Thinking Can Be Outsourced. Understanding Cannot.

---

Prologue: The Developer Who Doesn't Code

In December 2025, an event horizon arrived for AI coding.¹

Code blocks generated by models began building and running perfectly on the first try — with zero human intervention. Research, planning, code, testing, debugging — the entire pipeline completed autonomously.

Code writing has already been outsourced.

But the real insight begins here. In an interview with Sequoia Capital, the OpenAI co-founder, former Tesla Autopilot lead, and recent Anthropic joiner delivered the sharpest sentence I've heard all year:

Thinking can be outsourced. Understanding cannot be outsourced.

Why "understanding" of all things? In an era where everything gets automated, what is the last thing humans must retain?

---

I – Software 3.0: The Era When Code Disappears

The paradigm of software development has shifted once again.

Karpathy defines it as a progression through Software 1.0 → 2.0 → 3.0.

Version	Method	Example
1.0	Humans write code directly	C, Python, JavaScript
2.0	Humans build datasets and train neural nets	Image classifiers, recommendation systems
3.0	Humans write prompts for LLMs	"Overlay food photos on my nano banana menu"

It sounds paradoxical. In 1.0, humans wrote code. In 2.0, humans prepared data. But in 3.0, humans simply speak. One prompt creates an app.

Karpathy's MenuGen project proved it. When building an app that generates food photos for restaurant menus using an image generator, the process used to be long and painful:

Photograph menu → Extract items via OCR → Generate images → Build backend API → Deploy to Vercel

But now that app can be replaced by a single prompt. Code didn't just get faster — general information processing itself became automated. Where code disappeared, neural networks moved in directly.

graph LR
    subgraph "Software 1.0"
        A[Human writes code] --> B[Compiler runs]
        B --> C[Program output]
    end

    subgraph "Software 2.0"
        D[Human builds dataset] --> E[Neural net trains]
        E --> F[Model output]
    end

    subgraph "Software 3.0"
        G[Human writes prompt] --> H[LLM reasons]
        H --> I[Agent executes]
    end

    A -.->|Automated away| G
    D -.->|Automated away| G

Here comes the key insight. When code disappears, the important question is no longer "how do we code?" It's "what should we build?"

---

II – The Moats Built by Reinforcement Learning

Frontier AI labs train the highest-performance models through one core mechanism: massive reinforcement learning (RL) environments.

RL works by giving AI behavior positive or negative scores based on outcomes, then pushing the model to maximize its score. When a clear verification reward environment exists, RL loops run infinitely, explosively raising model capability.

This is why domains with clear answer verification — math, coding — see performance skyrocket, while subjective and ambiguously-verified areas show uneven, plateaued capability growth.

The fact that models are this uneven means we must never trust AI unconditionally, and humans must always stay in the loop to control them.

Karpathy's view on entrepreneurial opportunity lives here too. If a business operates in off-distribution areas that existing models fundamentally cannot solve, companies must build proprietary fine-tuning datasets and their own RL environments just to survive.

pie title The Uneven Shape of AI Capability (by verifiability)
    "Clear verification domains" : 85
    "Partial verification domains" : 45
    "Ambiguous verification domains" : 20

Specifically:

• Math, coding: Clear answer verification → Explosive RL-driven growth
• Chess: OpenAI intentionally injected massive chess board data into pre-training → Geometric rise from GPT-3.5 to GPT-4
• Literature, art: Subjective, but assembling multiple LLM judge panels enables reasonable verification

Among these, what gives startup founders a competitive edge is this: building differentiated, massive, and diversified RL environments directly within specific industry domains.

---

III – Human-in-the-Loop: The Human Domain

In the agent era, what human skills become more valuable?

Karpathy's answer is clear. Current-generation AI agents, no matter how impressive, are fundamentally intern-level entities in terms of skill.

You must continuously cultivate aesthetic sense, final judgment, engineering taste, and high-level supervisory ability — and the value of those abilities will skyrocket.

Human roles are changing. Instead of laboring over rote coding lines, humans must define things agents can never properly design alone:

• Clear, detailed product specifications (Specification)
• Structural masterplans (Plan)
• Rigorous technical documentation (Docs)

And at the deepest level. Existing LLMs and agent systems are brilliantly skilled at combining and generating "thought computation" — but they have never reached the domain of genuinely 'understanding' the essence of context the way humans do.

The fundamental reason I'm so obsessed with projects that autonomously build my own personalized knowledge wiki using LLMs lies here. Feed massive trend articles and raw fixed datasets to agents, and they synthesize and organize knowledge from every angle, continuously expanding my personal wiki dictionary.

In other words: leverage the increasingly cheap computational power of AI as a lever to explode your own human understanding horizon faster and more massively than anyone on Earth.

---

IV – Machine-Legible Infrastructure: The Unseen Revolution

What is the most fundamental change the agent era will bring?

Karpathy's argument is bold. All digital infrastructure, software, and network protocols built across humanity must be completely rewritten from the ground up so that agents can read them and take direct action.

My agent must be able to parse infrastructure specs in 0.001 seconds and control infrastructure automatically. Providing machine-legible, perfectly optimized specification files and API guidelines must become the basic standard.

One step further: in the near future, my agent and your agent will communicate directly through protocols, negotiating meetings and business deals in milliseconds, then presenting only final approval reports to humans. That world will become ordinary.

graph TD
    subgraph "Present: Human-Centric"
        A[Human reads] --> B[Human thinks]
        B --> C[Human inputs]
        C --> D[System executes]
    end

    subgraph "Future: Agent-Centric"
        E[Agent reads] --> F[Agent thinks]
        F --> G[Agent executes]
        G --> H[Human just approves]
    end

    A -.->|Rewritten| E
    style E fill:#4CAF50,color:#fff
    style H fill:#4CAF50,color:#fff

The way systems are designed changes completely. Currently, every system is strictly designed for the convenience of 'human users' — with vision, hearing, and fingers. They are relics of the past.

Agents being rewritten and able to read and act immediately — this is the unseen but unstoppable transformation.

---

💭 Questions to Ponder After Reading This

1. Can 80% of your work already be replaced by a single prompt? Or do you have a unique domain that cannot be replaced?

2. What exclusive data and experience do you uniquely possess to expand your "human understanding"?

3. Have you ever created a document that an agent can read and act on directly?

Share your thoughts in the comments.

---

Conclusion: The Paradoxical Survival Strategy

In an era where intelligence becomes extremely cheap, paradoxically, the learning ability to explode one's high-level understanding by leveraging machines, and the engineering comprehension that pierces through essential mechanisms will remain humanity's greatest and most valuable survival strategy.

"Thinking can be outsourced. But understanding cannot be outsourced. The final directing bottleneck of knowing what we must ultimately build remains entirely within the human brain."

In the age when machines think, what humans must do is not to get smarter. It is to understand more deeply.

---

Sources

---

If this article helped you, share it with one friend.

Footnotes

1. Andrej Karpathy, Sequoia Capital Interview, 2026.04.30 - https://www.youtube.com/watch?v=96jN2OCOfLs&t=349s ↩