"Train Your Own LLM from Scratch" open source project reaches the top HN: GPT can be trained on a laptop

Core conclusion

On May 5, 2026, the GitHub open source project "Train Your Own LLM from Scratch" topped the homepage of Hacker News. This project was created by Angelos P. and was inspired by Andrej Karpathy's nanoGPT, but it was boldly simplified - the goal is not to reproduce GPT-2 (124M parameters), but to build a mini-GPT model with about 10M parameters, which can be trained in less than 1 hour on an ordinary laptop.

Key Points

• Release time: 2026-05-05 (top of HN home page)
• Core selling point: No GPU cluster required, MacBook can be trained
• Technology stack: PyTorch + pure Python, no black box libraries
• Target users: developers/content creators who want to understand the underlying principles of LLM
• Project size: new project, rapidly gaining community attention

Background

The Hacker News community's curiosity about LLM has been growing, but the path to learning Transformer in the past was either too shallow (tuning the API) or too deep (reading papers). Karpathy's nanoGPT was the first resource to make "do-it-yourself training LLM" feasible, but GPT-2 reproduction of 124M parameters still requires a GPU.

This new project chose a more pragmatic path: reduce the model to 10M parameters, use Shakespeare texts as training corpus, so that an ordinary laptop can run the complete training process in 60 minutes.

Core content of the project

The project uses step-by-step document-driven learning. Each module has matching Markdown instructions in the docs/ directory, allowing you to understand the principles while writing code.

Why this is important

For readers of WayToClawEarn, the value of understanding the underlying principles of LLM is reflected in three aspects:

1. Content production is more efficient When you understand how Tokenizer cuts text and how Self-Attention understands context, you can design more effective Prompt templates and reduce unnecessary Token consumption. The billing of OpenAI and DeepSeek is based on Token, and understanding the underlying layer can directly save money.

2. Automated workflow is more reliable The bottom layer of AI Agent (such as Claude Code, Hermes Agent, n8n) is LLM. Understanding the training process helps understand why AI Agents sometimes "hallucination" and how to set reasonable verification points in the workflow.

3. Professional moat When "everyone knows how to adjust AI APIs", developers who truly understand the underlying principles will have a qualitative advantage. This project just provides a bridge from "caller" to "understander".

Get started quickly

Project installation is extremely simple, requiring only Python 3.12+ and the uv package manager:

# uv
curl -LsSf https://astral.sh/uv/install.sh | sh

#
git clone https://github.com/angelos-p/llm-from-scratch
cd llm-from-scratch
uv sync
mkdir scratchpad && cd scratchpad

#

# docs/ ， Part 1 Tokenizer

It can also be run directly on Google Colab without a local environment.

Tool entry (trigger tool floating card)

If you want to integrate LLM capabilities into your automated workflow, n8n and LangGraph provide low-code solutions. Claude Code and Hermes Agent are more advanced AI Agent frameworks. The APIs of OpenAI and DeepSeek are the entry points for model calls after training.

Internal link guidance

• Want to learn more about AI Agent automation? See: AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes
• Learn how to automate content production with Claude Code: Claude Code automated writing practice: build an AI content production pipeline in 30 minutes
• Real case: How an independent developer can earn $5,000 a month using AI tools: Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo