Andrej Karpathy

Sources#

• Andrej Karpathy: From Vibe Coding to Agentic Engineering
• llm-wiki
• The State of Agent Wikis

Summary#

Co-founder of OpenAI, former Director of AI at Tesla (got Autopilot working), now founder of Eureka Labs (AI-native education). The field's most prolific namer: coined "vibe coding," the "Software 1.0 / 2.0 / 3.0" taxonomy, the "animals vs. ghosts" framing for model intelligence, and — in this May 2026 Sequoia AI Ascent interview — "agentic engineering" as the serious discipline succeeding vibe coding. He also originated the LLM-as-compiler knowledge-base pattern this very vault runs on (the llm-wiki gist), and re-endorses it here as his own daily practice.

The "never felt more behind" admission#

The interview's startling opener: Karpathy — of all people — says he's "never felt more behind as a programmer." He locates a sharp phase transition in December 2026: agentic tools went from "good at chunks of code, sometimes you'd correct them" to "the chunks just came out fine, and I can't remember the last time I corrected it." He pushed people on X to re-evaluate AI as of December — many had frozen their mental model at "ChatGPT-adjacent," and the agentic-coherent workflow had since changed fundamentally. The admission is the emotional engine of Vibe Coding vs. Agentic Engineering: even the namer of the floor is scrambling to track the ceiling.

Frameworks he originated (each now its own page)#

• Software 3.0 — prompting/context as the program; the LLM as a programmable interpreter.
• Vibe Coding vs. Agentic Engineering — vibe coding raises the floor; agentic engineering preserves the quality bar.
• The Verifiability Thesis — LLMs automate what you can verify, the way classical computers automate what you can specify.
• Jagged Intelligence (Ghosts, Not Animals) — "we're not building animals, we're summoning ghosts"; spiky capability with no intrinsic motivation.
• Agent-Native Infrastructure — "why are docs still written for humans? what do I copy-paste to my agent?"
• Outsource Your Thinking, Not Your Understanding — the education thesis; understanding as the residual human bottleneck.

On his own knowledge-base practice#

He closes the interview by tying education back to LLM-as-Compiler Knowledge Base: "anytime I see a different projection onto information, I gain insight… whenever I read an article I have my wiki that's being built up from these articles, and I love asking questions about it." He frames it as synthetic data generation over fixed data — a tool to force information into his own head, because "you can't be a good director if you don't understand." This is the founder of the pattern using the pattern, and it is the direct lineage of this repository.

The gist became a category. Within roughly three months of its April 2026 publication, four industrial implementations shipped — Cognition's DeepWiki, Factory's AutoWiki, LangChain's OpenWiki, Garry Tan's GBrain — and mem0's July 2026 survey ("The State of Agent Wikis," practitioner-opinion) treats the gist as the founding document of the agent wiki, reading the four-way convergence on his three-layer sources/wiki/schema-file structure as evidence the structure is correct. Full landscape on LLM-as-Compiler Knowledge Base.

Key quotes#

• "You can outsource your thinking, but you can't outsource your understanding."
• "We're not building animals, we are summoning ghosts." (jagged, statistical, summoned entities)
• "Vibe coding is about raising the floor… agentic engineering is about preserving the quality bar of what existed before in professional software."
• "Why are people still telling me what to do? I don't want to do anything. What is the thing I should copy-paste to my agent?"
• On MenuGen: "all of my menu gen is spurious… that app shouldn't exist."

Connections#

• Software 3.0 — his taxonomy; the interview's conceptual spine
• Vibe Coding vs. Agentic Engineering — coined both terms a year apart
• The Verifiability Thesis — his attempt to explain why models are jagged
• Jagged Intelligence (Ghosts, Not Animals) — the "ghosts not animals" essay, applied
• Agent-Native Infrastructure — his "favorite pet peeve"
• Outsource Your Thinking, Not Your Understanding — his answer to "what's worth learning deeply"
• LLM-as-Compiler Knowledge Base — he originated the pattern (llm-wiki) and uses it daily
• Open-Ended Discovery Harnesses — autoresearch (2026) is his, and it defined the category: run an off-the-shelf coding agent in an experimental loop for hours or days, committing improvements and reverting regressions. It spawned systems-optimization, kernel-optimization and prompt-injection descendants, and it is the design every later harness in that class argues against — its two choices (one accumulating context, one editable program state) are named as the causes of idea collapse
• The Bitter Lesson — the Sutton principle his neural-net-as-host-process extrapolation rests on
• Claude Code — names "cloud code / codex / open claw" as the agentic-coding surfaces he lives in
• Boris Cherny — parallel "coding is solved" arc; both treat themselves as the case study
• OpenAI — the company he co-founded; origin of the Software 3.0 / vibe-coding lineage that recurs across this vault

Sources#

• Andrej Karpathy: From Vibe Coding to Agentic Engineering — Sequoia AI Ascent 2026 interview with Stephanie Zhan
• llm-wiki — his original gist defining the LLM-wiki pattern
• The State of Agent Wikis — mem0, In Context #17 (2026-07-21, practitioner-opinion): the gist's three-month industrialization into the agent-wiki category