As an experienced Large Language Model (LLM) user, I've been reflecting on my relationship with generative AI. Despite my critical stance on some aspects of modern generative AI, I've found ways to leverage LLMs effectively in my professional and personal endeavors. Over the past decade, I've explored various text generation tools, from char-RNNs to GPT-3 and beyond. While I'm adept at working with LLMs, I don't use them as frequently as one might assume, especially for certain tasks. How I Use LLMs Professional Work At BuzzFeed, where I work as a Senior Data Scientist, I often interface with LLMs via backend UIs and APIs rather than user-friendly frontends. These interfaces offer more control, allowing me to set system prompts and adjust the "temperature" of the generation. Setting the temperature to 0.0 ensures deterministic outputs, while a slightly higher setting (0.2 - 0.3) allows for some variance. One of the key benefits of LLMs in my professional role is their ability to quickly solve problems. For instance, I've used them for hierarchical labeling of articles, a task that would otherwise require extensive manual effort and R&D. These projects often start as offhand ideas during standups or Slack messages but are completed within hours, thanks to LLMs. They follow the Pareto principle, providing a quick 80% solution, though the remaining 20% still requires human iteration and testing to refine and catch any hallucinations. Writing I don't use LLMs to write my blog posts, despite the assumption many readers might make. My blog covers recent tech events and has a unique, often irreverent tone that LLMs struggle to emulate accurately. Even with sophisticated prompt engineering, the results tend to sound too much like generic Marvel movie dialogue. Ethically, I also value authorship integrity, preferring to write in my own words. However, I do use LLMs in a creative way: I feed drafts of my blog posts to Claude and ask it to generate critical comments as if from a cynical Hacker News user. This helps me identify weak arguments and improve my writing organically. Companionship While many people use LLMs for personal chat and companionship, I avoid this. Being an introvert, I find it difficult to form a genuine friendship with an entity trained to be overly friendly and prone to hallucinations. Despite my proficiency in prompt engineering, I remain skeptical of the practicality and ethical implications of using LLMs for emotional support. Coding LLMs have significantly enhanced my coding efficiency. I started using them to write regular expressions, a task that can be time-consuming. Now, I consult LLMs like Claude Sonnet for coding questions, particularly those involving specific constraints and less popular libraries. For instance, when I needed to write Python code to composite multiple images using the Pillow library, Claude's suggestions were accurate and helpful, saving me time and reducing the likelihood of errors. For more complex coding tasks, especially involving data science, LLMs are less reliable. They struggle with outputting the results of mathematical operations consistently, and I often find discrepancies when using them with newer or less documented libraries like Polars. In these cases, I prefer to rely on my own expertise and existing documentation. In-line code suggestions from tools like GitHub Copilot, while theoretically useful, can be distracting. Reviewing and integrating these suggestions breaks my flow, leading to a net neutral or even negative productivity impact. I find it more efficient to use LLMs for ad-hoc questions through a web UI. Agents and Vibe Coding Agents and Multi-Chain Processes (MCP) are rebranded versions of the Tools paradigm introduced by the ReAct paper in 2022. These advancements in context window size and prompt adherence have made agent workflows more reliable. However, I haven't identified any novel use cases that surpass the simplicity and effectiveness of LangChain, which I found adequate a few years ago. Simple MCP implementations have become more complicated and confusing, and I avoid vibe coding with Claude Code or Cursor due to the risk and cost associated with it. While vibe coding can provide quick, rough solutions for personal projects, it's unprofessional for high-stakes, serious projects. Industry Insights and Company Profiles AI critics like Ed Zitron argue that the LLM industry is unsustainable due to high operational costs and lack of revenue. While it's true that providers face economic challenges, this doesn't invalidate the usefulness of LLMs in practical applications. Open-source models like Qwen3 and DeepSeek R1, which perform comparably to commercial models, can be viable alternatives if major providers fail. These models can be hosted on dedicated LLM hosting platforms, ensuring continued availability and utility. The conversation around LLMs has become polarized, making it challenging to present neutral views. Nevertheless, LLMs are undeniably valuable tools when used appropriately. As a data scientist, my approach is to select the right tool for the job, recognizing that LLMs can be highly effective in certain contexts but may fall short elsewhere. Using LLMs is about balancing quick, efficient problem-solving with the need for precision and reliability. In the evolving tech landscape, staying flexible and open to new methods is crucial. While I currently use LLMs selectively, the future may bring innovations that change my perspective. However, for now, I'm content with my current workflow, where LLMs enhance productivity without compromising quality.