Modern Problems Require Modern Solutions
summary
All processing runs locally on a 5‑year‑old Intel MacBook with no cloud, no Python, using TypeScript plus Hugging Face pipelines to produce 512‑dim embeddings from images and text. Embeddings are stored and queried in Reddus vector DB with flat and HNSW index options (HNSW recommended past ~1M vectors). A TypeScript meme loader creates embeddings for dozens of memes, a server returns the top nearest neighbors via cosine/Euclidean/inner‑product, and the live demo finds sensible meme twins; Reddus benchmarks claim up to 50× speedups versus a competitor.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 3:20What are embeddings and vector DBs
- 10:05Unstructured data to embeddings
- 26:02Building embedding models
- 31:41How convolutional neural networks create embeddings
- 40:27Demo architecture and code
- 47:58Live demo: meme twins
- 50:31Resources and closing
related talks
54mPresentation: Finding Your Meme Twin with Embeddings & Vector Databases
50mA fun and absurd introduction to Vector Databases
46mA Fun & Absurd Introduction to Vector Databases
50mVector Ops: run vector embedding-powered apps in prod
30mScaling Embeddings in Real-Time
06mMachines and the magic of fast learning
33mTowards a deeper understanding of retrieval and vector databases
28m"Embeddings: How Computers Learned to Read"