Andrej Karpathy: Software Is Changing (Again)

Andrej Karpathy argues LLMs constitute ‘Software 3.0’ — a third programming paradigm where English is the programming language and we are in computing’s 1960s.

• LLMs are best understood as a new OS: context window = RAM, orchestrating memory and compute, with closed-source (OpenAI/Anthropic) and open-source (Llama/Linux) parallels.
• We are in the 1960s of LLM computing: centralized, expensive, time-shared — personal LLM computing has not happened yet.
• Unlike every prior transformative technology (electricity, GPS, computing), LLMs diffused to consumers first; governments and corporations are lagging adoption.
• Karpathy warns against full-autonomy agent hype: a 10,000-line diff is useless if the human is the verification bottleneck; keep AI on a short leash with small incremental chunks.
• Self-driving analogy: a perfect 30-minute Waymo demo in 2013 felt like imminent victory; 12 years later autonomy is still unsolved — 2025 is not the year of agents, it is the decade.
• Vibe-coding his MenuGen app took hours; making it production-ready (auth, payments, deployment) took a week of manual browser clicks — documentation and DevOps tooling must be rebuilt for agents.
• Vercel and Stripe are early movers offering docs in Markdown and replacing every ‘click’ instruction with an equivalent curl command so LLM agents can act without a GUI.
• LLMs have jagged intelligence: superhuman recall but will insist 9.11 > 9.9 and claim ‘strawberry’ has two R’s — emergent psychology with real cognitive deficits requires human audit loops.

2025-06-19 · Watch on YouTube