Are LLMs Really Redefining Our Technological Choices?

LLMs and Their Influence on Technological Choices

Large Language Models (LLMs) raise concerns about their potential influence on technological choices. Indeed, they could favor the tools that are best represented in their training data, which might hinder the adoption of newer, more efficient tools.

A few years ago, the results obtained by soliciting the help of LLMs for languages like Python or JavaScript were significantly better than for less common languages. However, with the latest models and their coding agent environments, this trend may be changing.

Promising Results with New Tools

New LLM models deliver impressive performance when used with recent tools. For example, by using commands like “use uvx showboat --help / rodney --help / chartroom --help,” the models can consume a large amount of documentation before solving a problem.

Integrating a coding agent into an existing codebase, even with recent or private libraries and tools, seems to work effectively. The agent is capable of consulting existing examples, understanding patterns, and then iterating and testing its own output to fill in the gaps.

The author thought that coding agents would be the ultimate embodiment of the "Choose Boring Technology" approach, but in practice, they do not seem to affect his technological choices in that way.

Study on LLM Biases

A recent study titled “What Claude Code Really Chooses” conducted by Edwin Ong and Alex Vikati tested the Claude Code model over 2,000 times. The results show a strong bias towards the build-over-buy principle and a preference for certain technologies like GitHub Actions, Stripe, and shadcn/ui, which dominate their respective categories.

The Importance of Skills in Coding Agents

The Skills mechanism is increasingly being adopted by coding agent tools. Projects are already publishing official skills to facilitate use by agents, with notable examples like Remotion, Supabase, Vercel, and Prisma.