Introduction and Context

• Former Tesla AI director Andrej Karpathy speaking to students entering industry
• Describes current time as "extremely unique and very interesting" for industry entry
• Software fundamentally changing for first time in 70 years, changed twice rapidly in recent years
• Massive opportunity: "huge amount of work to do, huge amount of software to write and rewrite"

Evolution of Software Paradigms

• Software 1.0: Traditional code written for computers
• Software 2.0: Neural network weights (introduced concept few years ago)
  • Parameters created by optimizer, not directly written
  • HuggingFace equivalent of GitHub for Software 2.0
  • Model Atlas for visualization
  • Flux image generator example (giant circle in visualization)
• Software 3.0: Programmable neural networks via LLMs
  • Prompts are programs written in English
  • "Remarkably we're now programming computers in English" (pinned tweet)
  • GitHub code now includes English interspersed with code

Tesla Autopilot Case Study

• Observed Software 2.0 "eating through" Software 1.0 stack
• Neural networks grew in capability/size while C++ code was deleted
• Functionality migrated from 1.0 to 2.0 (image stitching, temporal processing)
• Same pattern happening again with Software 3.0

LLMs as Operating Systems

• Utility analogy: LLM labs spend capex to train, opex to serve via APIs
  • Metered access (pay per million tokens)
  • Demand low latency, high uptime, consistent quality
  • Open Router as "transfer switch" between LLM providers
  • "Intelligence brownout" when state-of-the-art LLMs go down
• Fab analogy: Large capex requirements, deep tech trees, R&D secrets centralizing
• Operating system analogy (strongest):
  • LLM as CPU equivalent, context windows as memory
  • Orchestrates memory and compute for problem solving
  • Apps run on multiple LLM providers (like Windows/Linux/Mac)
  • Currently 1960s era - expensive, centralized, time-sharing model
  • Personal computing revolution hasn't happened yet (Mac minis early indication)

LLM Psychology and Limitations

• Capabilities:
  • Encyclopedic knowledge and memory (like Rainman's Dustin Hoffman)
  • Can remember SHA hashes, vast amounts of information
• Deficits:
  • Hallucination, poor self-knowledge
  • Jagged intelligence (superhuman in some areas, basic mistakes in others)
  • Examples: insist 9.11 > 9.9, count Rs in "strawberry"
  • Anterograde amnesia (don't learn/consolidate knowledge over time)
  • Context windows are working memory, not long-term learning
  • Security vulnerabilities (gullible, prompt injection, data leakage)
• Movie references: Memento, 50 First Dates for memory issues

Partial Autonomy Applications

• Cursor example:
  • Traditional interface + LLM integration
  • Context management, multiple LLM orchestration
  • Application-specific GUI for auditing
  • Autonomy slider: tab completion → command K → command L → command I (full repo)
• Perplexity example:
  • Packages information, orchestrates multiple LLMs
  • GUI for source citation auditing
  • Autonomy slider: quick search → research → deep research

Human-AI Cooperation Principles

• Generation-verification loop: AI generates, humans verify
• Speed up verification:
  • GUIs utilize human computer vision "GPU"
  • Visual representations more efficient than text reading
• Keep AI on leash:
  • Avoid 10,000-line diffs (human still bottleneck)
  • Work in small incremental chunks
  • Concrete prompts increase verification success probability
• Best practices: Vague prompts → verification failure → spinning

Vibe Coding and Democratization

• "Vibe coding" tweet became major meme, Wikipedia page created
• Everyone becomes programmer through natural language
• Kids vibe coding video - "gateway drug to software development"
• Examples: Built iOS app without Swift knowledge, Menu Genen app (menu.app)
• Code creation easy, DevOps/deployment still difficult
• Authentication/payments setup took week vs hours for core functionality

Building for Agents

• New category: agents as "people spirits" - computers but humanlike
• Infrastructure adaptations:
  • llm.txt files (like robots.txt) for LLM instructions
  • Documentation transitioning from human-readable to LLM-readable
  • Vercel/Stripe offering markdown documentation
  • Replace "click" instructions with curl commands
• Tools for LLM accessibility:
  • Change GitHub URLs to get-ingest for file concatenation
  • Deep Wiki for Devon-analyzed repo documentation
  • Model Context Protocol from Anthropic

Future Outlook

• Analogies: 1960s computing era, Iron Man suit (augmentation + autonomy)
• Recommendations:
  • Build Iron Man suits, not flashy autonomous demos
  • Focus on partial autonomy products with custom GUI/UX
  • Fast generation-verification loops
  • Autonomy slider in products
• Timeline: Decade of agents, not year of agents (referencing 12 years since perfect Waymo demo)
• Opportunities: Rewrite massive amounts of code, infrastructure adaptation for agents