🧵 @karpathy dropped the most compelling lecture at the AI startup school :) Here's my notes summarized with Opus. Main thesis: We're not in the "year of agents" — we're in the DECADE of agents. Here's the historical arc that explains why... ==== The Software Evolution Story ==== First, the foundational shift happening right now. Software is evolving through three distinct paradigms: • Software 1.0: Code programs the Computer • Software 2.0: Weights program the Neural Net • Software 3.0: Prompts program the LLM We're literally programming in English now. The same pattern that ate through Tesla's C++ autopilot stack is happening everywhere — Software 3.0 will start consuming existing codebases, but engineers will need to be fluent in all three paradigms. ==== Where We Are Historically ==== But where are we in this evolution? Karpathy's answer: the 1950s-70s timesharing era. Like computers in the past, models are centralized today: • Centralized expensive compute • OS (model) runs in the cloud • I/O streamed over network • Compute batched over users Chat is just the terminal — direct access to the operating system. The GUI hasn't been invented yet. ==== The Triple Nature of LLMs ==== LLMs are weird because they have properties of utilities, fabs, AND operating systems simultaneously: 🔌 Like utilities: CAPEX to train (equivalent to building the electrical grid), OPEX to serve intelligence, metered access at $/1M tokens 🏭 Like fabs: Massive R&D investment, deep secrets, with some companies being fabless (using NVIDIA chips) while others own their fab (Google/Amazon with custom GPUs) 💻 Like operating systems: Increasingly complex software ecosystems, switching friction due to different features, clear system/user space boundaries ==== LLMs are Savants with Cognitive Faults ==== LLMs are "stochastic simulations of people". They are superhuman in some areas, but still struggle will • Hallucination • Jagged intelligence (superhuman in some areas but make basic mistakes like thinking 9.11 > 9.9) • Suffer from anterograde amnesia with no long-term memory or continuous learning • Remain gullible to prompt injection They're essentially fallible "people spirits" we're learning to work with. ==== What Actually Wins ==== So what wins in this landscape? Partial autonomy apps — the "Cursor for X" pattern. Cursor works because it nails the Generation ↔ Verification loop with four key elements: • Manual UI where humans can work directly • AI that packages context automatically • App-specific GUI (like diffs) for easy human audit • Slider to choose autonomy level: Tab → Cmd+K → Cmd+L → background agent mode ==== Iron Man Analogy ==== The Iron Man suit is both an augmentation and an agent. It can run autonomously, but it's most powerful when it combines with Tony Stark. Partial autonomy that makes the Generation ↔ Verification loop incredibly tight, keeping humans firmly in control while gradually expanding AI capabilities.