Why you cannot “apt install” an LLM (yet)
summary
Debian cannot “apt install” current LLMs because models need their training data to meet Debian Free Software Guidelines, datasets are huge (GPT‑2’s original dataset ~40 GB GPT‑2 dataset) and mirrors lack that capacity, and reproducible builds force retraining—costly in time and money (original GPT‑2 training cost ≈ $10k; llm.c retrains the 124M GPT‑2 variant in 20 minutes, larger in ~24h for ~$600). Packaging also raises licensing, transparency, security and mirror-storage problems that smaller ML models or runtimes (OpenCV, llama.cpp) avoid.
AI-written summary · about
chapters (8)
- 0:00Introduction and Debian overview
- 3:21What is a Debian package
- 6:12Defining source code for AI models
- 8:43Technical challenges of LLM packaging
- 13:43Reproducible builds and security
- 16:32Transparency and ethical concerns
- 20:35Takeaways and Debian priorities
- 22:27Q&A session
related talks
28mAre LLMs the answer to all our problems?
47mLaptop-sized ML for Text, with Open Source
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
01h 00mPANEL - Data, Code, and Life with LLMs.
01h 02mA hacker's guide to open source LLMs
44mHow to build your own Tiny Language Model from scratch- 12mKeynote: How LLMs Have Changed the Way Organizations Do AI
27mProductionizing Open Source LLMs