Core Thesis
Andrej Karpathy's Sequoia AI Ascent 2026 keynote reframes software development around LLMs, framing it as Software 3.0: programming via prompts, context, tool calls, memory, and natural-language specs rather than typed code.
The Three Paradigms
- Software 1.0: Human-written rules (if/else, loops, functions) — the dominant paradigm for 70 years.
- Software 2.0: Neural networks trained on data, with humans as data curators and architecture designers (Karpathy's 2017 framing).
- Software 3.0: Humans write intent through prompts, context, tools, examples, memory, and instructions. The context window becomes the program.
- Writing a function, calling an API, handling an exception
- "Implement this feature," "refactor this subsystem," "research this library," "write tests, run them, fix failures," "compare approaches and recommend one"
- Vibe Coding raises the floor — designers, PMs, and students can ship prototypes. It does not guarantee security, maintainability, or taste.
- Agentic Engineering raises the ceiling — designing full systems where agents propose, code, test, and review while humans write specs, supervise plans, review diffs, manage permissions, and enforce quality.
- Code: compile, test, observe failures — immediate feedback
- Math proofs: step-by-step verification
- Creative writing: fuzzy boundaries, slow progress
- Medical diagnosis: long ethical/legal verification cycles
- When intelligence becomes cheap, understanding becomes scarce.
- Taste, judgment, oversight, and understanding remain uniquely human.
- Software 3.0 is not coming — it is already here for anyone using agentic tools productively.
- Karpathy, "Software 2.0" (2017) — https://karpathy.medium.com/software-2-0-a64152b37c35
- AI Ascent 2026 full video — https://www.youtube.com/watch?v=96jN2OCOfLs
- Karpathy tweet thread — https://x.com/karpathy/status/2049903821095354523
- Stephanie Zhan summary — https://x.com/stephzhan/status/2049518659513852109
- Sequoia Inference analysis — https://inferencebysequoia.substack.com/p/andrej-karpathys-software-30-and
- Karpathy Bear Blog notes — https://karpathy.bearblog.dev/sequoia-ascent-2026/
- Anthropic, "Effective Harnesses for Long-Running Agents" — https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
> "The context window is the new program."
> "LLM 就是一个高级的英文解释器" — an LLM is, literally, a high-level English interpreter.
Traditional code is precise but fragile; Software 3.0 is fuzzy but adaptive.
The December 2025 Inflection Point
Karpathy noted that around December 2025, agentic tools (Claude Code, Cursor, Codex) crossed a threshold: generated code became larger, more coherent, and more reliable. He no longer recalls when he last corrected the AI.
> "我从未像现在这样感到落后" — "I have never felt this behind."
The shift is not that programming became harder; the default workflow changed.
From Code Lines to Macro Actions
The unit of programming shifted from:
to delegating macro actions:
Karpathy describes "AI psychosis": spending 16 hours a day speaking to agents, starting a new task the moment one finishes.
Vibe Coding vs Agentic Engineering
MenuGen Bug Case Study
An agent matched Stripe buyer emails to Google login emails — logically plausible but architecturally wrong. Persistent user IDs were the correct approach. The agent remembered API details; the human needed to understand identity, security boundaries, and system shape.
LLMs as Ghosts, Not Animals
LLMs are jagged, statistical, summoned entities — strong on some tasks, inexplicably weak on others, prone to hallucination and confident wrong answers. Working with them requires taste, judgment, and oversight.
> "These are spiky entities. They are fallible and stochastic, but extremely powerful."
Verifiability: The Adoption Framework
AI lands fastest where outcomes are highly verifiable:
The December 2025 coding breakthrough happened because code verification is instant, automated, and cheap — enabling rapid model improvement.
Advice to founders: seek undervalued domains where outputs are easy to verify.
App Collapse: The MenuGen Pattern
Some app categories collapse into a single model call. A menu app — constraints in, complete formatted output out — needs no backend, database, API, or deployment stack.
> "Some app categories are collapsing into a single model call."
LLMs are not assisting software development; they are redefining what software is.
Hiring Must Change
Traditional LeetCode interviews are increasingly mismatched. Karpathy proposes:
> "Build a substantial project with agents, deploy it, make it secure, and then have adversarial agents try to break it."
This tests decomposition, specification, quality control, and security hardening. The "10x engineer" concept may become far more extreme.
Understanding Cannot Be Outsourced
> "You can outsource your thinking, but never your understanding."
Frontier skill is not memorizing APIs but grasping underlying concepts: storage, views, memory copies, invariants, identity, security boundaries, system shape.
Compute Architecture Inverting
The relationship between CPU and GPU is inverting. Future operating systems may be AI-native: model-driven rather than instruction-driven. The OS itself becomes a massive agent managing resources and user intent.