Seven click-through decks for the Diffusion-Model Rendering for Re-Themable Games onboarding: from a single neuron to the system we are building this year.
Return on effort, the ten pitch points, the pipeline in one picture, today's agenda, tomorrow's teach-back.
The 3Blue1Brown arc on MNIST, with a real 784-16-16-10 network running live: neurons, gradient descent, backprop and its calculus.
Tokens, embedding space, the residual stream, attention, MLPs as fact stores, superposition, temperature.
Prefill vs decode, the KV cache and quadratic attention, GPUs and parallelism, the build-out, RL from first principles, test-time compute.
Why this area, the denoising step function and its intuitions, the U-Net, how channels and theme enter the model.
The data factory, held-out compositions, leakage traps, sweeps, evaluation, the paper, and the call to the five streams.
Tensors and autograd, serving, renting GPUs on AWS with $5k of credits, the system design, failure points, next steps and mini-presentation topics.
Open any deck during the session and it moves with the presenter, slide by slide, on your laptop or phone. No account needed.
Use the arrow keys or swipe to look back. Click "Jump to live" in the top bar to catch up again.
→/Space next, ← back, G slide list, A slide-only view, F fullscreen.