Ask HN: Daily practices for building AI/ML skills?
Aspiring AI and machine learning practitioners are weighing how best to build skills with limited time, from one-hour-a-day routines to more immersive options like graduate study or switching jobs. Many advocate a project-driven, top-down approach using modern tools (fast.ai, PyTorch, Hugging Face, LLM APIs) while gradually backfilling math and statistics fundamentals, rather than starting purely from theory. Others stress that deep understanding, paper implementation, and knowledge of broader ML methods remain crucial—especially for research roles—while experienced engineers point to growing opportunities in MLOps and in simply becoming highly effective at using existing large models.
Overall approaches to learning AI/ML
- Many recommend “learning by doing”: pick a passion project (e.g., document understanding, audio, sports analytics) and iterate on it.
- Immersion is emphasized: ideally make AI/ML part of your full-time job or studies; otherwise, consistent side projects.
- Some suggest 4-hour focused weekend blocks may be more productive than 1 hour per day.
Personal projects vs formal education
- Formal degrees are seen as something you could replicate via self-study, but school provides structure, deadlines, and social legitimacy.
- The diploma is considered practically useful for getting interviews, even if the content is learnable independently.
- Others argue you can self‑study via textbooks, online courses, and paper implementations if you have enough discipline.
Foundations vs “just use the libraries”
- One camp: start with statistics, classic ML (regression, naive Bayes, SVM, boosting), and math; understand models from the ground up.
- Opposing view: bottom‑up is too slow for a fast-moving field; better to begin with high‑level tools and learn fundamentals as needed.
- Several note that fundamentals (especially stats) are broadly valuable, not just for ML.
Implementing models and papers
- Strong support for implementing simple models and research papers to gain deep understanding.
- Disagreement on whether to hand‑code neural nets/backprop:
- Some say it’s like writing your own HashMap and not worth it beyond basic intuition.
- Others insist doing a simple implementation at least once is critical for understanding gradients and architectures.
- Papers With Code and curated “awesome papers” lists are recommended starting points.
Tools, frameworks, and hardware
- PyTorch (with new compilation features) and JAX are favored; TensorFlow is described as legacy/technical debt except in niches.
- Free cloud tools (e.g., hosted notebooks) can offset weak local hardware.
- JS‑based ML is possible but most examples and tooling are Python‑centric.
LLMs and application-focused skills
- Distinction between training LLMs vs using them in products; many advise focusing on usage (APIs, RAG, embeddings, tooling).
- Prompt engineering and extensive hands‑on use of multiple LLMs are seen as major skills on their own.
- Some warn that the gap between model developers and users is widening, like chip design vs general software.
Careers, MLOps, and ecosystem skills
- Strong demand is reported for engineers who can productionize ML: data pipelines, testing, deployment, monitoring, and scaling.
- Kaggle, blogging, and open‑source contributions are suggested for building a portfolio.
- Several stress that ML in practice is largely about data quality, infrastructure, and simple baselines, not just fancy models.