DIY Personalization: How, when why to offer self-made AI
summary
A demo builds a local RAG in a Jupyter notebook: initial LLM-as-retrieval fails, switching to BM25 + sentence encoder plus reranking and a local LLaMA-style model succeeds for offline Q&A. Claims that on-device personalization is practical now cite ChatGPT hitting 1 million users in five days versus the internet taking ~10 years to reach 500 million. Other working projects include a scikit-learn TF‑IDF recommender that actually improves feed relevance, photo captioning with BLIP (Hugging Face) that needs polish, Nextcloud storage and Gadgetbridge watch exports.
AI-written summary · about
chapters (16)
- 0:00DIY personalization and AI anxiety
- 3:38Computer versus AI revolutions
- 13:15VizyCal moment for AI
- 16:23Building personal recommenders
- 20:16Local photo storage and search
- 23:07Health data and smartwatch privacy
- 24:52Local AI inference and scalability
- 27:18Filter bubbles and radicalization
- 29:25Tooling for personal AI
- 31:32Local-first data sovereignty
- 33:19Privacy risks in deep learning
- 35:06Future of communal AI
- 36:10Feminist AI LAN party Berlin
- 37:32Q&A: Hobbyist training
- 40:23Q&A: Open source UX challenges
- 44:33Q&A: DIY AI economics
related talks
01h 03mIt's All About AI
01h 06mScaling personalization with AI
39mBuilding Your Own Bot with GPT: Harnessing the Power of AI...
45mKeynote: Django Reimagined For The Age of AI with Marlene Mhangami
32mThe Inadequacy of Offline LLM Evaluations: A Need to Account for Personalization in Model Behavior
47mOperationalization of AI/ML models and use cases for ChatGPT Q&A
27mFrom RAGs to riches: Build an AI document interrogation app in 30 mins
41mThe Value of LLMs - Why Stupid Machines Are Smarter Than You Think