As an Experienced LLM User, I Actually Don't Use Generative LLMs Often
Source: minimaxir.com
The author, an experienced LLM user and data scientist, explains that while they find generative LLMs useful for specific, constrained tasks like text classification, summarization based on provided context (like a style guide), and generating code snippets for well-defined problems (especially regex or specific library functions), they don't use them frequently for core writing or complex coding tasks. They prefer API access over standard frontends for better control (system prompts, temperature) and use Claude Sonnet more than ChatGPT. They find LLMs less useful for their blog writing (due to unique style and ethics), complex data science work (especially with newer libraries like polars or R/ggplot2), and dislike distracting inline coding assistants like Copilot. They view LLMs as a useful tool but caution against over-reliance, hallucination, and hype, believing they have real but specific applications, distinct from AGI promises.
This blog post by Max Woolf explores his nuanced approach to using large language models (LLMs) as a Senior Data Scientist at BuzzFeed. Key points include:
-
Limited Use of Generative LLMs – Despite his experience, he rarely uses LLMs for writing or general text generation but finds value in specific professional applications.
-
Effective LLM Usage – He prefers API access over consumer-facing interfaces (like ChatGPT.com) to control parameters such as system prompts and temperature settings, which impact output quality and reliability.
-
LLMs in Professional Work – He has used LLMs at BuzzFeed for:
- Hierarchical Taxonomy Labeling – Assigning categories to articles without manual training data.
- Semantic Clustering Labels – Generating descriptions for clusters of articles.
- Grammar Checking – Using LLMs to validate writing against BuzzFeed’s style guide.
-
Text Embeddings vs. Generation – While generative models are useful for specific tasks, text embeddings (which map text into numerical vectors) are often more practical for recommendation systems and clustering tasks.
-
LLMs for Writing – He does not use LLMs to generate his blog posts, citing concerns about authorship ethics and hallucinations related to recent tech trends. However, he does use LLMs to simulate critical feedback (e.g., Hacker News comments) to improve his writing.
-
Coding Assistance – He finds LLMs useful for:
- Regex help and structured code generation (e.g., handling images in Pillow).
- Suggesting optimizations when writing code but remains cautious of hallucinations.
- Avoiding AI-powered autocomplete tools like GitHub Copilot, which he finds distracting.
-
Skepticism of Agents and "Vibe Coding" – He does not find AI-driven coding assistants or autonomous coding agents particularly useful, preferring more reliable workflows.
-
LLM Industry Uncertainty – He argues that LLMs are useful despite financial sustainability challenges faced by companies like OpenAI. Even if proprietary models disappear, open-source LLMs will continue to provide value.
The article emphasizes pragmatism—LLMs are tools that require careful application, and their effectiveness depends on thoughtful constraints and critical evaluation. Would you like me to highlight anything more specific?