Andrej Karpathy Says He’s Surrendered to AI Coding – A Workflow Revolution

Andrej Karpathy posted a long tweet describing his recent weeks using Claude Code for programming, declaring that a turning point has arrived.

Workflow transformation

From November to December he says his coding workflow changed fundamentally: the ratio moved from 80% manual + 20% AI assistance to 80% AI‑generated code + 20% manual edits.

He admits he now programs mostly by describing in English what he wants the LLM to write, which feels a bit embarrassing but is extremely useful for handling large code blocks.

AI’s new flaws

AI no longer makes simple spelling mistakes; its errors are now logical. It often makes unchecked assumptions, proceeds without asking for clarification, and fails to point out contradictions. It also tends to over‑engineer solutions, producing deeply nested, redundant code and rarely cleans up unused fragments. When asked to simplify, it can shrink a 1,000‑line implementation to about 100 lines.

Never‑tiring resilience

AI never gets tired or discouraged and will keep trying until a problem is solved, whereas humans often give up. Watching an LLM wrestle with a problem for 30 minutes and finally succeed gives a sense that AGI is approaching, highlighting physical stamina as a core bottleneck that LLMs eliminate.

Expansion more important than speed

Measuring pure speed gains is difficult, but Karpathy feels the impact is broader: AI enables him to write small utilities that previously weren’t worth the effort and to tackle code he previously couldn’t understand because of missing technical knowledge.

Write previously uneconomical small tools.

Handle code that was out of reach due to lack of expertise.

Practical usage tips

LLMs excel at repeated attempts toward a goal. Instead of telling the model exactly how to do something, give it a target criterion and let it figure out the steps. Have it generate test cases first, then the functional code. Start with a simple version that is guaranteed to run, then ask for optimization. Shift the prompt from "do this" to "what outcome should be achieved".

Unexpected enjoyment

Karpathy finds AI‑assisted programming more fun because repetitive grunt work disappears, leaving the creative part. He gets less stuck and always finds a way to make progress together with the AI.

Divergent programmer perspectives

He notes a possible split: programmers who enjoy the act of coding versus those who care mainly about delivering results.

Skill degradation

He observes his own hand‑coding ability declining, though his ability to read and review code remains solid, since reading engages different brain functions than writing.

2026 forecast

Karpathy predicts that by 2026 AI‑generated low‑quality content will flood code repositories, articles, papers, and social media, accompanied by more AI productivity showcases.

Questions to ponder

What happens to “10‑x programmers” when the efficiency gap widens?

Will AI‑augmented generalists surpass specialists, given AI’s strength in filling details?

Will future AI programming feel like a real‑time strategy game, a factory‑building simulation, or a musical performance?

How many jobs are blocked by the “need to understand technology” requirement?

Turning point

Karpathy concludes that AI coding tools such as Claude and Codex crossed a threshold in December of the previous year, triggering a fundamental shift in software development. AI’s intelligence level suddenly outpaced everything else, reshaping tool integration, workflows, and broader adoption. The industry will need to adapt to this new capability.