YC-backed founder Paul Graham @paulg recently shared some thoughts about startups and LLMs (large language models). He said, 'Someone asked me what I would do if I were 17. I would learn how to build large language models from scratch, then train the most powerful model I could with whatever hardware I had access to.' If you're still young (and if you're not, don't fall behind ), you can also try learning more about LLMs. Here are a few great resources I’ve come across during my own research: https://((github.com))/jingyaogong/minimind A great project for learning and practicing small models. This open-source project aims to train a super small language model, MiniMind, from scratch with just $3 and 2 hours of training time, resulting in a model around 64M in size. https://((github.com))/arman-bd/guppylm This project aims to prove that training your own language model isn’t hard. No PhD or massive GPU clusters needed. With just a Colab notebook and five minutes, you can build a functional language model from scratch—including data generation, tokenizer, model architecture, training loop, and inference. If you know how to run a notebook, you can train a language model. https://alisawuffles.notion.site/alisa-s-book-of-llms A Chinese girl, Alisa Liu, who successfully landed a job at OpenAI, shared her interview prep notes publicly. These include her LLM notes and another set of math interview notes. Many people find these notes solid and comprehensive, covering topics from Transformers and RL (like GRPO) to system scaling.