AI Circle Level 9 Earthquake! OpenAI Founder Karpathy Breaks Down Late at Night: I'm Falling Behind

Former OpenAI founder Andrej Karpathy posted on X: “As a programmer, I’ve never felt so behind.” He admits that new technologies over the past year can boost capabilities by 10 times, but he feels the gap for not keeping up. Boris Cherny, founder of Claude Code, provides an even more shocking case: completing 259 PRs, 80,000 lines of code in 30 days, all generated by AI.

Karpathy’s Attitude Turns 180 Degrees in Two Months

Karpathy’s tweet has caused a seismic shock in the tech community, not only because of his honesty but also because of his dramatic change in stance. Just two months ago, Karpathy mocked these AI models as “slop” in a podcast, believing they were still ten years away from real application. At that time, he was highly critical of AI assistants, questioning their practicality and reliability, showing a cautious attitude typical of a seasoned tech expert towards new technology.

But when he saw someone praising Opus 4.5’s performance, he directly added: “Not just good. If you haven’t kept up in the past 30 days, your worldview is already outdated.” From “laughing at” to “worldview outdated,” it only took two months. This rapid attitude shift itself illustrates the terrifying pace of AI evolution.

This change is not just personal emotion but a rational response to crossing a technological threshold. Opus 4.5 (Claude 3.5 Opus) has achieved a qualitative leap in reasoning ability, code generation quality, and contextual understanding compared to its predecessor. More importantly, the maturity of the Agent framework makes AI no longer just a Q&A tool but an autonomous system capable of planning, executing, and debugging itself. When experts like Karpathy realize the power of this combination, anxiety is real.

Developer Adam Wathan’s comment is even more brutal: “Honestly, if you don’t feel this anxiety, you’re truly done.” This highlights the essence of the current AI revolution: it’s not an optional upgrade but a forced paradigm shift. Those who still feel comfortable and believe AI won’t impact their work may already be unknowingly left behind by the times.

Industry analyst Aakash Gupta calls this a “Level 9 earthquake,” emphasizing that it’s not coming soon but has already happened, and the aftershocks will become the norm. This description accurately captures the current state: the big change has been completed, now is the adaptation period, and the speed of adaptation determines life or death.

259 PRs, Zero Hand-Written: Programmers Become Intelligent Operators

The most shocking proof comes from Boris Cherny, founder of Claude Code. He shared backend data showing that in the past 30 days, he completed 259 merge requests (PRs), submitted 497 pieces of code, involving nearly 80,000 lines of code added and deleted. The key point: not a single line was written by him; all generated by AI running on Opus 4.5.

This is no longer traditional “AI-assisted programming” but a true agent workflow. Boris’s dashboard shows a total consumption of 325 million tokens, with the longest session of AI running continuously for 1 day, 18 hours, and 50 minutes. This means he issues complex commands to AI, then goes to sleep or attends meetings, while AI autonomously debugs and modifies in the background until the task is completed. Boris straightforwardly states: “Code is no longer the bottleneck; execution and direction are.”

This working mode completely overturns the role of programmers. Karpathy makes a precise analogy: current AI tools are like “alien tools without manuals” suddenly shoved into your hands. You know they are insanely powerful, but don’t know which button to press because they not only lack manuals but are constantly changing themselves.

Previously, coding was about building with blocks, learning algorithms, data structures, with each line of code producing a predictable result. But now, programming is more like “taming beasts,” facing a probabilistic system—a “random, error-prone, incompletely understandable” black box. As Twitter user @SightBringer said, in the past your skill was “writing correct instructions,” but now it’s “orchestrating intelligence.” You are no longer the creator of code but a regulator of behavior.

This explains why tech professionals feel out of control. Because you not only need to understand code but also prompt engineering, agents, context, and how to connect these chaotic elements with hooks and plugins. This is not just an upgrade; it’s a change of track.

Three Core Capabilities of Intelligent Operators

Problem translation ability: Can you turn vague requirements into precise instructions that AI can understand? This requires deep understanding of business logic and prompt engineering skills.

Risk detection and debugging intuition: AI will hallucinate and make errors. Can you spot where it’s talking nonsense at a glance? This requires solid technical fundamentals and systemic thinking.

System integration and command ability: Can you coordinate multiple AI agents to work together on complex tasks? This demands architecture design skills and workflow optimization experience.

Foundational Engineering Skills as the Lifeline of the New Era

Google engineer Addy Osmani offers counterintuitive advice: testing, documentation, CI/CD—these “boring” foundational engineering skills are now your lifeline. Many mistakenly think that AI can write code, so these basics are no longer important, but exactly the opposite: AI produces code at such a fast pace that it can even “create chaos.”

You need extremely solid testing systems and engineering intuition to tame this wild horse, turning “random generation” into “stable output.” When AI generates thousands of lines of code in a day, without a comprehensive automated testing framework, you cannot verify correctness. When AI modifies a dozen files simultaneously, without good version control and CI/CD processes, the entire project could fall into chaos.

This also explains why middle-aged engineers might benefit from this wave of change. Having experienced the wilderness era without frameworks or tools, with a deep understanding of software engineering fundamentals, these seemingly “outdated” experiences become scarce skills in the AI era. Younger developers may be more adept at using the latest AI tools, but middle-aged engineers know how to establish order amid chaos.

Dr. Jim Fan’s prophecy is coming true: in 2024, AI will be your co-pilot; after 2025, humans will be the co-pilot. We must learn to let go, to become the ones issuing commands and hitting the brakes. Karpathy’s only advice: “Roll up your sleeves to avoid falling behind.” Anxiety produces no results; only by forcing yourself to use AI and practicing in real projects can you develop muscle memory.

Investor Anthony Pompliano’s summary may be the most accurate: “The world’s top engineers are telling you, a new game has begun. The gates are open, anyone can enter.” When everyone is pushed back to the starting line, why can’t you be the fastest?