First choose one learning path based on your goal
This page is a cross-author learning navigation. If you only want to answer “What is Karpathy’s LLM OS?”, first read LLM OS chapter in the Software 3.0 talk notes; this page is for reading software paradigms, hands-on programming, and knowledge workflows separately, and it no longer repeats or expands the same talk.
- Understand software paradigms and product design: Start from the talk notes, then return to the original role-based reading path and choose chapters relevant to your own goals.
- Build an AI gadget hands-on: Compare it with Karpathy’s own MenuGen retrospective, noting the integration and deployment that still need to be handled from prototype to real application. The Chinese-community practices on this page are extensions by their respective authors and are not uniformly rewritten as Karpathy’s own words.
- Understand model training and inference implementation: Start from the official nanochat repository to check the current README, then read this page’s historical learning clues. This time, only the resource entry points were verified; training was not run, and no hardware costs, training times, or reproduction benefits are reported.
- Organize your own knowledge materials: See the knowledge-base paradigm chapter. This is a cross-author workflow compilation; you need to verify sources, links, and update methods yourself, and it is not equivalent to a model-training tutorial or a ready-made operating system.
This guide was fact-checked on 2026-10-25. Two historical simplifications need to be distinguished: the current nanochat repository covers stages such as model training, evaluation, and inference, so the old draft’s “200 lines of PyTorch” should not be treated as a current commitment to the entire project’s scale; the llms.txt proposal is content navigation and description, and “Access Control” in the old table should not be understood as identity authentication or access authorization.
Figures and year-based periods in the original body text, such as “several times the value” and “what percentage of demos cannot be deployed,” were judgments and predictions in a historical compilation and were not verified as universal conclusions this time. When reading, distinguish among the original talk, Karpathy’s personal practice, and secondary inferences by other authors.
Original cross-author compilation and citations
When to open: You want to explain to someone Software 3.0 / the etymology of Vibe Coding / the Agentic Decade / the LLM analogy framework. This wiki consolidates the content by Karpathy cited in the Chinese community by the three camps of newtype / Wang Kai / Xiaohongshu vibecoding.
One-sentence core:Software 1.0 → 2.0 → 3.0 three-stage framework + LLM = OS/Utility/Fab + Agentic Decade (2025–2035) + Vibe Coding as natural-language programming + rewriting internet infrastructure for agents (robots.txt → llms.txt).
0. Four perspectives on “Karpathy”
| Perspective | Source | Focus |
|---|---|---|
| newtype Lao Huang (primary source) | Karpathy · Software in the era of AI(Software 3.0) | Complete Chinese compilation of the YC 2025-06 talk / Software 3.0 / Agentic Decade |
| newtype knowledge-system camp | newtype knowledge systems and the full evolution of Life OS | Karpathy LLM Wiki as a knowledge-base architecture paradigm |
| Wang Kai hands-on camp | Wang Kai’s Claude Code terminal workflow methodology (complete version) | “Just Do It” + nanochat learning |
| Xiaohongshu vibecoding | Xiaohongshu / Vibecoding aesthetics + creator-economy collection Xiaohongshu / AI tools + Skill tutorial collection | Reproducing an LM in 243 lines of code + etymology of Vibe Coding + zero-basics explanation |
1. Cross-author consensus (5 points)
- The second major change in software in 70 years, and two such changes arriving in succession within 5 years —— 1.0 → 2.0 → 3.0
- LLMs are new infrastructure rather than applications — the OS analogy fits best
- Vibe Coding is not a toy; it is the everyday form of Software 3.0 — newtype / Wang Kai / Xiaohongshu agree
- The Agentic Decade (2025–2035) is a 10-year-scale evolution; do not become addicted to making Agent demos — semi-autonomous > fully autonomous
- Rewriting internet infrastructure for agents (robots.txt → llms.txt) — protocol-camp consensus
2. Explicit disagreements (2 points — not smoothed over)
| Point of disagreement | Karpathy’s stated position | Chinese-community extension position |
|---|---|---|
| The seriousness of Vibe Coding | “fully give in to the vibes” / semi-joking | Wang Kai / newtype treat Vibe Coding as a serious working mode; Xiaohongshu treats it as aesthetics + democratization of creation |
| The value of Agent demos | “Don’t become addicted to making Agent demos” / warning | newtype Profile / Wang Kai multi-agent matrix directly do Agent orchestration (without waiting for Karpathy’s approval) |
3. Software 1.0 → 2.0 → 3.0 three-stage framework
Karpathy · Software in the era of AI(Software 3.0) §I:
3.1 Comparison of three types of software
| Paradigm | How programs are written | Author | GitHub analogy | Current status |
|---|---|---|---|---|
| Software 1.0 | Writing code (C / Python) | Programmers | github.com | 70-year traditional paradigm |
| Software 2.0 | Tuning neural-network weights | ML engineers | Hugging Face | Dominant in 2017–2022 |
| Software 3.0 | Writing natural-language prompts | Anyone | (No unified repository; Prompts as code) | Dominant from 2023 onward |
Key point:The three coexist; skillfully switch among the three + know when to use which.
3.2 Karpathy’s firsthand experience
- While working on autonomous driving at Tesla, he gradually replaced 1.0 code (lane lines / traffic lights / if-else) with 2.0 neural networks
- Final result: a huge 2.0 neural network does image analysis, and 1.0 only remains as stitching
- Now in 3.0: even the stitching is replaced by natural-language prompts
“When I saw Sun Microsystems servers being wheeled into a Tesla building, I realized for the first time that the 2.0 stack was replacing the 1.0 stack.”
3.3 Implications for entrepreneurs
- No need to “know how to write code” → learn to express requirements precisely in natural language + tune prompts
- But 1.0 is not dead → performance bottlenecks / determinism / high safety requirements = use 1.0; ambiguous requirements = use 3.0; large amounts of data that cannot be explicitly modeled = use 2.0
4. Three LLM analogies
Karpathy · Software in the era of AI(Software 3.0) §II:
4.1 LLM = new operating system (OS)
1960s mainframe-era analogy:
| 1960s mainframe era | Today’s LLM era |
|---|---|
| Unix / Multics | GPT-5 / Claude 4.5 / Gemini / Grok |
| Centralized mainframes | Remote APIs / data-center inference |
| Time-shared access via teletype machines | ChatGPT / API access |
| The personal computer (late 1970s) has not yet arrived | The “personal LLM revolution” will arrive within a few years |
| Closed-source IBM / open-source Linux | Closed-source OpenAI / open-source Llama, Mistral |
Conclusion: Today is the intermediate transitional stage of “minicomputers / terminals / remote,” neither the punched-card era nor yet the personal-computer era.
4.2 LLM = public utility
- OpenAI / Anthropic / Google / xAI are like four power plants
- APIs are power-grid dispatch, charged by Token
- Open-source options such as DeepSeek and Llama = “self-generation”
- When one has a “power outage” (API failure) = “intelligence brownout,” GDP falls
4.3 LLM = semi-formed fabrication plant (Fab)
- Capital-intensive + technology-stack analogy to TSMC / Intel
- The “hardware” is software (weights), so it is easier to copy
- Inference = shipping products, training = building the plant
- Marginal cost is almost 0, but upfront capital + intellectual barriers are enormous
Taken together, the three analogies mean: an LLM is simultaneously an OS, a Utility, and a Fab. The OS analogy fits best.
5. LLM psychology: human-like traits + cognitive defects
Karpathy · Software in the era of AI(Software 3.0) §III:
5.1 Human-like traits vs. defects
| Human-like traits | Defects |
|---|---|
| Natural-language conversation | Hallucination |
| Reasoning | Poor arithmetic / long-context consistency |
| Programming | Does not know what it does not know |
| Creativity | Goes along with what people want rather than with facts (sycophancy) |
| Multitasking | The longer the context, the easier it is to distort |
5.2 Recommended working mode
- Do not treat an LLM as an omniscient assistant → treat it as an “easily distracted genius intern”
- Semi-autonomous > fully autonomous → human in the loop + proofreading
- Shorten feedback loops → let the LLM output at a granularity where you can judge right or wrong within 100 ms
6. Agentic Decade(2025-2035)
Karpathy · Software in the era of AI(Software 3.0) §V:
6.1 Karpathy’s judgment
- Not the Agent Year, but the Agent Decade — don't expect it to be ready in 2025
- Semi-autonomous products are the ones that make money —— Cursor / Replit / Perplexity / GitHub Copilot
- Don't get addicted to making Agent Demos — 99% of demos will never be deployed in practice
6.2 Four principles for building semi-autonomous products
- Tight feedback loop — LLM output can be quickly verified by humans
- Predictability — Output format is predictable (JSON / Code Patch)
- Recoverability — If wrong, it can be rolled back / undone
- Generation/Verification speed matching — However fast the LLM generates, humans must verify just as fast
6.3 Extended application to newtype
newtype mindset · Super-individual methodology:
- newtype Profile = semi-autonomous multi-Agent orchestration
- Not "AI completes a blog post in one click" (fully autonomous)
- But "AI gives a draft → human proofreads → publish" (semi-autonomous)
7. Etymology and interpretation of Vibe Coding
7.1 Karpathy's original post
"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." — Karpathy, February 2025 X post
Karpathy · Software in the era of AI(Software 3.0) §VI:Vibe Coding means "software has shifted to natural-language programming" — It is not a toy; it is the everyday form of Software 3.0.
7.2 Extensions in the Chinese-language community
Xiaohongshu vibecoding school(Xiaohongshu / Vibecoding aesthetics + creator economy collection):
- Karpathy reproduces an LM in 243 lines of code — That's the entire algorithm; the hundreds of thousands of extra lines in GPT are only to make it run faster
- Zero-basis explanation + prompts — Enabling ordinary users to follow Karpathy's video all the way through
- Metaphor:Students and a teacher grading fill-in-the-blanks(explaining backpropagation)
Wang Kai + newtype school:
- Vibe Coding = a serious working mode, not "going by feeling"
- Used with Claude Code / Cursor / Spec → engineering
Vibe Coding and independent development (cross-author thematic synthesis) See for details.
8. Rewriting infrastructure for Agents
Karpathy · Software in the era of AI(Software 3.0) §7:
8.1 Core claim
"robots.txt → llms.txt; HTML → Markdown / API; search → Agent invocation"
The Agentic Decade requires rewriting the entire software infrastructure for Agents, rather than making Agents struggle to parse HTML designed for humans.
8.2 Practical extensions
| Old infrastructure | New infrastructure (for Agents) |
|---|---|
| robots.txt | llms.txt(Access Control + Inference Guidance) |
| HTML | Markdown / JSON / Open API |
| Google Search | MCP Server / Tool Call |
| OAuth Web Flow | Programmatic Auth Token |
| Mouse-click UI | Computer Use / Browser Tool |
| Captcha | Agent Identity / Whitelist |
SEO × LLM era (cross-author thematic synthesis) See the llms.txt proposal for details.
9. Karpathy LLM Wiki as a knowledge-base paradigm
The full evolution of the newtype knowledge system and Life OS §1:
Lao Huang exclusive:Karpathy LLM Wiki model = a paradigm reference for contemporary knowledge bases:
- Simple Markdown(no JS pollution)
- H1 + blockquote summary(LLM-friendly)
- Single source + cross-links(name wiki links)
- Can be read directly by an LLM Agent(no need to scrape HTML)
newtype wiki (wiki.newtype.pro) = one of the earliest wikis in China to engineer the Karpathy model.
10. Wang Kai / nanochat learning
Wang Kai's Claude Code terminal workflow methodology (complete version) + newtype mindset · Super-individual methodology:
10.1 Wang Kai quotes Karpathy's "Just Do It"
"Don't wait until you've finished learning before acting" — Karpathy post / one of Wang Kai's mindsets
10.2 nanochat as a learning path
- Karpathy 2025 nanochat = 200 lines of PyTorch / complete GPT training
- Path recommended by Wang Kai + Lao Huang:Study nanoGPT → nanochat → truly understand Transformer
- LLM fundamentals (Transformer / Attention / three types of PLM) (cross-author thematic synthesis) See for details
11. How to use it — a Karpathy path by identity
| Who you are | Primary method | Starting point |
|---|---|---|
| Want to understand the evolution of software in the AI era | Software 3.0 three-part framework | 1.0 / 2.0 / 3.0 coexist + when to use which |
| Want to make startup judgments | Three LLM analogies + Agentic Decade | LLM = OS; which layer should I work on? Four principles for semi-autonomous products |
| Want to do AI programming | Vibe Coding + nanochat | Karpathy’s tweet + 243-line reproduction + engineering implementation |
| Want to build an LLM-friendly site | robots.txt → llms.txt | llms.txt + Markdown structure |
| Want to build a knowledge base | Karpathy LLM Wiki paradigm | Clean Markdown + H1 + blockquote + wikilink |
| Getting started with the LLM kernel | nanochat path | 200 lines of PyTorch + Karpathy YouTube series |
General discipline:
- Not the Agent Year but the Agent Decade(Karpathy)
- Semi-autonomous > fully autonomous(4 principles)
- Don't get addicted to making Agent Demos
- Vibe Coding is not a toy but everyday practice
- Rewrite infrastructure for Agents(llms.txt / Markdown / API)
- Three types of software coexist + when to use which
12. Citation list
- Karpathy · Software in the era of AI(Software 3.0) — Complete Chinese compilation of the YC AI Startup School June 2025 talk / Software 3.0 / three LLM analogies / Agentic Decade / four semi-autonomous principles
- The full evolution of the newtype knowledge system and Life OS — Karpathy LLM Wiki paradigm / three-layer knowledge-base architecture
- newtype mindset · Super-individual methodology — Lao Huang's 15 mindsets / Just Do It
- Wang Kai's Claude Code terminal workflow methodology (complete version) — Wang Kai's practice / quoting Karpathy
- Xiaohongshu / Vibecoding aesthetics + creator economy collection — Beginner-friendly explanation / 243-line reproduction / student-and-teacher metaphor
- Xiaohongshu / AI tools + Skill tutorial collection — Karpathy nanochat / Skill system
Sources and related materials
- Karpathy · Software in the era of AI(Software 3.0)
- The full evolution of the newtype knowledge system and Life OS
- newtype mindset · Super-individual methodology
- Wang Kai's Claude Code terminal workflow methodology (complete version)
- Xiaohongshu / Vibecoding aesthetics + creator economy collection
- Xiaohongshu / AI tools + Skill tutorial collection
- Claude Code workflow (cross-author thematic synthesis)
- AI Agent architecture and multi-Agent orchestration (cross-author thematic synthesis)
- SEO × LLM era (cross-author thematic synthesis)
- Vibe Coding and independent development (cross-author thematic synthesis)