"Forget about that 100yr time horizon, I just want to build." — Greg
Why koding vs. math: in math you write a proof and maybe three people care; in coding you write a program and everyone benefits.
Speed via First Principles
Identify which steps are truly required vs. legacy process.
Collapse feedback loops (e.g. live-iterate on a call vs wait 9 months).
Focus on the few hard constraints; bulldoze through irrelevant ones.
Speed compounds because it unlocks more cycles of learning.
Independent Study that Works (lessons, backed by his examples)
Go deep; push through the “boredom walls.”
Learn by building end-to-end and shipping (Kaggle comps, GPU rig)
Create your own learning environment & tools.
If keep getting introed to same smart people, it's real. Use that as a signal to double down.
Turing (1950): you’ll never write all the rules—build a child machine that learns via rewards & punishments.
Mix “talk to people” with “do the work.” Conversations (e.g. Geoff Hinton) inform direction; mastery comes from doing.
Do you still believe great engineers can contribute as much as researchers?
“Absolutely—if not even more true today.”
Early example = fast conv kernels (Alex) + apply to ImageNet (Ilya) → AlexNet.
Today: eng for ~100k GPUs, complex RL orchestration.
If you don’t have the idea, there’s nothing to do; without engineering, the idea doesn’t live. You need both.
Tells every new OpenAI engineer to have technical humility — leave traditional intuitions at the door. Assume you’re missing context; deeply understand why; then change the abstraction.
Working pattern that scaled: bring 5 ideas; teammate says 4 are bad → “great—that’s all I wanted.”
Coding with Models
“How you structure your codebase determines how much you get out of Codex.”
Structure repos for models: small, well-documented modules, crisp entry points, and fast tests so models can fill in details and run tests repeatedly.
Design for models, not just humans: humans can hold big abstractions and skip tests; models can’t — so make it easy for them to succeed.
Pattern is likely durable as models improve and also great for human maintainability.
Vibe coding = empowerment and an early glimpse; next step is agents as coworkers — off-laptop, cloud-resident, tool-using, still working while you sleep.
Biggest ROI: transforming existing applications (migrations, library upgrades, legacy refactors), not just flashy greenfield demos.