Author argues LLMs disappoint on complex tasks while local open-weights models suffice for simple coding questions

The author uses LLMs daily to accelerate work but finds their answers 'half-baked' and requiring significant manual effort to reach completion, contrasting with their earlier skepticism about NLP projects in early 2017.

Open-weights local models (such as gemma3 or gemma4 via ollama) can handle simple coding questions adequately, allowing users to avoid search engines, StackOverflow, and paid LLMs while preserving privacy.

Large companies can deploy a local LLM filter layer for employees to query first, falling back to paid external providers like Claude or OpenAI only when needed, thereby reducing AI costs.