Between 2012 and 2026, Karpathy's public comments moved from warning how far machines were from real understanding to detailing what coding agents can already do under close supervision. The nine quotes below trace that arc, from his 2012 "really, really far away" post to his 2026 work directing agents on the No Priors podcast.
1. Why Does Andrej Karpathy Say English Is the New Programming Language?
"The hottest new programming language is English."Andrej Karpathy, Xataka En, June 2025
For Karpathy, English is taking over the role syntax used to play. Xataka En cited the line in June 2025 alongside tools like Cursor, Bolt, Replit, and V0, two years after ChatGPT made plain-language coding mainstream. The shift opens a real question for anyone outside engineering: if plain instructions produce working drafts, who is now eligible to build software?
2. Why Did Andrej Karpathy Once Say AI Was Still Very Far Away?
"The state of Computer Vision and AI: we are really, really far away."Andrej Karpathy, "The state of Computer Vision and AI: we are really, really far away." October 2012
In the post, Karpathy walked through a single photograph to show how much physics, social inference, and object knowledge people bring to one glance, while benchmark scores were still nowhere close. The argument carries forward into his later optimism about coding agents, which he treats as tools that need close human review on anything load-bearing.
3. What Does Andrej Karpathy Mean by Agentic Engineering?
"claim the leverage from the use of agents but without any compromise on the quality of the software."Andrej Karpathy, Observer, February 2026
Agentic engineering, in Karpathy's framing, is what separates production software from casual AI demos. Observer reported the phrase in February 2026, a year after vibe coding had spread from a tweet into investor pitch decks. His point is not that engineering matters less, but that the work has moved up the stack: review, orchestration, and QA rather than hand-typing each function.
4. What Is Andrej Karpathy's Definition of Vibe Coding?
"where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."Andrej Karpathy, TechRadar, September 2025
Vibe coding, in his telling, is the model writing the code while the human steers by feel and quick checks. TechRadar quoted the phrase in September 2025, after the term spread from a single tweet into enterprise discussion. Speed is the upside; the cost is that someone still has to decide whether the software actually works, which gets harder once security and compliance teams are in the room.
5. Why Does Andrej Karpathy Prefer Iron Man Suits to Iron Man Robots?
"it’s less Iron Man robots and more Iron Man suits that you want to build."Andrej Karpathy, AI Startup School transcript, June 2025
The Iron Man suit, in his June 2025 AI Startup School keynote, is Karpathy's case for augmentation: today's models are strong enough to assist but too unreliable to run unsupervised. The framing overlaps with Mustafa Suleyman's argument for AI containment, but Karpathy's emphasis stays on shipping useful tools rather than guarding against worst cases.
6. Why Does Andrej Karpathy Say Autocomplete Is Still AI's Sweet Spot?
"For now, autocomplete is my sweet spot."Andrej Karpathy, Dwarkesh Podcast, October 2025
Autocomplete keeps the human close to the work while the model supplies short bursts of acceleration. On the Dwarkesh Podcast in October 2025, Karpathy contrasted that with looping agents that drift, misread repo context, and confidently return bad code on longer tasks. Yann LeCun's case against LLMs lands in similar territory: fluency outruns dependable reasoning.
7. What Is Andrej Karpathy Trying to Build With AI Education?
"The goal here is to spark those moments in people's minds."Andrej Karpathy, VentureBeat, July 2024
With Eureka Labs, Karpathy wants to recreate the best part of one-to-one teaching: patient explanation at scale. VentureBeat quoted the line in July 2024, when he launched the company after leaving OpenAI for the second time. The pitch is not to replace teachers but to let one expert design a course that reaches far more students than any single instructor could teach in person.
8. What Does Andrej Karpathy's $100 Chatbot Say About Cheap AI?
"The best ChatGPT that $100 can buy."Andrej Karpathy, GitHub repository, October 2025
With nanochat, Karpathy argues useful AI does not always need a frontier-scale budget. He published the GitHub repository in October 2025 as a minimal training harness one builder could reproduce on ordinary hardware. The wager is the opposite of Demis Hassabis at DeepMind, who keeps the spotlight on the largest systems and what they can unlock for science.
9. How Does Andrej Karpathy Describe His Own Work With AI Agents?
"But I have to express my will to my agents for 16 hours a day."Andrej Karpathy, No Priors podcast transcript, March 2026
On the No Priors podcast in March 2026, Karpathy said he spends sixteen hours a day expressing his will to coding agents, not typing code. The job has become coordination: research runs and branches he steers in parallel, with the keyboard mostly idle. The English-as-programming thesis from 2023 had turned, by 2026, into a daily management problem.
What to Read Next
Karpathy is easiest to place alongside people arguing about whether today's models should be treated as copilots, agents, or systems that still need firmer limits. The links below show where his software-first view overlaps with other AI leaders and where it breaks from them.
- Ilya Sutskever quotes on AI: another OpenAI co-founder, but with more attention on research bottlenecks, pre-training limits, superintelligence, and what happens after internet-scale data runs thin
- Yann LeCun quotes on LLM limits: a sharper case that language models still lack the memory, planning, and world models people project onto them in coding, robotics, and daily use
- Demis Hassabis quotes on AGI: compare Karpathy's software-first framing with DeepMind's science-first view of general intelligence, automation, and AI as a discovery engine
- Mustafa Suleyman quotes on AI safety: the more cautious argument for containment, guardrails, and limits on what autonomous systems should do in public life, work, and state power
- All AI quotes: the full Aiifi collection of AI leader perspectives across work, automation, risk, and human-AI collaboration
