Recording intelligenceAI-generated brief · check the source for context

Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan

29:49 recording · AUTO · 2 speakers

Watch the original

Executive Summary AI
  • Andre says December was the clear break point: on a break with more time, he found the latest models produced whole chunks of code he never had to correct, and he ended up trusting the system and vibe coding full time.
  • He frames this as software 3.0 — after 1.0's explicit code and 2.0's trained weights, programming now means prompting, with the context window as the lever over the LLM interpreter, as in OpenClaw's install being text you copy-paste to your agent instead of a shell script.
  • His MenuGen app, which OCR'd a restaurant menu and generated pictures on Vercel, was made spurious when Gemini's nano banana rendered the dishes straight onto a photo of the menu, proving whole apps in between are no longer needed.
  • On verifiability, he argues LLMs automate what can be verified, which makes them jagged: state-of-the-art models refactor huge codebases or find zero-days yet tell you to walk to a car wash 50 meters away, and the chess jump from GPT-3.5 to GPT-4 came from data someone at OpenAI chose to add.
  • He separates vibe coding, which raises the floor so anyone can build, from agentic engineering, which preserves the professional quality bar, and says the strongest practitioners now gain far more than the old 10x, while hiring has not been refactored to test for it.

Brief overview

Coding has crossed into a software 3.0 paradigm where agents do the work and humans keep the taste.

  1. December was the real agentic turning pointThe chunks came out fine, he stopped correcting them, and trust in the system took over.
  2. Verifiable domains get the RL, so capability is jaggedModels peak at math and code yet fail the 50-meter car wash question.
  3. Humans still own spec, taste and oversightAgents matched Stripe and Google emails instead of using persistent user IDs.

Questions this recording answers

10 questions, each answered where it is said
What changed in December that made Karpathy feel behind as a programmer?

He had used agentic coding tools for about a year, editing the chunks when they were wrong. On a December break the latest models' chunks just came out fine; he kept asking for more, could not remember the last time he corrected one, trusted the system, and was vibe coding.

What are software 1.0, software 2.0 and software 3.0?

Software 1.0 is writing code. Software 2.0 is programming by arranging data sets, objectives and neural network architectures to train networks. Software 3.0 is prompting: trained implicitly on a large enough set of tasks from the internet, an LLM becomes a programmable computer, and the context window is the lever over that interpreter.

What is MenuGen, and why does Karpathy say the app shouldn't exist?

MenuGen lets you upload a photo of a restaurant menu; it OCRs the titles, uses an image generator for pictures of the dishes and re-renders the menu, running on Vercel. Then he saw the software 3.0 version: give the photo to Gemini and ask nano banana to render the dishes into the menu pixels.

Why are frontier models so jagged in what they can do?

Labs train these LLMs in giant reinforcement learning environments with verification rewards, so capability peaks in verifiable domains like math and code and stagnates elsewhere. Beyond verifiability it depends on what the labs happen to put into the data distribution, because some capabilities are far more valuable in the economy.

What example shows state-of-the-art models still getting simple things wrong?

Asked whether to drive or walk to a car wash 50 meters away to wash his car, state-of-the-art models say walk because it's close. The same Opus 4.7 will refactor a hundred-thousand-line codebase or find zero-day vulnerabilities. The older example was counting the letters in strawberry, which models now patch.

What is the difference between vibe coding and agentic engineering?

Vibe coding raises the floor: everyone can build software. Agentic engineering preserves the quality bar of professional software — no vulnerabilities introduced, you stay responsible, but you go faster. It is coordinating agents that are spiky, fallible and stochastic yet extremely powerful, and its ceiling is well above the old 10x engineer.

How should hiring change to find strong agentic engineers?

Most people have not refactored hiring; handing out puzzles is the old paradigm. Instead give a really big project — write a Twitter clone for agents, make it secure, have agents simulate activity — then attack the deployed site with ten Codex 5.4x-high instances that should not be able to break it.

What human skills become more valuable as agents do more?

Aesthetics, judgment, taste and oversight. Agents are intern-like entities with good recall for API details he no longer remembers — keepdims versus keepdim, dim or axis, reshape, permute, transpose — but they make spec mistakes, like tying MenuGen funds to matching Stripe and Google emails instead of a persistent user ID.

How far could computing be extrapolated from where it is now?

He imagines completely neural computers: a device that takes raw video or audio into a neural net which uses diffusion to render a UI unique to that moment. The neural net becomes the host process and CPUs the coprocessor, with tool use a historical appendage for deterministic tasks.

What is still worth learning deeply when intelligence gets cheap?

Understanding. A tweet he keeps thinking about says you can outsource your thinking, but you can't outsource your understanding. He feels like a bottleneck on knowing what is worth building and how to direct agents, and builds an LLM wiki from the articles he reads to process information and gain insight.

Key Quote

“And then I just trusted the system more and more. And then I was vibe coding.”

— Andre, Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan, at 1:37▶ watch at 1:37
Key Quote

“It's just like, what is the piece of text to copy paste to your agent? That's the programming paradigm now.”

— Andre, Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan, at 4:46▶ watch at 4:46
Key Quote

“this blew my mind because actually all of my menu gen is spurious it's working in the old paradigm that app shouldn't exist”

— Andre, Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan, at 5:58▶ watch at 5:58
Key Quote

“And state-of-the-art models today will tell you to walk because it's so close.”

— Andre, Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan, at 11:36▶ watch at 11:36
Key Quote

“Yeah, so I would say vibe coding is about raising the floor for everyone in terms of what they can do in software”

— Andre, Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan, at 15:57▶ watch at 15:57
Key Quote

“But then I would say agentic engineering is about preserving the quality bar of what existed before in professional software.”

— Andre, Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan, at 16:07▶ watch at 16:07

Cite this

“Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan.” https://www.youtube.com/watch?v=96jN2OCOfLs. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/andrej-karpathy-from-vibe-coding-to-agentic-engineering-w-st.html.

Report this page
What is wrong with this page?