
    • Code editing (Cursor)
    • Code hosting (Github)
    • Web deployment (Vercel, Namecheap)
    • Database (Supabase)
  • Why web?
    • Web is widely distributed, every device you work with has a web interface, lowest common denominator
    • SaaS distribution model is amazing — enables realtime updates for your service
    • Bypasses app store gatekeeping (on mobile) or extra download step (on desktop)
    • Possible upstarts: VR, AR, BCI
  • buni vs Cursor, Devin, Replit Agent, etc:
    • These are designed for code silos. Each app doesn’t talk very much to other apps
    • buni is aiming for an extremely composable world, think NPM or wikipedia.
      • Should derisk: maybe composability doesn’t matter, maybe archipelagoes is how code wants to be?
  • Make it fast!
    • On every level, make the feedback loops quick
      • Build, deploy the app asap
      • Make sharing instantaneous
      • Get user feedback and suggestions easily?
        • record all user interactions, to replay/train on?
  • open > closed
    • Want to experiment with a paradigm everything is public readonly; writing is reserved for keyed systems
      • Helps the context flow freely, for LLMs and humans
      • (better aligned with fundamental economics of information)
    • Revealed preference of internet is that privacy, personal data just doesn’t matter, compared to convenience or systems/tools that do things you want. E.g. Google
  • multiplayer > single player
    • today “ChatGPT”, tomorrow “SlackGPT”
    • But also, multiplayer like Twitter or Wikipedia are multiplayer
    • new features and improvements could propagate throughout an app, bottom-up
      • today, web apps are gatekept by dev teams
  • down the line: marketplace for agents (or: NPM for agent infra, prompts)
    • And also $/credits attached, like Github sponsors kinda
    • Theory is that agent infra is bottlenecked on good UI & code deployment
    • Can stick in a bunch of cool crypto mechanisms, smart contract stuff (prediction markets, perp swaps, etc)
    • Can bootstrap into agents (or corporations of agents), with long-running identities
    • Beyond building prototypes of web apps, I'd eventually like to figure out how to build an open and composable web; how LLM agents live and interact through this; even economic & governance mechanisms for open source development. It's a ridiculous amount to consider, which is exciting but also distracting/scary/suspicious from an outside view, and also flies in the face of standard startup advice.

Appendix

  • Q: What’s hard for Austin to build buni?
    • Building app vs platform
      • Building a web app: you get to be dictator, monolith/monorepo, everything stays up to date
        • Managing multiple versions is hell
      • Building an npm-like platform or ecosystem: all kinds of weird versioning stuff, complexity of references explode. Probably need to straddle the line
        • Maybe LLMs can parse all your downstream deps and tell you when things work…?
        • Also: reliability matters much more
          • (cf Streamlit on reliable platforms)
    • Scope is ginormous — kind of encoding my entire philosophy of how code or the web “should” work
    • many degrees of freedom when designing a whole ecosystem from scratch
    • still hard to frame the killer MVP
      • Maybe “prompting Twitter for changes” isn’t actually that compelling?
        • Ideally your apps kind of figure out what you need? Coming up with good prompts or searches is expensive, hard
          • A bit easier after you’ve used an app for a bit and think something is wrong — that’s product sense though, maybe not everyone has product sense.