Overview
Andrej Karpathy observes that programming has fundamentally changed in just the past two months, specifically December, due to AI coding agents that previously didn’t work but now do. This represents a sudden shift rather than gradual progress, with models showing significantly higher quality, long-term coherence and ability to handle complex tasks.
Key Arguments
- AI coding agents have crossed a critical threshold in December, transitioning from basically non-functional to basically functional: Karpathy emphasizes this is not gradual progress but a specific inflection point where coding agents that previously didn’t work now do work effectively
- The improvement represents a qualitative leap in AI capabilities, not just incremental progress: Models now demonstrate significantly higher quality, long-term coherence, and tenacity - they can power through large and complex programming tasks that were previously beyond their capabilities
- This change is extremely disruptive to traditional programming workflows: The breakthrough is substantial enough to fundamentally alter how programmers work, moving beyond the typical ‘progress as usual’ pattern to something transformative
Implications
This observation suggests we may be witnessing a pivotal moment in software development where AI becomes a genuinely capable programming partner rather than just a helpful tool. For developers, this could mean reimagining their entire workflow and skillset requirements in the near term, not over years of gradual change.
Counterpoints
- Karpathy mentions there are ‘a number of asterisks’ to his observation: This suggests there are still limitations or caveats to the AI coding capabilities that weren’t detailed in this brief quote
- The assessment may be premature given the recency of the changes: Since the changes occurred just in December, it may be too early to fully evaluate their long-term impact and sustainability