🧠 Andrej Karpathy shared his thoughts on coding with AI after several weeks of intense work with Claude Code.
Main observations:
The transition happened over weeks. From November to December, the ratio changed from 80% manual code / 20% agents to the opposite. "The biggest change in my workflow in 20 years of programming."
Agents still make mistakes. But the errors have changed - they are no longer syntax errors, but conceptual ones, like a hasty junior. Models make assumptions instead of clarifying questions, do not show trade-offs, complicate code unnecessarily. They might write 1000 lines where 100 would suffice.
Unyieldingness. "Feel the AGI moment" - watching an agent struggle with a problem for 30 minutes and succeed where a human would have long given up. Endurance is a bottleneck in work, and LLMs dramatically expand it.
Not acceleration, but expansion. The main effect is not doing the same thing faster, but doing much more - code that previously was not worth the effort, and code that could not be written before.
Leverage. Don’t tell it what to do - give success criteria. Let it write tests first, then pass them. Connect it to the browser via MCP.
Programming has become more fun. The routine is gone, leaving the creative part. But the skill of writing code manually is starting to atrophy.
The year 2026 will be the year of "slopacolypse" - slop everywhere: on GitHub, arXiv, social media. And the year when the industry will digest new opportunities.