OpenAI Co-Founder Abandons AI Tools to Build Chatbot by Hand

More than a year after leaving OpenAI, company cofounder Andrej Karpathy has returned with a hands-on demonstration of what it takes to build a working AI chatbot from scratch. His new open-source project, called “nanochat,” offers developers a stripped-down, full-stack chatbot interface that can be trained in just a few hours, and for as little as $100.

Karpathy describes nanochat as a minimal training and inference pipeline designed for those who want to build a large language model with a basic ChatGPT-style interface. The system covers everything from tokenizer training and transformer architecture in PyTorch to a JavaScript-powered web interface.

With roughly 8,000 lines of code, nanochat aims to remain clean, readable, and easy to modify. According to Karpathy, all of the code was written manually, a decision he made after discovering that popular AI tools like Claude and Codex could not produce the quality or specificity he needed for the task.

Not ‘Vibe Coded’

Karpathy coined the phrase “vibe coding” to describe the growing practice of delegating coding tasks to AI tools. But despite popularizing the term, he admitted that nanochat didn’t benefit from those techniques. He initially attempted to use AI agents to assist with coding but ultimately found them unhelpful, choosing instead to rely on tab-autocomplete and write the code himself.

He shared that the AI tools often failed to understand what he needed and slowed his progress. As a result, nanochat became a deliberate return to human-led development, a move that contrasts sharply with current trends in automated software generation.

Fast and Affordable Training

In terms of resources, Karpathy estimates that users can train a working chatbot in about four hours using an 8x H100 GPU setup, which typically costs around $24 per hour, bringing the total to roughly $100. A longer 12-hour run could push performance further, allowing the model to handle math, code, and basic reasoning tasks. The model is trained using publicly available datasets such as FineWeb-Edu-100B and later fine-tuned on instructional data, including MMLU and GSM8K.

Designed for Education and Experimentation

While nanochat is not intended to compete with industry giants in terms of performance or scalability, its strength lies in simplicity. Karpathy designed the project to be forkable and educational — an accessible baseline for developers and students who want to learn how AI chatbots work from the ground up.

He acknowledged that the model may not outperform larger systems, but said it performs well enough to carry on basic conversations and serve as a learning platform. The pipeline also supports optional reinforcement learning and includes a UI for testing chatbot responses.

The post OpenAI Co-Founder Abandons AI Tools to Build Chatbot by Hand appeared first on ProPakistani.

More than a year after leaving OpenAI, company cofounder Andrej Karpathy has returned with a hands-on demonstration of what it… Read More

The post OpenAI Co-Founder Abandons AI Tools to Build Chatbot by Hand appeared first on ProPakistani.