Software Engineer Details Paradox of Using LLMs Despite Valid Criticisms

Software engineer Theo Charis has published an article titled "The LLM Critics Are Right. I Use LLMs Anyway," detailing the cognitive dissonance experienced by developers who utilise Large Language Models while agreeing with their ethical and practical criticisms. Inspired by observations at the Local-First Conf in Berlin, where attendees criticised AI technologies while actively using tools such as Claude Code, Charis outlines concerns regarding the erosion of trust in open-source software, the potential devaluation of junior engineers, geopolitical risks, and environmental impacts.

Charis describes specific concerns regarding LLMs, including the degradation of open-source software trust, the devaluation of junior engineers, geopolitical risks, and environmental impacts. He notes that projects like Zig and Gentoo are refusing to accept LLM-generated pull requests, highlighting a broader industry shift towards filtering automated contributions. Armin Ronacher, creator of Flask and founder of Earendil, stated at Local-First Conf that his company auto-closes almost all pull requests and issues generated by LLMs to protect their workflow, while encouraging humans to continue contributing.

The author argues that LLMs can enhance human thinking when used selectively for brainstorming, refinement, and acting as a "rubber duck," provided the human retains control over core ideas. Charis disclosed spending nearly 10,000 USD on LLM tokens in June 2026, a figure he described as extreme, leading him to adopt cheaper models like GLM 5.2 via OpenRouter for code execution and use Fable more selectively. He emphasizes that written text should remain from humans to humans, with LLMs serving only to sharpen existing human thoughts rather than generate them from scratch.

Specific techniques mentioned include the "grill-me" skill to force critical thinking and using local open-weight models to maintain independence from major corporations. The article references specific technical techniques adapted from other developers, such as Matt Pocock’s "grill-me" skill and Anselm Eickhoff’s method of using LLM hallucinations to test user expectations. Charis notes that projects like Zig and Gentoo are refusing to accept LLM-generated pull requests, highlighting a broader industry shift towards filtering automated contributions.

The US government issued an export-control directive on June 12, 2026, forcing Anthropic to abruptly disable its Fable 5 and Mythos 5 models for all non-US customers. Charis argues that while the AI bubble may burst, open-weight models running locally on personal hardware offer a safeguard against such geopolitical disruptions. He concludes that the value of LLMs lies in their ability to amplify human intent, provided the user maintains rigorous control over the output and retains the ability to distinguish quality from "AI slop."