Blog

Interactive tools, long-reads, and deep dives into ML, GPU architecture, and inference

RL for Diffusion Models: Part 2

Reward models for diffusion RL: annotation, yes/no checklists, VLM judges, agentic critics, and balancing several rewards in one advantage (GDPO, GD²PO, DAPO).

RL for Diffusion Models: Part 1

How to deliver a reward to a diffusion model: gradients through the trajectory (ReFL, DRaFT, DRTune, ShortFT) and online policy gradient (Flow-GRPO, DiffusionNFT, AWM).

Direct Preference Optimization (DPO) for Diffusion Models

Teaching visual preferences to diffusion models: DPO pairs, FiFA, and training without losing generation quality.

Age of Autoresearch

Expertise moves from "find the solution" to "build the system that finds it". Oracles, loops, pitfalls — and what remains for the human.

GPU Story: from triangles to AI chips

How a triangle-drawing chip became the most important processor on the planet. GPU history from first principles — parallelism, shaders, CUDA, Tensor Cores, and what comes next.

Real-Time Video Generation Calculator

First-principles compute estimator for diffusion and AR streaming video models. Adjust model size, steps, VAE compression, sparse attention, hardware, and precision to see if realtime generation is achievable.