skip to content
reelikklemind

🎥 Software in the era of AI 🧠 by Andrej Karpathy


🎥 Software in the era of AI 🧠

Summary of Andrej Karpathy’s lecture.

💡Main Thesis

We’re not in the “year of agents”, we’re in the decade of agents.

Karpathy walks us through a historical arc showing how software is evolving and why we’re at the beginning of a major transformation.


📜The Software Evolution Story

Software is transitioning across three paradigms:

  1. Software 1.0 – 👨‍💻 Code programs the computer
  2. Software 2.0 – 🧠 Weights program the neural net
  3. Software 3.0 – ✍️ Prompts program the LLM

We’re now literally programming in English.

The same shift that replaced Tesla’s C++ Autopilot stack is happening everywhere.

➡️ Software 3.0 will begin to absorb existing codebases, but engineers must still be fluent in all three paradigms.


🕰️Where We Are in History

Karpathy compares our moment to the 1950s–70s timesharing era. Like early computers, today’s models are centralized:

  • 💰 Expensive, centralized compute
  • ☁️ Model (OS) runs in the cloud
  • 🌐 I/O is streamed over the network
  • ⏱️ Compute batched across users

🖥️ Chat is the terminal.

🖱️ The GUI hasn’t been invented yet.


🔄The Triple Nature of LLMs

LLMs are strange hybrids: they simultaneously resemble:

🔌 Utilities

  • Massive CAPEX to train
  • Ongoing OPEX to serve intelligence
  • Metered access ($ per 1M tokens)

🏭 Fabs

  • R&D-intensive
  • Secretive IP
  • Some companies are fabless (use NVIDIA), others own the fab (e.g., Google, Amazon)

💻 Operating Systems

  • Growing software ecosystems
  • Switching friction (different features)
  • Distinct system/user space layers


🧠LLMs are Savants, with Cognitive Faults

LLMs = “Stochastic simulations of people” 🤯

They’re superhuman in bursts, but still very flawed:

⚠️ Hallucinate

⚠️ Jagged intelligence (e.g., thinks 9.11 > 9.9)

⚠️ No long-term memory (anterograde amnesia)

⚠️ Vulnerable to prompt injection

They’re like fallible “people spirits” we’re learning to collaborate with.

🛠️


What Actually Wins?

🏆 Apps with partial autonomy: the “Cursor for X” model.

What makes Cursor powerful is its tight Generation ↔ Verification loop, supported by:

  • 🖱️ Manual UI for human control
  • 🤖 Context packaging by AI
  • 👀 GUI for visual diff & audit
  • 🎚️ Adjustable autonomy:

Tab → Cmd+K → Cmd+L → full background agent mode


🦾The Iron Man Analogy

The Iron Man suit = Agent + Augmentation.

It can act autonomously but is most effective when paired with Tony Stark.

➡️ The future lies in tight human–AI feedback loops, where autonomy is adjustable, and humans stay in control while AI augments their abilities.



Crepi il lupo! 🐺