How I Use "AI"
Large language models are described as a major productivity boost for many programmers, especially when used as a patient “smart coworker” for boilerplate code, configuration files, debugging, and learning unfamiliar tools. Commenters emphasize that these systems work best when the user already has domain knowledge, treats outputs as drafts to be verified, and uses them to navigate large codebases or dense documentation rather than as an unquestioned authority. Others remain skeptical, pointing to frequent hallucinations, poor performance on niche or high‑stakes tasks, environmental and ethical concerns around training data, and the potential impact on developer jobs and professional responsibility.
Overall sentiment
- Many commenters say the article closely matches their own experience: LLMs are major productivity boosts for coding and research, but not magic or fully reliable.
- Others report the opposite: despite repeated efforts, they haven’t found LLMs consistently useful for “serious” or complex tasks.
Effective use cases
- Coding “glue” and boilerplate: shell scripts, YAML/config (Docker, k8s, Terraform), MVC models, spreadsheet formulas, unit tests, test data.
- Learning and comprehension: explaining unfamiliar APIs, frameworks, math notation, kernel subsystems, hardware interfaces, CLI flags, and suggesting search keywords.
- Brainstorming and ideation: exploring variations on ideas, generating hints, outlining approaches rather than final answers.
- Real‑world troubleshooting: washing machines, cars, odd icons, reverse engineering small problems.
- Many frame LLMs as “smart coworker / intern / rubber duck” whose output they verify and edit.
Limitations and failure modes
- Hallucinations are a core problem: fabricated academic papers, wrong technical details (e.g., calling conventions, math notation), unsafe C code, bogus dependencies.
- Particularly bad at: niche research paper search, tasks requiring exact truth, or domains with sparse training data.
- Some worry early exposure in a new field may plant subtly wrong fundamentals.
- Others emphasize that, like any fallible tool or coworker, outputs must be tested, reviewed, and treated as non‑authoritative.
Ethical, environmental, and social concerns
- Strong concern about:
- Training on unlicensed data and “polluting the commons” with AI‑generated sludge.
- Climate impact and dubious CO₂ accounting that compares “being a human” vs. running a model.
- Corporate ownership and “intelligence as a service” non‑competes.
- Some find the tech so ethically tainted or “icky” that they refuse to use it despite utility.
Impact on work and jobs
- Many professionals say LLMs let them avoid tedious RTFM work and tackle more ambitious or enjoyable problems.
- Others fear widespread automation of “80% bullshit tasks” will justify large layoffs, concentrating gains with employers and model vendors.
- A contrasting view is that productivity gains will expand demand for software, possibly increasing the need for skilled engineers.
Meta: how to prompt and integrate
- Experience and domain knowledge are seen as crucial to getting value and catching errors.
- Some rely on chat interfaces; others use IDE integrations, CLI tools, or local/alternative frontends.
- There’s debate over whether LLMs are overhyped or simply being “held wrong” for inappropriate tasks.