Just now! OpenAI co-founder Karpathy erupts: "Proxy engineering" boosts programmers' efficiency by over 10 times, leaving human's final barrier only in aesthetics? 99% of code will be rewritten!

On April 29th, a prominent figure in the AI field, Andrej Karpathy, spoke candidly at a closed-door event. This guy was the one who built Tesla Autopilot from scratch and holds a significant position at OpenAI. He clearly stated that since December last year, workflows centered around AI agents have become truly usable — this is not just a concept, but a substantial technological leap. He mentioned that many people’s impression of AI still lingers on ChatGPT, but you must reassess — a fundamental change has already occurred.

Karpathy introduced a new concept: “agentic engineering,” to distinguish it from the “vibe coding” he named last year. The latter raises the baseline for all developers, while the former is a continuation and acceleration of professional software quality standards. He straightforwardly said that a large portion of existing code and applications “should not exist” under this new paradigm. Currently, most organizations’ recruitment processes, development tools, and infrastructure are still designed for humans, not for agents.

The underlying computing architecture is undergoing a power transfer. December last year marked a critical turning point, and Karpathy admitted that he was deeply shocked by the latest AI models — system-generated code blocks are becoming increasingly perfect, and he can no longer remember the last time he manually modified code. He trusts the system more and more, yet has never felt so behind as a programmer. This is not optimization; it’s a paradigm shift.

We are bidding farewell to “Software 1.0 (writing code)” and “Software 2.0 (organizing datasets to train neural networks),” and are officially entering “Software 3.0.” In this new era, large language models themselves are a new type of computer. Programming has become writing prompts, and the context window is the lever you use to control this interpreter. Karpathy predicts that in the future, neural networks will become the main process, with CPUs turning into co-processors. Neural networks will handle most of the heavy lifting. This means the strategic importance of “intelligent computing power” will be further solidified.

Everything must be rewritten. Currently, the frameworks, library documentation, and descriptions on the internet are still “written for humans,” which frustrates Karpathy immensely. He said, “Why do I still need someone to tell me what to do? I don’t want to do anything. What text should I copy and paste to my AI agent?” The big opportunity ahead lies in building infrastructure that prioritizes “agents.” Systems will be broken down into sensors that perceive the world and actuators that transform it. Data structures should be designed to be highly readable by large language models, and machine agents will represent individuals and organizations interacting in the cloud.

In this highly automated future, human core scarcity will revert to aesthetics, judgment, and the deepest business understanding. Karpathy quoted a phrase he keeps chewing on: “You can outsource your thinking, but you cannot outsource your understanding.”

Regarding productivity explosion, Karpathy clearly distinguished between “vibe coding” and “agentic engineering.” Vibe coding raises the lower limit, while agentic engineering maintains the upper limit of professional software quality. Agentic engineering is not just about speeding up; it requires developers to coordinate those “somewhat error-prone, stochastic but extremely powerful” AI agents to move forward at full speed without sacrificing quality. People used to talk about “10x engineers,” but 10x is no longer enough to describe the speedup you gain. In his view, those who perform well in this field have output peaks far exceeding 10x.

Faced with this productivity surge, organizational structures and talent evaluation logic must be reconstructed. He recommends that companies abandon traditional algorithmic problem-solving interviews and instead assess how candidates utilize multiple AI agents to collaboratively build large projects, and their ability to defend against attacks from other AI agents.

For entrepreneurs and investors, Karpathy offers a highly practical evaluation framework: verifiability. Currently, AI capabilities exhibit an extremely strange “sawtooth” pattern — the most advanced models can reconstruct 100k lines of code or find zero-day vulnerabilities, yet they tell you to walk 50 meters to a car wash. This is crazy. The reason for this disconnect is that leading labs have poured massive reinforcement learning resources into fields like “mathematics” and “code,” where results are easier to verify. Therefore, as long as you are in a result-verifiable business scenario, AI can exert tremendous power. There are still many high-value, yet underexplored, verifiable reinforcement learning environments that top labs have not focused on. This is a huge blue ocean for startups to fine-tune and commercialize.