The AI Curse — Dan's Musings

Dan's Musings

The AI Curse

Published on June 10, 2026

The Lisp Curse is, briefly, that Lisp is so powerful and expressive that it makes it easy to write code without collaborating with others and just write it yourself. If that sounds familiar, it's because that's what's happening with AI right now.

Hundreds of stars instead of tens of thousands or thousands on GitHub for projects, dozens of implementations for the same thing, abandoned projects everywhere but that's okay because we can just use the code from them anyway because it's really easy to fix the code ourselves. Am I talking about vibe coding, or Lisp? (Both. The answer is both.)

How to fix this? Well, we can start by our attitude. The Common Lisp community came together and preferred a patchy, messier language in favor of unity through political compromise. That is how the ANSI Common Lisp standard was achieved. It still suffers from the Lisp Curse, but there is still some semblance that the community is on the same page. There are still a handful of implementations for json parsers for example, but a couple of them are very high quality and widely used.

In this regard, the Common Lisp folks fared much better than their Scheme cousins, whose whole language family was founded on the premise that if you have a better idea for how to make the language you can just make a new language and do your own thing instead of joining the heard (this is what Guy Steele did with his original paper). Such a premise makes for an extremely fractured community, and together with the Lisp Curse itself, means that in my opinion, Scheme is the most fractured of all language communities within the programming universe.

This is already happening with AI. Go ahead. Search on GitHub for a local-RAG MCP server (one that's vector database based). For such a simple universal concept you'd think we'd find a single implementation with over 10,000 stars. Instead, we find a handful of implementations with less than 1,000, max. My brother does local RAG, but he just rolled his own. I downloaded one of them, but got a segfault when it ran.

This similarity to the Lisp Curse is alarming enough, but we also see the tendency for the use of AI to run counter to collaboration. When I Google Search now, it's just easier to interact with Gemini, which means I'm not interacting with the actual blogs, which means the bloggers have less incentive to write the blogs. When I use Claude, as above mentioned, it's easier to just write my own code because then I don't have to get it to work with somebody else's slop. It makes me think there's a Liapanov stability function between AI and collaboration on the internet. The more collaboration, the more AI can feed on the tokens generated by that collaboration. The more AI, the less people are incentivized to collaborate. The less they are incenitivized to collaborate, the less tokens for AI training. The less tokens for AI training, the less AI we use and the more we collaborate. And so on.

Clearly, prioritizing community is an important step, but I think it's going to become important to make tools that make it easy to collaborate with AI. Imagine a git clone but it also stores vector indexed knowledge graph data so that every collaborator has access both to the code and to data that AI can directly use to work on that code, so that everyone that uses AI is all on the same page. We need more of those things to make AI contacts more viral. This idea of virality is disgusting the essay Worse is Better which is sort of the antithesis of the Lisp Curse: the most viral thing wins. We need to make collaboration tools that make sharing and improving a specific context engineering environment absolutely frictionless.

Until then, good luck navigating an AI-driven, fractured community that all does the same stuff but not together, a cesspool of Not-Invented-Here Syndrome and supercharged walled garden "Platforms" written by start-ups and has-beens alike, driven by seemingly straightforward business reasons that makes everyone want to keep everything in-house (again, am I talking about the AI landscape in the 2020s or the AI Lisp landscape in the 80s?). I know I'm going to need that luck.