Things I've Made
Products and small things I have made, all in one place.
Products and small things I have made, all in one place.
Downloading the originals behind 2.98 billion image URLs in OmniCorpus-CW, I left almost all operations to an AI agent: patrols, rate tuning, creating and deleting machines, incident handling and billing. I only made the decisions. In 79 hours it saved 99.6% of the images, 744 TB, for about $175 in machines.
On CaptionExam, even GPT-6 Astra only just passes at describing images, with a score of 61.2%.
Training a text-to-image model from scratch means finding, inspecting and filtering samples across billions of image–text pairs. Using a retrieval system over 1.4B samples as an example, this post covers data storage and management, distributed embedding extraction, building and querying an IVF-PQ index, and online retrieval and cluster browsing.
This is a fully annotated guide to Lilian Weng’s post A (Long) Peek into Reinforcement Learning.