Stenobird

Episode

D2DO291: From Politics to Machine Learning and AI Engineering

Podcast
Day Two DevOps
Published
Jan 7, 2026
Duration seconds
2503
Processing state
processed
Canonical source
https://packetpushers.net/podcasts/day-two-devops/d2do291-from-politics-to-machine-learning-and-ai-engineering/
Audio
https://feeds.packetpushers.net/link/20975/17248695/D2DO291.mp3
JSON
/v1/public/podcasts/day-two-devops/episodes/d2do291-from-politics-to-machine-learning-and-ai-engineering
Markdown
/podcast/day-two-devops/d2do291-from-politics-to-machine-learning-and-ai-engineering.md

Actions

  • POST https://stenobird.com/v1/public/podcasts/day-two-devops/episodes/d2do291-from-politics-to-machine-learning-and-ai-engineering/transcription-requests
    Idempotently request low-priority transcript generation for this episode.
  • GET https://stenobird.com/podcast/day-two-devops/d2do291-from-politics-to-machine-learning-and-ai-engineering.md
    Read the agent-friendly Markdown representation of this episode resource.

Summary

Learn the fundamental distinction between Machine Learning Engineering and AI Engineering through the lens of a Senior Applied Scientist at Twitch. This episode explores how to transition from using pre-trained APIs to building production-ready applications and managing data pipelines.

Topics

  • AI Engineering
  • Machine Learning
  • Data Pipelines
  • Airflow
  • LLM Evaluation
  • Data Science Careers
  • AWS SageMaker
  • Applied Science

Highlights

  • Main idea: AI Engineering focuses on building products using pre-trained models and APIs, whereas ML Engineering involves training models from scratch
  • Practical takeaway: Use Jupyter Notebooks for early-stage prototyping and data exploration, but transition to Python scripts and orchestrators for production
  • Failure mode: Relying solely on certifications rather than building and deploying production-ready, user-facing applications
  • Technical insight: Data orchestration tools like Airflow are essential for managing complex DAGs and triggering training jobs in large-scale environments
  • Practical takeaway: Evaluate LLM outputs using rubrics, specific rules, or 'LLM-as-a-judge' patterns to handle subjective content

Chapters

  1. 1:00 Defining the Role of an Applied Scientist: An introduction to Marina Wyss's role at Twitch and the scope of her work across ML infrastructure and recommendation systems.
  2. 4:10 AI Engineering vs. Machine Learning Engineering: A breakdown of the distinction between using pre-trained models via APIs versus training custom models from scratch.
  3. 10:25 Career Paths and Skill Acquisition: Discussing the non-linear paths into data science and the importance of foundational statistics and domain knowledge.
  4. 16:40 Demystifying Large Language Models: Understanding LLMs as probabilistic next-token predictors and how to approach model selection.
  5. 19:45 Data Orchestration with Airflow: An overview of using Airflow for managing data pipelines, DAGs, and integrating with services like AWS SageMaker.
  6. 25:50 Advice for Aspiring AI Engineers: Practical recommendations for learning, including essential reading and the value of self-directed projects.
  7. 29:00 Evaluating Autonomous AI Agents: Strategies for evaluating the performance of LLMs in subjective tasks like email drafting.