Bbarath.ai
New · Theo Browne

Notes from people worth listening to.

I watch the talks and interviews that change how I build, then distill them — key takeaways, lessons learned, and the tricks worth stealing. Skip the hour-long video; keep the ideas.

sessions distilled
3
sessions distilled
speakers
3
speakers
key takeaways
37
key takeaways
minutes of notes
39
minutes of notes
  • Define "done", then let it run — Theo Browne
  • NVIDIA's founding technology was flat-out wrong — Jensen Huang
  • Opus 5 runs for days, weeks, even months without scaffolding — Boris Cherny
  • Every good prompt has three parts — Theo Browne
  • The fix was three textbooks — Jensen Huang
  • Prompt injection is close to a solved problem in their stack — Boris Cherny
  • Pre-approve the scary-but-necessary steps — Theo Browne
  • The idea that was right: accelerate algorithm domains, not chips — Jensen Huang
  • Claude Code deleted over 80% of its system prompt for Opus 5 — Boris Cherny
  • Say what you don't want — Theo Browne
  • Great companies come from a unique perspective you deeply believe in — Jensen Huang
  • Prompts are rebuilt by ablation, not intuition — Boris Cherny
  • Give it permission to ask — Theo Browne
  • The Sega story is a masterclass in honest failure — Jensen Huang
  • Evals outlive harnesses — but not by much — Boris Cherny
  • Delete "think carefully" lines — Theo Browne
  • AlexNet wasn't AlexNet — it was a universal function approximator — Jensen Huang
  • "Product overhang" is the opportunity — Boris Cherny
  • Don't use Max reasoning — Theo Browne
  • Coarse recursive self-improvement already exists — Jensen Huang
  • Give the model harder tasks than you think it can handle — Boris Cherny
  • Steer mid-run instead of restarting — Theo Browne
  • Agents don't need to be 100% right to be useful — Jensen Huang
  • Verification is the single most-missed skill — Boris Cherny
  • Design needs negatives — Theo Browne
  • AI automates tasks, not jobs — Jensen Huang
  • Dynamic workflows are a new axis of test-time compute — Boris Cherny
  • Name the stops in CLAUDE.md — Theo Browne
  • Physical AI is next, and it's already a ~$10B business for NVIDIA — Jensen Huang
  • Codebase maintenance is being automated with routines — Boris Cherny
  • Ask for subagents explicitly — Theo Browne
  • Keep the task list in a file — Theo Browne
  • Run a review pass before a person does — Theo Browne
  • Ask it to mark what it couldn't confirm — Theo Browne
  • Attach the chart or screenshot itself — Theo Browne
  • Know your flag escape hatches — Theo Browne
  • Use fast mode (/fast) — Theo Browne

Latest learning

Theo Browne

Founder of T3 · YouTube (Theo – t3.gg)

Theo Browne: How to Get the Most Out of Opus 5.5

What I took away from Theo's walkthrough of Addy Osmani's Opus 5.5 guide, plus the details from the guide itself — define done, stop saying "think hard", never touch Max reasoning, name the stops, keep the task list in a file, and verify everything.

  1. 1Define "done", then let it run
  2. 2Every good prompt has three parts
  3. 3Pre-approve the scary-but-necessary steps
Read the notes24 min read

The library

Browse by speaker

Know a talk worth distilling?

Send me the interview, keynote or deep-dive that changed how you build — I'll watch it and write it up.

Suggest a talk