Super Data Science: ML & AI Podcast with Jon Krohn 970: The “100x Engineer”: How to Be One, But Should You?
47 snips
Feb 27, 2026 A fast look at how code-generation tools are creating super-productive engineers who program through AI agents. Stories about Karpathy moving to agentic, English-first coding and Steinberger’s viral multi-agent workflow that outpaced teams. Discussion of why more planning and specs matter when agents do the typing and the risks of skill atrophy and overreliance on AI.
AI Snips
Chapters
Transcript
Episode notes
Agent Errors Are Conceptual Not Syntactic
- Code-gen mistakes have shifted from syntax errors to subtle conceptual errors and wrong assumptions.
- Karpathy observed agents now avoid simple syntax mistakes but often overcomplicate or assume incorrectly, requiring human judgment to simplify.
Parallel Agent Workflow That Generated 6,500 Commits
- Peter Steinberger achieved massive output by running multiple AI coding agents in parallel.
- He ran 3–8 agents in a 3x3 terminal grid, each working 5–15 minutes while he rotated attention and committed directly to main branch.
Voice First Spec Process That Enables Agentic Build
- Steinberger builds very detailed specs using voice-first dictation and adversarial AI review before coding.
- He dictates stream-of-consciousness, has AI structure it, then asks the AI to find 20 underspecified points until the spec grows ~500+ lines.