Why Even Top AI Leaders Feel Outpaced: The Rise of AI‑Native Programming

On December 27, OpenAI co‑founder and former Tesla AI director Andrej Karpathy posted on social media that, as a programmer, he has never felt more behind, noting that the "bits" contributed by human developers are becoming increasingly sparse.

He identifies a new "programmable abstraction layer" that includes Agents, Subagents, Prompts, Contexts, Memory, Tools, MCP (Model Context Protocol), LSP, and others, describing the situation as receiving a powerful alien tool without a manual. The tool is stochastic, fallible, and sometimes unintelligible, yet it is causing a "Liskov‑level 9" quake in the software industry.

Agents, Subagents, Prompts, Contexts, Memory, Tools, MCP, LSP…

Karpathy’s tweet prompted a vivid response from fellow technologist Boris Cherny, author of *Programming TypeScript*. Cherny shared a concrete debugging case where Claude Code fixed a memory‑leak issue.

Old‑fashioned way: Connect a profiler, run the app, pause, manually inspect heap allocations, and hunt down lingering objects—a skill set accumulated over a decade.

AI‑native way: Ask Claude to produce a heap dump, read it, and identify stray objects. Claude resolved the issue in a single step and submitted a pull request.

According to Cherny, such AI‑driven fixes now occur almost weekly. He highlights a counter‑intuitive phenomenon: recent graduates often use these tools more effectively than senior engineers because they lack "legacy memories" that constrain their mental models of what AI can or cannot do.

Senior engineers still habitually open IDEs, configure environments, and set breakpoints, whereas newcomers have learned to describe problems to AI and let the model solve them. Experience, once the most valuable asset, can become a limiting chain.

Cherny revealed that during a month in which he did not open an IDE at all, the Opus 4.5 model wrote roughly 200 pull requests for him, with every line of code generated by AI.

This shift suggests that software engineering is moving from "writing code" to "reviewing code" and "orchestrating intent". As Karpathy put it, the tool may misfire like a small projectile, but when used correctly it can "spray a powerful laser that instantly melts your problem".

Both Karpathy and Cherny advocate a stance of "re‑adjust" rather than resignation, urging engineers to "roll up their sleeves and not fall behind".