Vanishing Gradients

Episode 59: Patterns and Anti-Patterns For Building with AI

25 snips
Sep 23, 2025
In this engaging discussion, John Berryman, Founder of Arcturus Labs and an early engineer on GitHub Copilot, dives into the real-world challenges of building AI applications. He highlights the 'seven deadly sins' of LLM development, offering practical solutions to keep projects moving. John explains why aspiring for perfect accuracy may hinder progress and shares insights on context management and retrieval debugging. Treating an LLM like a forgetful intern, he emphasizes starting simply and avoiding unnecessary complexity for successful deployment.
Ask episode
AI Snips
Chapters
Books
Transcript
Episode notes
ADVICE

Don't Demand 100% Accuracy

  • Reframe requirements that demand 100% accuracy and compare them to human expectations.
  • Make outputs transparent and let users review and sign off so the system saves time without promising perfection.
ADVICE

Keep Users In The Loop With Agents

  • Avoid giving agents large closed tasks to run unattended; keep the user in the loop during the process.
  • Ask clarifying follow-ups and present intermediate checks so the agent narrows ambiguous requirements.
INSIGHT

Use Context Responsibly

  • Longer context windows don't guarantee better results and can degrade accuracy, cost, and latency.
  • Tier potential context by relevance and trim to fit your token budget for better model performance.
Get the Snipd Podcast app to discover more snips from this episode
Get the app