AI luminary Andrej Karpathy, former director of AI at Tesla and OpenAI co-founder, has released nanochat—a comprehensive toolkit that lets anyone build their own ChatGPT-like AI assistant for approximately $100. Released on October 13, 2025, this open-source project represents the most accessible path yet for creating a functional large language model (LLM) from scratch.
Unlike previous AI-building tools that required extensive technical expertise or substantial financial investment, nanochat provides a complete “full-stack” solution. This means it handles every step of the AI creation process, from initial data processing to deploying a web interface where users can chat with their custom AI assistant.
However, this democratization of AI development comes with important caveats. While the tools are free, you’ll need cloud computing resources that cost around $100 for a basic model, plus technical skills to navigate the process successfully. For businesses considering custom AI solutions or technical teams wanting to understand how modern AI systems work, nanochat offers an unprecedented learning opportunity—but it’s not quite plug-and-play.
Nanochat builds on Karpathy’s earlier project, nanoGPT, which only handled the initial training phase of AI development. The new version covers the entire pipeline: data preparation, model training, conversation fine-tuning, and deployment. As Karpathy explained in his announcement, “You boot up a cloud GPU box, run a single script and in as little as 4 hours later you can talk to your own LLM in a ChatGPT-like web UI.”
This represents a significant advancement in AI accessibility. Previously, building even a simple language model required cobbling together multiple tools, managing complex dependencies, and navigating numerous technical hurdles. Nanochat consolidates everything into what Karpathy calls a “single, cohesive, minimal, readable, hackable, maximally-forkable codebase.”
The project is hosted on GitHub, the world’s largest repository for software code, where developers can access the complete nanochat toolkit for free. The $100 cost comes from renting cloud-based graphics processing units (GPUs)—specialized computer chips designed for the intensive calculations required to train AI models.
Creating a functional language model through nanochat involves six critical phases, each serving a specific purpose in developing an AI that can understand and generate human-like text.
The first step involves establishing a tokenizer—a program that converts human-readable text into numerical codes that AI systems can process. Think of this as creating a translation dictionary between human language and computer language.
Every word, punctuation mark, and even parts of words get assigned specific numerical values called tokens. For example, the word “artificial” might become token 1247, while “intelligence” becomes token 3891. Modern tokenizers can handle roughly 50,000 different tokens, allowing them to process virtually any text input.
Nanochat includes pre-built tokenization tools, eliminating the need to create this foundational layer from scratch. This step typically requires minimal computational resources and completes within minutes.
This phase represents the most resource-intensive part of the entire process. The AI system analyzes massive amounts of text data to learn patterns in human language—how words typically follow each other, common phrases, and the structure of different types of writing.
Nanochat uses a dataset called FineWeb-EDU, a curated collection of educational content scraped from across the internet. This 24-gigabyte dataset contains millions of high-quality text examples that serve as the foundation for your AI’s knowledge base.
During training, the system processes this data repeatedly, adjusting internal mathematical parameters to better predict what word should come next in any given sequence. This is where most of your $100 budget gets consumed—the intensive calculations require powerful GPU processors that cloud providers rent by the hour.
After initial training, your AI can predict text sequences but can’t engage in back-and-forth dialogue. This step teaches the model how to participate in conversations, understand context across multiple exchanges, and maintain coherent discussions.
The process involves training the AI on thousands of example conversations, showing it how to respond appropriately to questions, maintain topic consistency, and recognize when a conversation has shifted to new subjects. Without this phase, your AI would only generate single responses without understanding conversational flow.
Raw language models often generate inappropriate, biased, or harmful content because they’ve learned from unfiltered internet text. Fine-tuning addresses this through two primary techniques:
Supervised Fine-Tuning (SFT) involves training the model on carefully curated examples of appropriate responses. Human trainers provide thousands of example interactions that demonstrate proper tone, helpful responses, and appropriate boundaries.
Reinforcement Learning from Human Feedback (RLHF) goes further by having human evaluators rate different AI responses, then using those ratings to mathematically reward the AI for generating preferred outputs. This creates a feedback loop that gradually improves response quality.
Systematic testing reveals your AI’s strengths and limitations across different types of tasks. This includes evaluating reading comprehension, mathematical reasoning, creative writing, and factual accuracy.
One crucial test involves multiple-choice questions—a format that doesn’t occur naturally in conversational training data. Your AI must learn to recognize when it should select from provided options rather than generating free-form responses.
Testing also reveals knowledge gaps, reasoning errors, and areas where additional training might improve performance. This evaluation phase helps determine whether your basic model meets your intended use case or requires further development.
The final step involves deciding whether your basic AI assistant meets your needs or requires additional investment for enhanced capabilities. The initial $100 creates what Karpathy describes as a “kindergarten-level” AI—functional but limited in knowledge depth and reasoning ability.
Scaling up requires additional training data, more powerful computing resources, and potentially thousands of dollars in additional cloud computing costs. However, for learning purposes or specific narrow applications, the basic version may suffice.
Successfully implementing nanochat requires comfort with command-line interfaces, basic understanding of cloud computing platforms, and familiarity with software development concepts. While Karpathy has simplified the process considerably, this isn’t a point-and-click solution.
You’ll need to set up an account with a cloud computing provider like Amazon Web Services, Google Cloud Platform, or Microsoft Azure. These platforms rent GPU-equipped servers by the hour, typically charging $20-30 per hour for the computational power needed for AI training.
The provided script automates most technical details, but understanding its operations proves valuable for troubleshooting and customization. The script handles data downloading, model configuration, training execution, and web interface deployment—but you’ll need to monitor progress and resolve any technical issues that arise.
The FineWeb-EDU dataset that nanochat uses by default consists of educational content collected from across the internet. While this provides high-quality training material, it raises important questions about data usage rights and potential copyright concerns.
Major AI companies face ongoing legal challenges regarding their use of internet-scraped training data. While using existing datasets for personal learning projects likely falls under fair use provisions, commercial applications of models trained on such data could face legal scrutiny.
Organizations considering nanochat for business applications should consult legal counsel about data usage rights and consider alternative training datasets with clearer licensing terms.
The basic AI created through nanochat serves primarily as an educational tool and proof-of-concept rather than a production-ready system. Its limited training means it lacks the breadth of knowledge and reasoning capabilities found in commercial AI assistants like ChatGPT or Claude.
However, this limitation also represents an opportunity. The basic model provides a foundation that can be enhanced with domain-specific training data. A medical practice could add medical literature to create a specialized health information assistant. A law firm could incorporate legal documents to build a case research tool.
The modular nature of nanochat makes such customizations possible, though they require additional technical expertise and computational resources beyond the initial $100 investment.
For businesses evaluating custom AI solutions, nanochat offers valuable insights into the complexity and costs involved in AI development. The $100 baseline demonstrates that basic AI functionality is becoming increasingly accessible, but also highlights the substantial additional investment required for production-ready systems.
Technical teams can use nanochat as a learning platform to understand AI development processes before committing to larger projects. The hands-on experience provides practical knowledge about data requirements, computational costs, and technical challenges that inform future AI strategy decisions.
The project also illustrates the growing democratization of AI technology. As tools like nanochat become more accessible, the barrier to entry for custom AI development continues to decrease, potentially enabling smaller organizations to develop specialized AI solutions previously available only to well-funded technology companies.
Nanochat represents a significant milestone in making AI development accessible to a broader audience. While it’s not a replacement for commercial AI services, it provides an invaluable educational tool and foundation for understanding how modern AI systems work. For $100 and some technical curiosity, anyone can now build their own ChatGPT clone—even if it won’t be ready to take on OpenAI just yet.
Join hosts Anthony, Shane, and Francesca for essential insights on AI's impact on jobs, careers, and business. Stay ahead of the curve – listen now!
Join hosts Anthony, Shane, and Francesca for essential insights on AI's impact on jobs, careers, and business. Stay ahead of the curve – listen now!
No hype. No doom. Just actionable resources and strategies to accelerate your success in the age of AI.
