Hacking Diffusion into Qwen3 for the ARC Challenge
Exploring the fundamental tradeoffs between autoregressive and diffusion approaches for ARC solving - an empirical analysis of speed, accuracy, and architectural constraints.
Read more →You Might Not Need RAG
We built a life insurance classification bot that matches clients with carriers based on medical conditions. Our initial implementation used what we thought was a temporary solution—letting the LLM pick and read files directly. After implementing a 'proper' RAG system, we discovered our simple approach wasn't just easier to maintain—it performed better. Here's what we learned about the tradeoffs.
Read more →