AI Thinkers·AI researcher, educator, founder
Andrej Karpathy
Software is undergoing a third fundamental shift. Software 1.0: humans write explicit instructions in code. Software 2.0: humans write datasets and let gradient descent find the weights. Software 3.0: programs are written in English and executed by large language models. LLMs are not tools — they are a new kind of computer whose "CPU" is the transformer, whose "RAM" is the context window, and whose instruction set is natural language. The next decade is the decade of agents, but genuine AGI remains 10+ years away, and the path runs through careful orchestration, not magic.
Key positions
Neural networks are a new programming paradigm — gradient descent finds the program, humans curate the data.
His 2017 "Software 2.0" essay argued the most valuable programs are no longer written by humans specifying logic but by optimisers searching program space via backpropagation. Software 3.0 extends the arc: prompts in English are source code, LLMs are the CPUs. The "programmer" role shifts from writing if-statements to curating training data and intent.
https://karpathy.medium.com/software-2-0-a64152b37c35 (Nov 2017); YC AI Startup School keynote, Jun 2025
LLMs should be understood as a new operating system — not a chatbot interface.
LLM = CPU, tokens = bytes, context window = RAM, retrieval = file system, agents = long-running processes, tool calls = system calls. This framing implies LLMs should be understood through their computational constraints as architectural properties, not quirks to work around with prompts. It also implies they are multi-tenant substrates, not single-purpose assistants.
X post @karpathy, Sep 2023 (x.com/karpathy/status/1707437820045062561)
Vibe coding is valid for prototyping, risky in production — the risk depends on whether you can verify the output.
Karpathy coined "vibe coding" in Feb 2025: "fully give in to the vibes, embrace exponentials, and forget that the code even exists." He presented it as his mode for personal weekend projects. The tension: he also teaches neural networks from first principles (Zero to Hero) and advocates knowing what your code does professionally. Skilled practitioners who can catch failures benefit most; the dangerous middle ground is semi-competent practitioners who trust outputs they cannot evaluate.
X post @karpathy, Feb 6 2025 (x.com/karpathy/status/1886192184808149383)
2025–2035 is the decade of agents — but the right product shape is an "Iron Man suit," not an army of robots.
Karpathy explicitly pushed back on annual "year of agents" claims. He advocates for "partial autonomy products with custom interfaces that make the human-AI collaboration loop fast and effective" and an "autonomy slider" calibrated to the cost of failure. Fully autonomous agents that never fail are not ready for most production use cases.
YC AI Startup School keynote, Jun 2025; Dwarkesh Patel podcast, Oct 2025
Reinforcement learning extracts supervision through a very thin straw.
"You're sucking supervision through a straw, because you've done all this work that could be minutes of rollout and you're sucking the bits of supervision of the final reward signal through a straw, and you're broadcasting that across the entire trajectory." A long agentic rollout receives one binary signal at the end, broadcast across thousands of steps. He proposes system prompt learning from iterative interaction as a more scalable path.
Dwarkesh Patel podcast, Oct 2025 (dwarkesh.com/p/andrej-karpathy)
There is a dangerous AI literacy divide widening — power users and casual users have incompatible mental models of what AI can do.
AI capability gains are concentrated in technical domains — coding, maths, research — that casual users rarely probe. This creates both overblown dismissals and overblown evangelism. Tooling optimised for expert workflows widens rather than closes the gap. Non-experts consuming AI outputs without the context to evaluate them has real governance implications.
Multiple X threads, @karpathy, 2025
In their own words
“There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
“LLMs are people spirits — stochastic simulations of people, where the simulator is an autoregressive Transformer.”
“You're sucking supervision through a straw... you're broadcasting that across the entire trajectory.”
“It's less about building an army of autonomous Iron Man robots and more about building the Iron Man suit.”
Tensions with other thinkers
Where this view genuinely conflicts with others in the field.
Karpathy says 10+ years; Altman and Amodei lean mid-to-late 2020s.
LeCun considers LLMs largely obsolete within five years; Karpathy sees them as the foundation for the coming decade.
Karpathy advocates building foundational understanding (Zero to Hero); Cherny says fundamentals will stop mattering "in a year or two."
His "sucking through a straw" critique challenges the dominant narrative that RL is the primary path to superhuman capabilities.