When to open: You need to judge "When will AGI arrive / Does Scaling Law have diminishing returns or not / Is Software 3.0 natural language programming the future / Should we believe AGI by 2027". This wiki integrates the perspectives of 4 schools (Karpathy / Aschenbrenner / Domestic three-school debate / Wang Kai + Embiricos) on the AGI topic.
One-sentence core:All schools agree that LLM = OS / Software 3.0 (natural language programming), but AGI timeline disagreements span a 3–15 year gap. Aschenbrenner bets on 2027, Kai-Fu Lee bets "Scaling Law is already slowing down", Zhu Xiaohu bets "I won't believe in AGI before controllable nuclear fusion", Karpathy bets on "Agentic Decade 2025–2035", Embiricos bets "AGI bottleneck = human typing speed rather than compute".
0. The 4 schools' "AGI" positions
| Perspective | Source | Position | Timeline forecast |
|---|---|---|---|
| Brute-force aesthetics school | Wang Kai Collection / Close Reading of Cutting-Edge AI Papers and Reports 2024–2026 §1 Aschenbrenner | "AGI by 2027 is strikingly plausible" | 2027 |
| Software 3.0 school | Karpathy · Software in the era of AI(Software 3.0)(Karpathy 2025-06) | "Agentic Decade slow 10-year evolution" | Mature in 2032–2035 |
| Domestic three-school debate | newtype AI Business, Industry and Investment Observations §3 | Zhu Xiaohu / Cohere Aidan / Kai-Fu Lee | 2030+ / Wait for controllable nuclear fusion |
| Typing speed bottleneck school | Lenny's Newsletter / 2026 Three-Person Talk on Organizational Operations in the AI Era (Anthropic + OpenAI) Embiricos | "AGI bottleneck = human typing speed, not GPU compute" | A few years later |
1. Cross-author consensus (5 points)
1.1 LLM = OS (Karpathy + Jensen Huang consensus)
Karpathy · Software in the era of AI(Software 3.0) §2.1:LLM = new type of operating system:
| 1960s mainframe era | Today's LLM era |
|---|---|
| OSes such as Unix / Multics | GPT-5 / Claude 4.5 / Gemini / Grok |
| Centralized mainframes | Remote API / data center inference |
| Time-shared access via telegraph machines | Access via ChatGPT web / API |
| Personal computers (late 1970s) haven't arrived yet | "Personal LLM revolution" will arrive within a few years |
| Closed-source IBM / open-source Linux | Closed-source OpenAI / open-source Llama, Mistral |
Jensen Huang's May 2024 post 227 corroborates:
- Memory management = context window
- File system = conversation history + knowledge base
- Drivers = Function Call
- User interface = natural language interaction (replacing CLI / GUI)
1.2 Software 1.0 / 2.0 / 3.0 triad
Karpathy · Software in the era of AI(Software 3.0) §1 Core:
| Paradigm | Way of writing programs | Writer | 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 (writing datasets → optimizer produces weights) | ML engineers | Hugging Face (GitHub for weights) | Dominant 2017–2022 |
| Software 3.0 | Writing natural language prompts, LLM executes directly | Anyone | (No unified repository yet, Prompts as code) | Dominant from 2023 onward, full-stack penetration within 5 years |
3 coexisting: Today's software engineers must be proficient at switching among the three, knowing the strengths and weaknesses of each paradigm, and when to use which.
Karpathy's own experience: Tesla Autopilot gradually replaced 1.0 code (lane detection, traffic light recognition) with 2.0 neural networks;now in the 3.0 era, even stitching (gluing 2.0 outputs together) is starting to be replaced by natural language prompts.
1.3 LLM psychology: Human-like but with "cognitive defects"
Karpathy · Software in the era of AI(Software 3.0) §3:
| Human-like traits | Defects |
|---|---|
| Has encyclopedic knowledge, a "polymath" trained on massive tokens | Hallucination: Makes up stories and is confident when there is no ground truth |
| Has contextual understanding, reasoning, code-writing ability | No long-term memory: Forgets everything just said across sessions |
| Capable of multimodal conversation, role-playing | Arithmetic / counting / geometric spatial perception: Often wrong |
| Able to organize / summarize / translate | Sycophancy bias: Agrees with whatever the user says |
| Able to perform step-by-step reasoning | Completely unaware of the world after the knowledge cutoff |
Karpathy's advice to developers:
- Don't treat LLMs as "absolutely reliable programs": they are "knowledgeable but slightly mentally unstable interns"
- For things that 1.0 code can handle, such as ground truth, counting, geometry, strict logic,still use 1.0 to do it
- Prompts are not commands, they are "assigning tasks to an intern" — the more context provided, the more specific the requirements, the higher the completion quality
- Don't trust a single output.Use evals + automatic review mechanisms
1.4 Iron Man Suit 4 principles (semi-autonomous product consensus)
Karpathy · Software in the era of AI(Software 3.0) §4.3 Karpathy summarizes 4 principles for LLM Apps:
- Context management: Don't stuff all information into the prompt; let the user decide what to stuff, when to stuff it, and to what depth
- Multi-LLM orchestration: Use cheap models for first pass, expensive models for verification; use tool models for retrieval, reasoning models for planning
- GUI ≠ chat box: Don't make users do everything through a chat box;Build a GUI, make "autonomy" a visual slider(like layer opacity in Photoshop)
- Rapid validation loop: The UI must let users judge whether the LLM output is correct within 5 seconds;the longer it takes, the deader the product
Iron Man Suit vs Robot analogy:
| Dimension | Fully manual | Semi-autonomous (Iron Man Suit) | Full Agent (Robot) |
|---|---|---|---|
| Decision power | 100% user | User sets direction + LLM executes details | 100% LLM |
| Trust requirement | 0 | Medium | Extremely high |
| Currently feasible | Yes | Yes | No |
| Karpathy recommendation | ❌ Wastes LLM | ✅ Golden zone | ❌ Demo runs, productization dies |
Embiricos corroborates(Lenny's Newsletter / 2026 Three-Person Talk on Organizational Operations in the AI Era (Anthropic + OpenAI)):Mixed-Initiative(Mixed-initiative) — Tesla Autopilot is a good example of mixed-initiative software, empowering humans while providing assistance. Codex is also designed in the same direction.
1.5 Vibe Coding = Software has shifted to natural language programming
Karpathy · Software in the era of AI(Software 3.0) §1 + newtype · AI Coding Practice and Toolstack Evolution §9:
"You don't need to 'know how to write code' to start a business.Learn to express requirements precisely in natural language + tune prompts, and you can have LLMs write code and deliver software for you. "
But1.0 is not dead. Key judgment:Where to use which:
- Performance bottlenecks, determinism, high-safety parts → 1.0
- Fuzzy requirements, natural language interaction parts → 3.0
- Parts with abundant training data but no explicit modeling → 2.0
See AI Coding Toolstack Evolution (Cross-Author Thematic Synthesis) + Vibe Coding and Independent Development (Cross-Author Thematic Synthesis).
2. Explicit disagreements (3 points — not smoothed over)
2.1 Disagreement 1: AGI timeline (2027 vs 2030+ vs 2035)
| Perspective | Timeline forecast | Key argument |
|---|---|---|
| Aschenbrenner(June 2024 / 165 pages) | AGI by 2027 strikingly plausible | GPT-2→GPT-4 = preschooler→smart high schooler in 4 years. compute 0.5 OOM/year + algorithm efficiency 0.5 OOM/year + unhobbling = 2024→2027 another 4–5 OOMs |
| Wang Kai(Brute-force aesthetics school) | AGI before 2035 | Leans toward Aschenbrenner / Viewing the present from an AGI perspective / "Assume AGI arrives on 2035-01-01" |
| Karpathy(2025-06) | Agentic Decade 2025-2035 | "Don't be fooled by 'Year of the Agent' / Agent is not 2025 year one, it's a slow 10-year evolution / 2032–2035 maturity period" |
| Kai-Fu Lee(2025-01) | Scaling Law is slowing down | "Only big tech can continue making super-large models / 2025 accelerates commercialization and also accelerates elimination" |
| Zhu Xiaohu(2024-03) | Before controllable nuclear fusion is achieved, I don't believe in AGI | "OpenAI's technology iteration curve is still relatively steep / open source lags behind non-open source by one year" |
| Cohere Aidan(2024-10) | General + vertical coexist | "Scaling Law still holds and will remain valid for a considerable period. But most companies' businesses cannot yet support GPT-4-level deployment costs" |
| Embiricos(2026-05 OpenAI) | A few years later, big companies catch up | "Next year (2026) — early adopters see agent self-sufficiency.AGI is reached = when productivity growth feeds back to AI labs" |
Aschenbrenner quote(2024-06):
"By 2025/26, these machines will outpace college graduates. By the end of the decade, they will be smarter than you or I; we will have superintelligence, in the true sense of the word."
"Hundreds of millions of AGIs could automate AI research, compressing a decade of algorithmic progress (5+ OOMs) into 1 year."
Objection: Yann LeCun / Gary Marcus criticize it as linear extrapolation fallacy — the OOMs scaling assumption may hit walls in data / algorithms / physics.
Real disagreement:Fundamental difference in betting on time windows:
- 2027 school → bets on "Intelligence Explosion / fast takeoff"
- 2030+ school → bets on "Scaling Law diminishing returns"
- 2035 school → bets on "Agentic Decade 10-year slow evolution"
- "Never arrives" school → bets on "controllable nuclear fusion / physical limits"
2.2 Disagreement 2: What is the real bottleneck of AGI?
| Perspective | Position |
|---|---|
| Aschenbrenner | Compute + algorithm efficiency(OOMs counting)/ trillion-dollar cluster |
| Kai-Fu Lee | Scaling Law diminishing returns(small enough, fast enough, cheap enough, capable enough) |
| Embiricos(OpenAI Codex) | Human typing speed + multitasking management speed(Not GPU compute!) |
| Karpathy | From Demo to product(In the LLM era it's not 10x difficulty, it's 1000x difficulty) |
| Jensen Huang synthesis | Models are eating applications, LLM = complete application |
Embiricos original words:
"The current underappreciated limiting factor is literally human typing speed or human multitasking speed."
Translation: The currently severely underestimated limiting factor is human typing speed, or human multitasking processing speed.
Humans need to:
- Write prompts
- Manually verify AI work(code review)
- Multitasking management
Outlook:AGI is reached = when productivity growth feeds back to AI labs(When productivity growth feeds back to AI labs, it triggers recursive acceleration).
2.3 Disagreement 3: Has Software 3.0 already become widespread?
| Perspective | Position |
|---|---|
| Karpathy | 3.0 dominant, full-stack penetration within 5 years — "You don't need to know how to write code to start a business" |
| Jensen Huang's Practical Experience(newtype · AI Coding Practice and Tool Stack Evolution §10) | Programming thinking is required, programming experience is not —— "Don't believe those UP hosts who say you can develop with zero foundation" |
| Wang Kai's Practical Experience(Wang Kai's AI Coding Hands-On Guide: From Cursor to Claude Code Terminal Workflow) | Token consumption will increase a hundredfold starting in 2026 —— Claude Code terminal is the primary tool, browsers are only for reading information |
| Xiaohongshu vibecoding practice | 0-80 vs 80-100 life-or-death line —— 80→100 is what makes a product truly usable |
| Vibe Coding cleansing prophecy(Jensen Huang) | Code written by AI may contain many hidden pitfalls —— It's fine at small scale, but explodes when complexity grows |
Real disagreement:Whether Software 3.0 is already "end-to-end usable":
- Karpathy's position:Natural language programming is already dominant; 1.0 is not dead but 3.0 is mainstream
- Practitioner position:3.0 makes 0→80 fast, but 80→100 still requires 1.0 + programming thinking
3. Detailed Explanation of Aschenbrenner's Situational Awareness (Exclusive)
Wang Kai Collection / Close Reading of 2024-2026 Cutting-Edge AI Papers and Reports §I:
3.1 Core Claims of the 5 Major Chapters
| Chapter | Core |
|---|---|
| I. Counting the OOMs | GPT-2→GPT-4 = preschool child → smart high school student over 4 years.2024→2027 will bring another 4-5 OOMs = an equivalent leap |
| II. Intelligence Explosion | Hundreds of millions of AGI-automated AI researchers,compressing 10 years of algorithmic progress into 1 year (5+ OOMs in 1 year). Human-level → superhuman-level will happen within 1 year |
| IIIa. Trillion-Dollar Cluster | $10B → $100B → $1T compute clusters / US electricity production will increase by dozens of percentage points by the end of this decade |
| IIIb. Lock Down the Labs | Safety at US AI labs is a joke / They are giving away key AGI secrets to the CCP for free |
| IIIc. Superalignment | Controlling AI that is far smarter than us is an unsolved technical problem |
| IIId. The Free World Must Prevail | Superintelligence = decisive economic and military advantage |
| IV. The Project | There will be some kind of government AGI project in 2027/2028 / "in a SCIF, the endgame will be on" |
3.2 Wang Kai's Applications
Repeatedly cited in posts as the source for "viewing the present from an AGI perspective":
- Wang Kai's Decision-Making System for the AI Era "Viewing the present from an AGI perspective" — a way of thinking directly derived from this paper
- Wang Kai's Observations on AI Industry Judgments and Big Tech Competition Landscape "The Three Scaling Law Debates" — corresponding to the paper's "counting the OOMs"
- Wang Kai's RSI AI Investment Strategy and the "Good Asset, Good Price, Good Volatility" Framework "Interrogating Claude's Value in the AI Era" — using Aschenbrenner's timeline as a stress test
3.3 Key Sentences (Original Text)
"By 2025/26, these machines will outpace college graduates."
"Hundreds of millions of AGIs could automate AI research, compressing a decade of algorithmic progress (5+ OOMs) into 1 year."
"The most extraordinary techno-capital acceleration has been set in motion."
4. Detailed Explanation of Karpathy's Agentic Decade (2025-2035)
Karpathy · Software in the era of AI(Software 3.0) §5:
4.1 Timeline Predictions
2025 — "打孔卡时代" —— 你能做 Demo,但跑不动复杂任务
2026-2028 — "小型机时代" —— 专业 Agent 在窄场景跑得动(代码、客服、合规)
2028-2032 — "PC 时代" —— 多 Agent 编排成熟,大量企业 deploy
2032-2035 — "成熟期" —— Agent 跟当年的 SaaS 一样普及
Implication: Don't bet all your time and money on "we must build an AGI Agent in 2025", you will hit a wall.
Corresponding approach:Build semi-autonomous products (suit), so that your product benefits at every stage, instead of gambling on "the day of full autonomy".
4.2 Don't Be an "Agent Demo Addict"
Karpathy repeatedly saw inside OpenAI:
- teams build an Agent Demo
- that gets hyped up on Twitter / in internal demos
- but deploying it to a real product is 1000x harder
- and it eventually dies quietly
Counterexamples: Demos like Devin / AutoGPT / GPT Researcher went viral but no one pays for the actual product — because they promised "full automation", but full automation has extremely poor robustness.
5. Detailed Explanation of Jensen Huang's Three Factions of Scaling Law
newtype Observations on AI Business, Industry, and Investment §3:
5.1 Zhu Xiaohu (2024-03 / Realist)
"Right now OpenAI's technology iteration curve is still relatively steep, and open source is definitely one year or even a year and a half behind closed source. But once the closed-source technology iteration curve slows down, open source will catch up."
Core judgments:
- Large models are already in a price war (tens of millions for deployment in early 2023 → five million in June → less than one million by year-end)
- Midjourney can't hold on, because to-C is too low-frequency
- Before controlled nuclear fusion is achieved, I don't believe the Earth can truly reach AGI
5.2 Cohere CEO Aidan Gomez (2024-10 / Pragmatist)
"Scaling Law still holds, and will remain valid for a considerable period of time. But the reality is,most companies' businesses cannot yet support the deployment cost of GPT-4-level models. "
The trend for a long time to come = general + vertical coexistence:
- First use general large models for prototype design
- Then do vertical fine-tuning once it works
5.3 Kai-Fu Lee (2025-01 / Evolutionist)
"1. Scaling Law is slowing down. 2. Only big tech companies can continue to build super-large models; startups have an extremely low probability of success. 3. The models that AI applications need are small enough, fast enough, cheap enough, and capable enough, not necessarily the largest and most general. 4. If super-large models (such as Claude 3.5 Opus) cannot be made cheap, they don't need to be released; they can stay internally as 'teachers' to train 'students'. 5. In 2025, AI will accelerate both commercialization and elimination."
Golden quote:"When opportunities come, be brave in making decisions; the same is true when opportunities disappear."
5.4 Jensen Huang's Comprehensive Judgment (2025-03-27)
"Many people in China think that even if the models are not good enough, they can still compete on applications. But the current situation is,models are evolving into complete applications, eating up various vertical tracks and traditional jobs. "
→ Model = application, they are not separate.This is a further extension of LLM = OS.
6. Model Evolution Timeline Verification (Jensen Huang newtype-model-bench)
newtype Model Evaluation, Fine-Tuning, and Hardware Evolution —— 505 posts covering the complete 2024-2026 timeline:
2024-04 Cohere Command R+
↓
2024-05 GPT-4o + Claude 3.5 Sonnet
↓
2024-07 OpenAI AGI 5 等级(行业标准)
↓
2024-09 OpenAI o1(慢思考)
↓
2024-11 Mac mini M4 集群跑 Qwen 32B 18 tok/s
↓
2025-01 DeepSeek R1 APP(国民级)
↓
2025-03 Gemini 2.5 Pro
↓
2025-05 Claude 4(SWE-bench 72.5%)
↓
2025-08 GPT-5 + Claude Code Sub-agents
↓
2025-10 Anthropic Claude Skills
↓
2025-11 Kimi K2 Thinking + Gemini 3 Pro + Claude Opus 4.5
↓
2025-12 GPT-5.2(首次达到人类专家水平)
↓
2026-01 OpenCode 黑卡 $200 + Anthropic Cowork + NVDA 收 Groq $200 亿
↓
2026-04 Anthropic 要身份验证 + GitHub 暂停 Copilot 注册(算力扛不住信号)
↓
2026-05 Claude Opus 4.7
Key milestones compared against Aschenbrenner OOMs:
- What has already happened from 2024→2026 ~2 OOMs(compute + algorithm efficiency)
- What is still needed from 2026→2027 2-3 OOMs to reach the AGI predicted by Aschenbrenner
- Current actual pace is faster than Aschenbrenner's median prediction(LeCun-faction critics have been partially proven wrong)
7. Academic Foundations of Foundation Models (Jensen Huang Exclusive)
Foundation Models for NLP (Close Reading of Springer Textbook) —— Springer open textbook, 436 pages(Paaß + Giesselbach 2023-02 arXiv:2302.08575):
8-chapter structure:
- Ch1 Classical NLP Foundations
- Ch2 Three major PLM architectures(BERT / GPT / Transformer) + training / regularization / uncertainty / interpretability
- Ch3 6 major improvement directions(RoBERTa / ELECTRA / T5 / PaLM / MoE / few-shot / CoT)
- Ch4 GLUE / SuperGLUE / BIG-bench
- Ch5 IE Information Extraction
- Ch6 Text Generation (QA / Translation / Summarization / Dialogue LaMDA)
- Ch7 Multimodality (ViT / DALL-E 2 / NÜWA / GATO / DNABERT)
- Ch8 5 major harms + Grounding + Fast/Slow Thinking + Planning
Why academic foundations matter:
- Repeated reading will let you know that "GPT did not appear out of thin air"
- The architectural difference between BERT and GPT determines capability boundaries
- MoE / CoT / few-shot are algorithmic advances outside of Scaling Law
8. How to Choose — AGI-Perspective Paths by Identity
| Who you are | Primary method | Starting point |
|---|---|---|
| Investor / Focus on valuation | Aschenbrenner 2027 + Wang Kai AGI perspective | Situational Awareness 165 pages / Looking back from 2035 / Don't do the mainstream, do supporting infrastructure |
| AI entrepreneur | Karpathy Agentic Decade + Iron Man Suit | Don't build Robot, build Suit / 4 principles (context + orchestration + GUI + verification) |
| Product manager | Karpathy LLM psychology + Embiricos | Treat LLMs as "knowledgeable but unstable interns" / Mixed-Initiative design |
| Want to understand Scaling Law | Jensen Huang's three-faction debate | Read Zhu Xiaohu vs Cohere vs Kai-Fu Lee in comparison |
| Want to understand Software 3.0 | Karpathy's three-stage framework + Jensen Huang LLM=OS | 1.0/2.0/3.0 transitions / Where to use which |
| Want to see the model evolution timeline | newtype-model-bench 505 posts | Complete inflection points for 2024-2026 |
| Academic foundations | Foundation Models 436 pages | BERT/GPT/Transformer + MoE + CoT + few-shot |
Universal disciplines(for any identity):
- Don't treat LLMs as absolutely reliable programs(Karpathy LLM psychology)
- Don't be an Agent Demo addict(Karpathy)
- Build Suit rather than Robot(Iron Man 4 principles)
- Verify outputs within 5 seconds(rapid verification loop)
- View the present from an AGI perspective(Wang Kai / Don't do the mainstream, do supporting infrastructure)
- Pay attention to three types of OOMs: model / algorithm / unhobbling(Aschenbrenner)
- AGI before controlled nuclear fusion is still possible(a hedge against Zhu Xiaohu's view)
9. Reference List
Main original author wikis:
- Karpathy · Software in the era of AI(Software 3.0) —— Karpathy 2025-06 talk complete Chinese translation / Software 1.0/2.0/3.0 / LLM=OS/Utility/Fab / Iron Man Suit / Agentic Decade
- Wang Kai Collection / Close Reading of 2024-2026 Cutting-Edge AI Papers and Reports §I —— Aschenbrenner Situational Awareness 165 pages / AGI by 2027 / OOMs / Trillion-dollar cluster
- newtype Observations on AI Business, Industry, and Investment §3 —— Three-faction Scaling Law debate (Zhu Xiaohu / Cohere / Kai-Fu Lee) + Jensen Huang's synthesis
- newtype Model Evaluation, Fine-Tuning, and Hardware Evolution —— 505 posts of 2024-2026 model evolution timeline
- Foundation Models for NLP (Close Reading of Springer Textbook) —— Foundation Models for NLP 436-page academic foundation
- Wang Kai's Observations on AI Industry Judgments and Big Tech Competition Landscape §5 —— Wang Kai's three-faction positions / Leans toward the brute-force aesthetics faction / AGI before 2035
- Wang Kai's RSI AI Investment Strategy and the "Good Asset, Good Price, Good Volatility" Framework —— Viewing the present from an AGI perspective / Interrogating Claude's value in the AI era
- Lenny's Newsletter / 2026 Trio Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI) Embiricos —— AGI bottleneck = human typing speed / Mixed-Initiative / Productivity feedback triggers recursive acceleration
Related topic wikis (this series):
- AI Industry Judgments and Corporate Competition (Cross-Author Thematic Synthesis) —— AI Industry Judgments (Batch 2)
- AI Investment and Valuation (Cross-Author Thematic Synthesis) —— AI Investment Valuation (this batch, Batch 4)
- AI Coding Tool Stack Evolution (Cross-Author Thematic Synthesis) —— AI Coding Tool Stack (Batch 2)
- AI Agent Architecture and Multi-Agent Orchestration (Cross-Author Thematic Synthesis) —— Agent Architecture (Batch 2)
History
- 2026-05-18:Topic-dimension wiki first release. Synthesized and rewritten from 8 author wikis. Explicitly marks 3 disagreements: AGI timeline (2027 vs 2030+ vs 2035) / real bottleneck / whether Software 3.0 is already widespread
Sources and related materials
- Karpathy · Software in the era of AI(Software 3.0)
- newtype Observations on AI Business, Industry and Investment
- newtype Model Evaluation, Fine-Tuning and Hardware Evolution
- Foundation Models for NLP (Close Reading of Springer Textbook)
- Wang Kai Collection / Close Reading of Cutting-Edge AI Papers and Reports 2024–2026
- Wang Kai’s Judgments on the AI Industry and Observations on the Competitive Landscape of Large Companies
- Wang Kai RSI AI Investment Strategy and the "Good Asset, Good Price, Good Volatility" Framework
- Lenny's Newsletter / 2026 Three-Way Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI)
- AI Industry Judgments and Corporate Competition (Cross-Author Thematic Synthesis)
- AI Investment and Valuation (Cross-Author Thematic Synthesis)
- AI Agent Architecture and Multi-Agent Orchestration (Cross-Author Thematic Synthesis)
- Claude Code Workflow (Cross-Author Thematic Synthesis)