AI A god-tier figure in the field, also a key driver behind developing Tesla’s autonomous driving neural networks and promoting deep learning, and one of the former co-founders of OpenAI. Andrej KarpathyAfter leaving OpenAI, previously worked at On social media in early 2025Proposed and promoted a new concept:「vibe coding」referring to using AI tools to assist in writing code, allowing developers to build programs and applications through natural language even with zero prior knowledge. It is also known as“Father of Vibe Coding”. However, recently he overturned his own theory through his actions: in launching a new open-source project Nanochat At the time, he admitted that all 8,000 lines of code in the entire system were “purely hand-written,” with no AI used at all.
Nanochat: A Mini Chat Model Built from Scratch
Karpathy is at Released on GitHub. NanochatIt is a “from-scratch, complete end-to-end (full-stack)” training and inference system that enables users to train a ChatGPT-like chat model within hours and at a cost as low as $100.
According to him, Nanochat only contains about 8,000 lines of clean, clear code, enough to present the complete skeleton of how a large language model (LLM) operates. Its purpose is not to challenge mainstream models, but to provide an open, transparent learning example so that more people can understand the internal logic of language models.
nanochat d32, i.e. the depth 32 version that I specced for $1000, up from $100 has finished training after ~33 hours, and looks good. All the metrics go up quite a bit across pretraining, SFT and RL. CORE score of 0.31 is now well above GPT-2 at ~0.26. GSM8K went ~8% -> ~20%,… pic.twitter.com/8UcpefaSJN
— Andrej Karpathy (@karpathy) October 16, 2025
Surprisingly, all 8,000 lines of code were written by him line by line, without relying on any AI coding assistants such as ChatGPT, Claude, or Copilot.
“The entire project was almost completely hand-written (relying only on basic autocomplete).” Karpathy explained in a community post, “I tried using Claude and Codex agents, but they were completely useless and actually slowed things down.”
The Rise of “Vibe Coding”: The Wave of AI Writing Code
This statement forms a stark contrast with his past remarks. In early 2025, he posted on X to promote the so-called “vibe coding” , a development approach that relinquishes control and lets AI take over.
There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper…
— Andrej Karpathy (@karpathy) February 2, 2025
At the time, he described it as: “I barely touch the keyboard and just go by feel. Error messages? I just paste them straight to the model. It doesn’t matter if the code bloats to the point where I can’t understand it—it mostly works anyway.” This attitude once became a signature of the AI coding community’s culture (and spawned many teachers who taught vibe coding), encouraging developers to use AI to generate entire applications or websites, with the focus not on code quality but on rapid experimentation and creation.
However, Karpathy also cautioned that “vibe coding” is only suitable for “toy projects thrown together on a weekend.” As it turns out, he has indeed stayed true to his word: Nanochat is clearly not that kind of project.
Limitations of AI Coding Tools: The Gap Between Theory and Reality
Karpathy’s experience is not an isolated case. According to cloud companies Fastly The survey earlier this year, as high as 95% of developers They stated that after using AI-generated code, they had to spend extra time fixing errors; some even noted that the time spent fixing errors was greater than what they had originally saved.
another research institution METR The analysis also shows that developers using AI tools actually take longer to complete tasks than those writing code the traditional way. Some companies have even started hiring “AI code maintenance specialists” specifically to clean up the logical confusion and security vulnerabilities in auto-generated code. In other words, while AI tools can produce results quickly, they may also generate code that is “confidently wrong.” For less experienced developers, these errors are often harder to spot and fix.
The Meaning of Handwriting Code: Grasping the System’s True Context
The development process of Nanochat revealed a neglected fact:True understanding still requires writing it yourself.For Karpathy, writing every line of code himself is a way to regain mastery over the model’s internal logic. In his post, he emphasized that Nanochat’s code is “clean, concise, and educational,” with the goal of helping developers truly understand the complete flow of a chat model—from data loading to response generation.
This contrasts with the fuzzy development of the “vibe coding” style: the latter emphasizes speed and creativity, while the former emphasizes structure and understanding. Karpathy’s shift reflects a more mature view: AI can assist creation but cannot replace deep thinking.
Although AI can help write code, human engineers remain irreplaceable.
This self-contradictory experiment ultimately sent a clear signal: AI is a collaborator, not a leader. AI coding assistants still hold value in lightweight tasks like rapid prototyping, UI design, and automated testing; but when it comes to projects involving algorithm design, architecture planning, or data efficiency, experienced senior human engineers remain indispensable. This also serves as a reminder to the industry that when we over-rely on generative AI, we may simultaneously lose control over our systems, and we may also leave behind unmanageable technical debt for serious bugs that could surface later—there was recently a case of a vibe coding instructor crashing their own project. But in the editor’s personal opinion, vibe coding isn’t entirely worthless; it can indeed handle small projects or personal tools. However, for larger projects, the team should at least have senior engineers who can handle integration and debugging; otherwise, problems may become truly unsolvable once they arise.
Source: KOCPC Chinese