Deutschภาษาไทย中文
中文
Sequoia AI Ascent 2026 Deep Dive — Series 1/3
"AGI is not the future — it's now. And you only have 18 months."
Deep Research

Sequoia AI Ascent 2026 (Part 1):
AGI Is Here — Now What?

9 AI leaders, 150 founders, one world-changing conclusion
Independent Research | LittleX Research Lab | 2026-05-02
In 2025, Sequoia said: "AI is not the future — it's now."
In 2026, they revised that: "Functional AGI has arrived — the question is no longer 'does it exist?' but 'who will capture that $10 trillion?'"

This is not science fiction. This is the verdict of Silicon Valley's most influential venture capital firm, delivered to 150 top founders and backed by real money. When Pat Grady said "This is AGI" on stage, nobody in the audience laughed.

Abstract

In April 2026, Sequoia Capital held its fourth annual AI Ascent conference. Three partners — Pat Grady, Sonya Huang, and Konstantine Buhler — made a stunning declaration in the Keynote: from a functional standpoint, AGI has arrived. Their definition is not academic — it's commercial: when you can dispatch an agent to execute a task and it can recover from failure and persist until the job is done, that is functional AGI.

This article synthesizes the full Keynote transcript, plus highlights from six talks by Demis Hassabis (DeepMind), Greg Brockman (OpenAI), and others. Using first-principles thinking, we break down: what functional AGI means, the exponential acceleration of the METR curve, the $10 trillion shift from Copilot to Autopilot, the three pillars of the Agent economy, and Taiwan's 18-month window of opportunity.

Functional AGI METR Curve Agent Economy Copilot vs Autopilot $10 Trillion Services Market Cognitive Revolution MAD Strategy Sequoia Capital Taiwan AI Transformation
Table of Contents
  1. What Is Functional AGI? — Grady/Huang's Definition and the METR Curve
  2. Copilot vs Autopilot: The $10 Trillion Question
  3. Hassabis Says We're 75% There — What's the Remaining 25%?
  4. Human Attention as the Bottleneck — Brockman's Perspective
  5. Three Pillars of the Agent Economy — Buhler's Perspective
  6. What Should Taiwan Do? The 18-Month Window
  7. Historical Parallels: AI Bubble vs the 2000 Dot-Com Bubble
  8. Business Insights: 5 Decision Questions for Business Leaders
  9. Conclusion
  10. Series Guide
  11. References

I. What Is Functional AGI?

In academia, there is still no consensus on the definition of AGI. But Sequoia's three partners are not academics — they are investors who have reviewed tens of thousands of companies, and their definition comes from business reality.

Pat Grady's Definition: Self-Completing Work

"If you can dispatch an agent to do a job and it can recover from failure and persist until that job is done — that feels pretty much like AGI."
— Pat Grady, Sequoia Capital Managing Partner

In plain terms: if you can send an AI to do something, it can solve problems on its own, won't give up halfway, and keeps going until it's done — that's AGI.

Grady specifically emphasized that he is not proposing a "technical definition." He studied economics and researches the collision between founders and markets — in other words, "business." From a commercial, practical, functional standpoint, the emergence of long-horizon agents is a watershed moment.

Three Turning Points: From ChatGPT to AGI

November 2022
The ChatGPT Moment — The world saw the power of pretraining. Stunning, but essentially a "smarter search engine."
2024
O1 Reasoning Models — A second Scaling Law emerged: inference-time compute. AI doesn't just memorize answers — it begins to "think."
2025–2026
Claude Code / Opus 4.5–4.7 — The world saw the power of long-horizon agents. Not conversation, but execution. This is a "discontinuous leap" from the previous two.

The first two were "a faster horse" — boosting your productivity by 10–40%, but not changing how you work. The third is "the arrival of the car" — boosting productivity by 10–40x, fundamentally changing the nature of work itself.

The METR Curve: Iron-Clad Proof of Exponential Acceleration

In the Keynote, Sonya Huang presented research data from METR (Model Evaluation and Threat Research) — arguably the most important slide of the entire event:

Core Metric

The METR curve measures: how long an AI model can persist on complex tasks without "going off the rails."

AI task duration doubles every 7 months.

~30 minutes
2025: How long an agent could sustain work
~Several hours
2026 (now): Reached "hour-scale" capability
~A full day
2028 forecast: An agent can complete an entire workday
~An entire year
2034 forecast: An agent can operate continuously for a year
First Principles

What is the essence of the METR curve? It doesn't measure "intelligence" — it measures reliability multiplied by persistence. An average human employee can maintain focused work for 4–6 hours. When an agent's METR score reaches that level, it is functionally equivalent to a full-time employee — except it doesn't need sleep, never takes days off, and can be infinitely replicated.

At the rate of doubling every 7 months, we reach this crossover point around 2028. This is not speculation — it's an extrapolation based on an already-observed exponential curve.

From AutoGPT to Claude Code: Why Now?

Sonya specifically mentioned AutoGPT and BabyAGI from 2022 — which also wrapped GPT-3 in a loop and let it run autonomously. The result?

"Kind of cute, kind of endearing, but completely useless."
— Sonya Huang, describing the 2022 agent experiments

Three years later, why does the same architectural concept suddenly work? The answer is that three components matured simultaneously:

Sonya used a brilliant analogy: the SaaS tools we built for humans over the past two decades have now become "prosthetic limbs" for agents. SaaS isn't dead — it's about to see the biggest user explosion ever, except the users are no longer human.

II. Copilot vs Autopilot: The $10 Trillion Question

This is the most commercially impactful framework of the entire Keynote. Pat Grady opened with a chart that silenced the room:

Traditional Software Market (Copilot)

Over the past 15 years of cloud transformation, the software TAM grew from $350 billion to $650 billion. Cloud accounted for $400 billion of that.

This is your budget for selling tools: $1 spent on software for every...

Services Market (Autopilot)

The services market AI agents can address: $10 trillion. U.S. legal services alone are $400 billion — equal to the entire software market.

This is your budget for selling outcomes: $6 spent on services.

Key Insight

Copilot sells tools — "This software makes your lawyer more efficient." The customer spends $1.

Autopilot sells outcomes — "This agent litigates and settles your case." The customer spends $6.

Same customer, 6x the addressable budget. That's why the AI wave dwarfs the cloud era in scale.

The Four Stages of Agent Evolution

Using "writing code" as an example, Sonya illustrated the four stages from assistance to full autonomy:

StageYearFormHuman Role
Tab Autocomplete2023AI autocompletes beside youTypist
Agentic Dev2025You direct the agentManager
Async / Background2026Agent runs in the backgroundExecutive
Dark Factory2027+Fully unsupervisedOwner (occasional reports)

"Dark Factory" is a term from manufacturing — referring to fully automated factories that don't even need the lights on. Sonya said she has already seen this pattern in production environments, including cybersecurity companies.

Concrete Scenarios of "Services as the New Software"

Sonya listed agent applications that are already happening:

First Principles

Why is "services as the new software" more than just marketing speak? Because the fundamental cost structure of enterprises is changing. Scaling with humans is hard (hiring, training, management), while agents scale infinitely with compute. Humans need salaries; agents only need tokens.

Today, humans are still smarter at most things — but as Scaling Laws continue, agents will soon surpass humans in many domains. This isn't humans being replaced — it's the human role shifting from "the one who does things" to "the one who decides what to do."

The MAD Strategy: A Survival Guide for Application-Layer Founders

Pat Grady proposed the MAD framework — for every founder building on top of foundation models:

"You cannot pass 15 cars in the sun, but you can pass 15 cars in the rain. And right now, there is a torrential downpour of new capabilities coming out of the foundation models — which means no lead is safe. But it also means anybody can win."
— Pat Grady, Sequoia Capital

III. Hassabis Says We're 75% There — What's the Remaining 25%?

Google DeepMind CEO Demis Hassabis made a specific yet bold assessment during his AI Ascent interview:

"We're three-quarters of the way to AGI."
— Demis Hassabis, CEO of Google DeepMind

By his estimate, we will reach true AGI around 2030. But what's more interesting is his breakdown of the "75% completed" and the "remaining 25%."

The 75% Already Completed

The Remaining 25% — The Hardest Part

Hassabis's Deeper Perspective

Hassabis offered a more profound philosophical insight: information may be more fundamental than energy and matter. If the universe is fundamentally about information processing, then the significance of AI is even more far-reaching than we currently imagine — "and it's already pretty profound."

He used a precise analogy: our current AI is like the 120 years of tinkering before the steam engine was invented — something is moving, but we haven't yet discovered the "laws of thermodynamics" behind it. The foundational science of AI — why it works, why Scaling Laws exist — we still don't truly understand.

"Build the Tool First, Then Cross the Next Rubicon"

"We should build a tool first — an incredibly intelligent and useful and precise tool — and then cross the next rubicon. That's already profound enough."
— Demis Hassabis

This is Hassabis's pragmatic advice for the AGI race: don't rush toward "superintelligence" — first perfect the "super tool." That step alone is already enough to fundamentally change the world.

The Founder's Time Window

Hassabis also shared a lesson from his own experience running a game company (Elixir Studios):

Founder Wisdom

Be 5 years ahead, not 50.

Technology that's too far ahead can't find a market; technology that's too far behind can't find an edge. Five years is the sweet spot — early enough to build a moat, but not so early that you burn through your capital waiting for the market to mature.

IV. Human Attention as the Bottleneck

OpenAI co-founder Greg Brockman offered a deceptively simple yet profoundly important insight during his interview:

"Human attention is going to be this incredibly scarce resource. The doing of things now is easy. The 'is this a good thing? Is this what I wanted?' — that is going to become the single most important bottleneck."
— Greg Brockman, Co-founder of OpenAI

The Shift from "Doing" to "Judging"

The traditional workplace bottleneck was execution capacity — not enough people to do the work. In the agent era, execution capacity approaches infinite supply. The new bottleneck becomes judgment — "Should this be done at all? Is the result what I actually wanted?"

The Old Bottleneck

Not enough people, not enough skills, not enough time.

"I know what to do, but I can't find anyone to do it."

The New Bottleneck

Not enough attention, not enough judgment, not enough direction.

"AI can do anything, but I'm not sure what it should do."

Brockman's Two Practical Recommendations

"Why are you explaining to your computer what's going on? That makes no sense."
— Greg Brockman, questioning why we still manually input information that AI could obtain on its own
First Principles

Brockman's point boils down to this: "doing" is being commoditized, and "thinking" becomes the only source of differentiation.

In this framework, the most valuable capability of the future isn't "being able to use AI" — it's "knowing what to have AI do." This echoes Karpathy's point from another talk: "You can outsource thinking, but you can't outsource understanding."

V. Three Pillars of the Agent Economy

Konstantine Buhler covered the third segment of the Keynote — "What happens next?" He drew a sweeping analogy: the cognitive revolution will reshape the world just as the Industrial Revolution did.

A Parallel Universe to the Industrial Revolution

Industrial Revolution (Physical Work)

Before 1700: 99%+ of physical work was done by humans and animals.

Waterpower → Steam engine → Internal combustion → Electric motor

2026: 99%+ of physical work is done by machines. Planes carry you; factories manufacture all goods.

Cognitive Revolution (Mental Work)

Before 2020: 99%+ of cognitive work was done by humans.

Calculators → Digital computing → Neural networks → ???

The future: 99.9% of cognitive work will be done by machines. We are at this inflection point right now.

"Today 2026, you could estimate that 99 plus percent of all the physical work done on planet Earth for humans is done by a machine. We believe a similar pattern is going to happen in cognition. We're just a little earlier on."
— Konstantine Buhler, Sequoia Capital

When Agents Become "Economic Actors"

Before handing off to Buhler, Sonya posed a series of unsettling questions:

For the agent economy to function, three pillars are needed:

Pillar 1
Identity
Agents need verifiable digital identities. Who built it? Who does it represent? What are its permissions? This amounts to building an "AI passport" system.
Pillar 2
Protocol
Agents need standardized communication and transaction protocols. MCP (Model Context Protocol) is a first step, but a complete infrastructure for payments, contracts, and arbitration is still needed.
Pillar 3
Trust
How do we verify agent outputs? Who is liable when things go wrong? An entirely new "machine trust mechanism" is needed — likely combining cryptographic verification, reputation systems, and insurance frameworks.
First Principles

Looking back at history: the Industrial Revolution required new economic infrastructure to unlock its full value — limited liability companies, central banks, patent law, labor law, antitrust law. The cognitive revolution will likewise require entirely new infrastructure.

Whoever builds these three pillars first will control the gateway to the next economic era. This is not a technology problem — it is an institutional innovation problem.

Buhler's Four Stories: The Lesson of Aluminum

Buhler used an elegant historical story to illustrate the impact of the cognitive revolution: in 1884, the United States capped the Washington Monument with the most precious metal in the world at the time: aluminum. 100 ounces of pure aluminum sat atop the monument, worth more than gold.

Then the electrolytic process for producing aluminum was invented. The same metal crashed in price by 99.5%. Today you wrap your sandwich in aluminum foil and throw it away.

Core Metaphor

AI's impact on "cognitive work" is like the electrolytic process's impact on aluminum. The expertise you pay millions for today may become as cheap as aluminum foil tomorrow. But the ubiquity of aluminum didn't destroy the world — it created entirely new industries: aviation, construction, packaging, and more.

The question isn't "will cognitive work lose its value" — it will. The question is "can you use cheap cognitive capacity to create unprecedented value?"

VI. What Should Taiwan Do? The 18-Month Window

Sequoia's framework holds special significance for Taiwan. Taiwan sits at the core of the AI supply chain yet on the periphery of the AI application layer — a dangerous asymmetry.

Taiwan's Structural Advantages

TSMC
Over 90% of the world's advanced process chips — the physical foundation of AI
Hardware Ecosystem
Foxconn, Quanta, Wistron — the world's factory for AI servers

Taiwan's Structural Risks

Application Layer Gap
Of the 12 AI-first verticals Sequoia identified, Taiwan has virtually no globally competitive AI application companies
Services Sector Lag
Taiwan's legal, healthcare, and financial services digitization lags far behind the U.S. and Singapore

The Logic Behind the 18-Month Window

Why 18 months? This comes from cross-referencing the METR curve with Sequoia's market analysis:

Now — 2026 Q2
Agents can handle hour-scale tasks. Enterprises are still "testing the waters." The best time to build localized AI infrastructure.
Mid-2027
Agents can handle full-day tasks. Global AI application companies begin "invading" local markets. Language and regulations are no longer a moat — agents will learn the language themselves.
After 2028
Agents can handle full workdays. Regions without agent infrastructure will be directly "covered" by globalized AI service companies — just as Uber covered local taxi companies.
Five Action Items for Taiwan
  1. Legal Tech: Taiwan's legal system is unique (civil law tradition + Traditional Chinese) — a natural moat. Build a legal agent ecosystem within 18 months.
  2. Medical AI: The National Health Insurance database is one of the world's most complete single-payer datasets. Leverage this advantage to train localized medical agents.
  3. Manufacturing Agents: Taiwan's manufacturing know-how is tacit knowledge — ideally suited for encapsulation and scaling through agents.
  4. Agent Infrastructure: The three pillars of identity, protocol, and trust — Taiwan could follow Estonia's e-Residency model and build an agent identity system for the Asia-Pacific region.
  5. TSMC's Extension: Expand from "chip foundry" to "AI inference foundry" — not just hardware, but computational services as well.
First Principles

Taiwan's core challenge isn't "whether it has AI technology" — TSMC is the best AI technology there is. The question is whether Taiwan can upgrade from "selling shovels" to "mining the gold."

In a gold rush, selling shovels is good business — but when the shovels start mining gold on their own, the shovel seller who doesn't transform can only watch as agents carry the gold away.

VII. Historical Parallels: AI Bubble vs the 2000 Dot-Com Bubble

Historical Perspective

Every time a new technology wave arrives, "bubble talk" follows. Many people are now asking: is AI a repeat of the 2000 dot-com bubble? Pat Grady's framework provides a clear answer.

Similarities

Fundamental Differences

Dimension2000 Dot-Com Bubble2026 AI Wave
Type of RevolutionCommunications revolution (information distribution)Compute revolution (information processing)
TAM$650B software market$10T+ services market
Revenue GrowthPets.com burned cash with no revenueMultiple AI companies have surpassed $1B ARR
Technology MaturityBroadband not yet widespreadFoundation models already production-ready
Scaling LawsDid not existPretraining + inference-time + RL — three Scaling Laws still holding
First Principles

The core lesson of the 2000 dot-com bubble isn't "the internet had no value" — it's "the value was priced in too early." The bubble burst, but Amazon, Google, and Facebook grew from its ashes.

Will AI have a bubble correction? Possibly. Some companies will die. But Grady's "compute revolution vs communications revolution" framework tells us: AI's fundamentals are stronger than the internet's, because it doesn't just change how information flows — it changes how information is processed. That's a deeper transformation.

Bubbles aren't scary. What's scary is avoiding the market out of bubble fear, only to realize — after it pops — that you missed the next Amazon.

VIII. Business Insights: 5 Decision Questions for Business Leaders

5 Questions You Should Ask Yourself Today

Question 1: Is your company selling tools or selling outcomes?

If you're still selling software licenses (Copilot model), your addressable market is $1. If you can reposition to sell outcomes (Autopilot model), your addressable market is $6. Same customer, 6x the difference.

Action: List your top 10 customers. For each one, ask: "If we stopped selling tools and directly sold the outcome they want, how much could we multiply our pricing?"

Question 2: Which of your employees' tasks can be replaced by agents within 6 months?

This isn't about layoffs — it's about redeployment. One employee plus one agent can match the productivity of 10 people in the past. But this requires redesigning workflows, not simply "adding an AI tool."

Action: Conduct a "task verifiability audit" (borrowing Karpathy's framework). Prioritize automating high-verifiability tasks (code, data analysis, compliance checks).

Question 3: Have you made your context investment?

Brockman said context is a one-time investment. Is your company's knowledge base, decision-making process, and customer data structured to a degree that agents can directly use?

Action: Assign one person (or one agent) to convert your company's core knowledge into an agent-consumable format within 90 days.

Question 4: Where are you weakest in the MAD framework?

Modes (customer stickiness), Affordance (ease of use), Diffusion (market penetration). If all three are weak, you'll be displaced soon.

Action: Score your product on MAD (1–10 each), identify the weakest dimension, and concentrate resources on strengthening it.

Question 5: Will your moat still exist in 18 months?

If your moat is "our engineers are excellent" or "our software is mature," these can be leveled overnight in the agent era (Notion rewrote 8 million lines of code in 6 weeks; Brett Taylor rewrote Sierra over a single weekend). Durable moats are customer relationships, unique data, and institutional trust.

Action: List your 3 biggest moats. For each one, ask: "If agents could do everything a human can, would this moat still exist?"

IX. Conclusion: The Car Is Here — Are You Getting In?

Let's return to Pat Grady's metaphor: the AI applications of the past few years were "a faster horse." Now, "the car has arrived."

At AI Ascent 2025, Sequoia's message was "AI is not the future — it's now."

In 2026, they revised it: "Functional AGI has arrived. The question isn't whether AGI exists, but who will capture that $10 trillion."

The core messages across six talks:

Pat Grady
AI is a compute revolution, not a communications revolution. The $10 trillion services market is wide open.
Sonya Huang
The agent era has officially begun. The METR curve doubles every 7 months. Services are the new software.
Konstantine Buhler
Cognitive revolution = Industrial Revolution 2.0. In the future, 99.9% of cognitive work will be done by machines.
Demis Hassabis
We're 75% of the way there. Build the tool first, then cross the rubicon. Full AGI by 2030.
Greg Brockman
Human attention is the new bottleneck. Context is a one-time investment. Judgment becomes the scarcest resource.
Andrej Karpathy
Software 3.0 = LLM as computer. You can outsource thinking, but you can't outsource understanding.
Final Takeaway

Sonya Huang's closing remark is the best summary of the entire conference:

"Whatever you can imagine building over the next hundred years, we think is now possible in a hundred days thanks to agents."

What you planned to build over a hundred years can now be done in a hundred days with agents.

This isn't hyperbolic marketing. This is the true story of the Zed founder completing three years' worth of plans solo with Claude Code over a holiday. This is the real case of the Notion team rewriting eight million lines of code in six weeks.

The car is here. You can keep riding your horse. But the car won't wait for you.


Sequoia AI Ascent 2026 Deep Dive — Article Series

[1/3] Sequoia AI Ascent 2026 (Part 1): AGI Is Here — Now What? (This article)
[2/3] Sequoia AI Ascent 2026 (Part 2): Karpathy on Software 3.0 — From Vibe Coding to Agentic Engineering [3/3] Sequoia AI Ascent 2026 (Part 3): Jim Fan on the Robotics Endgame — The Grand Parallelism Theory and Physical AI

References

  1. Sequoia Capital, "This is AGI: Sequoia AI Ascent 2026 Keynote" (YouTube: LRo33rnv6rQ), April 2026
  2. Sequoia Capital, "Demis Hassabis: We're Three Quarters of the Way to AGI" (YouTube: AFpeWo1GTeg), April 2026
  3. Sequoia Capital, "OpenAI's Greg Brockman: Why Human Attention Is the New Bottleneck" (YouTube: bBS93A0BeNI), April 2026
  4. Sequoia Capital, "Andrej Karpathy: From Vibe Coding to Agentic Engineering" (YouTube: 96jN2OCOfLs), April 2026
  5. Sequoia Capital, "Nvidia's Jim Fan on the End Game for Robotics" (YouTube: 3Y8aq_ofEVs), April 2026
  6. Sequoia Capital, "Waymo's Dmitri Dolgov: How Autonomous Cars Got 13x Safer Than Humans" (YouTube: I_0Kuf6Aa2c), April 2026
  7. METR (Model Evaluation and Threat Research), AI Task Duration Benchmarks, 2025–2026
  8. Sonya Huang & Pat Grady, "Generative AI's Act Two", Sequoia Capital blog, 2024
  9. CnYes, "Sequoia AI Ascent 2025 Key Takeaways", 2025
  10. eimba-sequoia-2026.vercel.app, Sequoia AI Ascent 2026 Conference Analysis, 2026