AI Vibecheck, Q2 2025

We’re solidly in the AI backlash backlash. Here are some thoughts about using large language model-based tooling (like Claude Code, Cursor and Goose) for software development in 2025.

On determinism

One thing that drew me to computing in the first place was the sense that, however complex computers were, they were fundamentally understandable. Computers aren’t magic. I always find it extremely empowering when I can figure out a tricky bug by peeling back a few layers of abstraction.

LLMs are closer to magic. Unless I’ve missed some new development in interpretability, you can’t stare at the matrices powering them to figure out why a model has given you a certain response. If you’re having trouble with a tool, the solution is fuzzy: try a different model; try to stuff more relevant code in the context window; IDK, have you tried writing your prompt in all caps or asking more nicely? I would find it frustrating to be responsible for the quality of such a system. Maybe someone will figure out how to make useful assertions about the quality of their output.

Nondeterminism isn’t new: probabilistic data structures are useful, and everything these days is asynchronous and distributed and horrifyingly complex. But do we have to inject more chaos into everything? I am heartened by newer tools that use LLMs to write a script or plan for invoking deterministic tools.

Also, just an aesthetic objection: dealing with tools that are confidently and cheerily wrong is annoying! Have some shame!

AI as anti-expertise

LLM tools are great at creating greenfield, prototype apps. I support the democratization of software engineering: I’m excited for more people to try building software, even if they vibe-code their way to victory.

But some of the excitement around AI tools feels like it’s trying to devalue software engineering broadly. If a CEO can clone a SaaS app in a few days, who needs those highly-paid engineers? Part of what I love about writing software is understanding a topic deeply and crafting something that will stand the test of time, and I think that will still require deep expertise.

There’s undoubtedly some “kids these days” energy here — people have long made similar arguments whenever a new technology makes the practice of programming more approachable. “Real engineers write in assembly code.” “How can you call yourself a programmer if you’ve never done manual memory management?” “These web developers are ruining the desktop with their Electron apps.” I don’t want to gatekeep software development.

I just hope that the vibe-coders still have curiosity about the systems they’re building on top of. In the near future, I see a lot of haphazardly built software, in desperate need of people to understand and maintain it. Maybe not all that different from today.