By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Sequoia CapitalAndrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Why “Software 3.0” turns prompting into programming and engineering
- Andrej Karpathy describes a personal inflection point in December where agentic coding tools began producing reliably correct code, leading him to feel “behind as a programmer” while also enabling rapid side projects.
- He explains “Software 3.0” as a paradigm where prompting and context windows become the programming interface, enabling workflows that bypass traditional scripts and even entire apps.
- Using examples like OpenClaw installation and his Menu Gen project, he argues many existing software patterns are artifacts of an older mindset and will be replaced by more direct model-driven transformations.
- He frames current model performance as “jagged,” often strongest in verifiable domains because reinforcement learning rewards what can be checked, while weaker in out-of-distribution everyday reasoning.
- He distinguishes vibe coding from “agentic engineering,” emphasizing that professionals must coordinate stochastic, intern-like agents to go faster without lowering security and quality standards, and he suggests hiring should test large, real deployments rather than puzzles.
IDEAS WORTH REMEMBERING
5 ideasThe recent jump in agentic coding felt like a sudden phase change, not a gradual improvement.
He describes a sharp shift around December when newer models began producing larger chunks of correct code consistently, reducing how often he needed to intervene and making him trust the system enough to “vibe code” continuously.
Software 3.0 shifts the lever of control from code to context windows and prompts.
In his framing, Software 1.0 is hand-written code, Software 2.0 is “programming” via datasets/training, and Software 3.0 is programming by shaping context/prompt for an LLM that interprets it to perform work in an information space.
Agents can replace complex cross-platform glue code by using “copy-paste to your agent” as the new interface.
He points to installation instructions that are now best expressed as text to give an agent, because an agent can adapt to the user’s specific environment, debug in-loop, and avoid brittle cross-platform scripts.
Some apps built in the old paradigm may be rendered spurious by direct model-to-output workflows.
He built a full app to transform a photographed menu into items and generated images, then saw a “raw” approach where giving the photo to a model and prompting it to overlay results made the app feel unnecessary in hindsight.
Verifiability helps explain why models can be brilliant at code yet oddly poor at simple real-world judgment.
He argues LLMs automate what you can verify, because training uses reinforcement learning with verification rewards, producing “jagged” capabilities that peak in verifiable domains (like math/code) while failing strangely elsewhere.
WORDS WORTH SAVING
5 quotesAnd then I can't remember the last time I corrected it.
— Andrej Karpathy
Actually, all of my Menu Gen is spurious. It's working in the old paradigm. That app shouldn't exist.
— Andrej Karpathy
How is it possible that state-of-the-art Opus 4.7 will simultaneously refactor a hundred thousand like code base, uh, line code base or find zero-day vulnerabilities, and yet tells me to walk to this car wash?
— Andrej Karpathy
This is insane.
— Andrej Karpathy
So I would say vibe coding is about raising the floor for everyone in terms of what they can do in software.
— Andrej Karpathy
SPEAKERS
Stephanie Zhan
hostPartner at Sequoia Capital and host/interviewer for AI Ascent.
Andrej Karpathy
guestAI researcher and engineer known for work on deep learning and large language models.
High quality AI-generated summary created from speaker-labeled transcript.
