※ You can check Andrej Karpathy's complete insights on AI-era software through the full presentation video.

Key Messages at a Glance

  • Software is fundamentally changing for the first time in 70 years. Two rapid changes have occurred in recent years.
  • Three programming paradigms coexist: Software 1.0 (code), 2.0 (neural network weights), 3.0 (English prompts)
  • LLMs are the new operating system. They have characteristics of utilities, fabs, and operating systems all at once.
  • Currently similar to 1960s computing era. We’re using time-sharing methods on centralized clouds.
  • LLMs are “human souls”. They have encyclopedic knowledge but also many cognitive flaws.
  • Partially autonomous apps are the future. Human-AI collaboration is key rather than full autonomy.
  • The era of programming in English has arrived. Everyone can become a programmer.
  • Infrastructure for agents needs to be built. New digital information consumers have emerged.

Three Evolution Stages of Software

Software 1.0: Traditional Coding

  • Code that directly commands computers
  • Written in programming languages like Python, JavaScript
  • Basic paradigm unchanged for 70 years

Software 2.0: Neural Network Weights

  • Neural networks trained on data
  • Weights and parameters replace explicit code
  • Revolutionary change in how we solve problems

Software 3.0: English Prompts

  • Natural language instructions to LLMs
  • Programming through conversation
  • Democratization of software creation

LLMs as Operating Systems

Three Characteristics of LLMs

1. Utility (Like Electricity)

  • Basic infrastructure everyone uses
  • Invisible but essential service
  • Standardized interface and access

2. Fab (Like Semiconductor Manufacturing)

  • Extremely expensive to build and operate
  • Requires massive capital investment
  • Centralized production, distributed consumption

3. Operating System

  • Manages resources and provides interfaces
  • Enables applications to run on top
  • Abstracts complexity for developers

Current State: 1960s Computing Revisited

  • Centralized mainframes → Cloud-based LLMs
  • Time-sharing systems → API rate limits and queues
  • Terminal access → Chat interfaces and APIs
  • Batch processing → Prompt-response cycles

The Nature of LLMs: “Human Souls”

Encyclopedic Knowledge

LLMs possess vast knowledge across domains:

  • Literature and history
  • Science and mathematics
  • Culture and languages
  • Technical expertise

Cognitive Limitations

But they also have human-like flaws:

  • Hallucinations and confabulation
  • Inconsistent reasoning
  • Bias and prejudice
  • Lack of true understanding

Implications for Development

This dual nature requires:

  • Careful prompt engineering
  • Robust verification systems
  • Human oversight and validation
  • Graceful error handling

The Future: Partially Autonomous Applications

Why Not Full Autonomy?

Full autonomy faces several challenges:

  • Reliability concerns
  • Ethical implications
  • User trust issues
  • Regulatory requirements

Human-AI Collaboration Model

The optimal approach combines:

  • AI for routine tasks
  • Human for critical decisions
  • Seamless handoffs
  • Transparent processes

Examples of Partial Autonomy

Email Management:

  • AI drafts responses
  • Human reviews and approves
  • AI learns from feedback

Code Generation:

  • AI writes initial code
  • Human reviews and refines
  • AI assists with debugging

Content Creation:

  • AI generates drafts
  • Human edits and polishes
  • AI helps with research

Programming in English: Democratization of Development

Natural Language as Code

Programming is becoming more accessible:

  • No syntax to learn
  • Intuitive instructions
  • Immediate feedback
  • Iterative refinement

New Developer Categories

Traditional Programmers: Still needed for complex systems Prompt Engineers: Specialists in AI communication Citizen Developers: Domain experts who can now program AI-Assisted Developers: Hybrid approach users

Skills for the New Era

Essential skills include:

  • Clear communication
  • Problem decomposition
  • System thinking
  • Quality assessment

Infrastructure for AI Agents

New Digital Consumers

AI agents represent a new category of information consumers:

  • Different access patterns
  • Unique authentication needs
  • Specialized data formats
  • High-volume requests

Required Infrastructure

APIs for Agents:

  • Machine-readable formats
  • Structured data access
  • Real-time capabilities
  • Scalable architectures

Authentication Systems:

  • Agent identity management
  • Permission frameworks
  • Usage tracking
  • Security protocols

Data Formats:

  • Standardized schemas
  • Semantic markup
  • Structured outputs
  • Version control

Examples of Agent-Ready Services

Web Services:

  • RESTful APIs with clear documentation
  • GraphQL endpoints for flexible queries
  • Webhook systems for real-time updates

Data Providers:

  • Structured data feeds
  • Real-time streaming APIs
  • Semantic web standards

Tool Integrations:

  • Function calling interfaces
  • Plugin architectures
  • Workflow orchestration

Practical Implications for Developers

Adapting to the New Paradigm

Learn Prompt Engineering:

  • Understand LLM capabilities and limitations
  • Master effective communication techniques
  • Develop testing and validation strategies

Build Hybrid Systems:

  • Combine traditional code with AI capabilities
  • Design for human-AI collaboration
  • Implement robust error handling

Prepare for Agents:

  • Design APIs with agents in mind
  • Implement proper authentication
  • Provide structured data access

Tools and Frameworks

Development Environments:

  • AI-assisted IDEs
  • Prompt testing platforms
  • Hybrid development tools

Deployment Platforms:

  • Serverless AI functions
  • Container orchestration
  • Edge computing solutions

Monitoring and Analytics:

  • AI performance metrics
  • User interaction tracking
  • Cost optimization tools

Challenges and Opportunities

Technical Challenges

Reliability: Ensuring consistent AI behavior Scalability: Managing computational costs Integration: Combining different paradigms Security: Protecting against AI-specific threats

Business Opportunities

New Markets: AI-native applications Efficiency Gains: Automated development processes Innovation: Novel problem-solving approaches Democratization: Broader access to technology

Societal Implications

Education: Need for new curricula Employment: Changing skill requirements Ethics: Responsible AI development Governance: Regulatory frameworks


Conclusion: Embracing the Transformation

Andrej Karpathy’s vision of software evolution represents a fundamental shift in how we think about programming and application development. The coexistence of three programming paradigms—traditional code, neural networks, and natural language—creates unprecedented opportunities and challenges.

The key insight is that this isn’t about replacement but about augmentation and collaboration. Traditional programming skills remain valuable while new capabilities emerge. The most successful developers and organizations will be those who can effectively combine all three paradigms.

As we build the infrastructure for AI agents and develop partially autonomous applications, we’re not just creating new tools—we’re reshaping the entire software landscape. The future belongs to those who can navigate this complexity while maintaining focus on human needs and values.

The era of programming in English has begun, but it’s not the end of programming as we know it—it’s the beginning of a new chapter where technology becomes more accessible, more powerful, and more aligned with human communication patterns.


Ready to Build the Future of Private AI Cloud?

If you resonate with Thaki Cloud’s mission—”AI Compute & Software for Everyone”—please contact us via email right away. 📧 info@thakicloud.co.kr