They gave me a Claude subscription and told me to get tokenmaxxing. AI adoption has been mandated from management, unlike past silver bullets. This triggered my contrarian side—someone who never wrote code telling engineers what to use. Over 3 months I found Claude somewhat useful, but don't ask it to write code.

AI isn't intelligent—it's a large neural network for NLP. "Agentic workflow" is just adding indirection. It's good at search: answering pointed questions, summarizing texts, running parallel searches. But LLMs flood the internet with word salad, diluting information. Freya Holmér published a video on AI-generated content in the SEO arms race.

A particular area where LLM loops prove useful is internal company knowledge bases—wikis, Slack, Confluence. I've found answers written before my time because Claude turns natural language into search queries for synonyms. This isn't new tech; Google did synonym generation 20 years ago. But from a user perspective, it's better than searching "UI" and getting no results because the page says "User Interface."

Hallucinations are inherent—token generation is based on statistics, not logic. I've experienced them on all models, from free Copilot to paid Claude. Always check primary sources.