Untangling How BLIP-2 Matches Images to Text Queries
A patient back-and-forth unpacks how BLIP-2's joint image-text embeddings work, then traces the idea's evolution through 2024 vision-language research.
Related
A from-scratch guide to porting a diffusion model to WebGPU
Claude clones a browser-based diffusion model port and writes a from-scratch guide covering ONNX export, WebGPU execution, and every design choice in the code.
Asking three times for a random number: 42, then 8, then 7
A request for a random number returns the Douglas Adams answer, then a figure from code that may not have run, then a number chosen with code switched off.
Auditing grok-build's Silent Repository Upload Code
Claude reverse-engineers xAI's coding agent after reports it exfiltrated repositories, tracing a server kill switch and vestigial upload code in the source.
Building a verified career-minutes chart for footballers
Designing a career-minutes chart for four footballers, with a transparent projection model and an honest pass to verify data pulled from sports sites.