Slides & Demos

The canonical LLM paper trail (Transformer → BERT/GPT → scaling laws → RLHF → LoRA → FlashAttention → MoE → test-time scaling → Muon), one slide per paper.

• What’s actually in the paper — the ablations everyone skips — and what’s used in practice today
• Live calculators: Chinchilla-optimal scaling, LoRA parameter counts, GQA KV-cache sizing, speculative-decoding speedup
• A draggable RoPE rotation demo
• Navigation: ←/→ keys, G overview grid, / full-text search, D dark mode

The engineering underneath: parallelism axes and the device mesh, collectives, ZeRO/FSDP, context/tensor/pipeline parallelism, MoE + expert parallelism, and the serving stack (PagedAttention, continuous batching, KV-cache economics).

• Mesh planner and per-GPU training-memory calculator
• Step-through continuous-batching simulator
• Ring-attention animation with a causal-bubble toggle

A side project: a full web arcade where every game is one self-contained HTML file with synthesized audio, and all of them share a coin economy, a collectible card gacha (46 legends, each with unique seeded generative art), daily missions, and 18 cross-game achievements. Progress lives in your browser — no accounts, no real money. Agent-built, verified by headless simulation tests plus real-browser runs.

• Echo Hunter — you’re blind; the world renders only where your sonar ping sweeps it (and the creatures hear your pings)
• Terrarium & Ant Empire — zero-player simulations: an ecosystem whose creatures measurably evolve, and real pheromone-trail stigmergy
• Star Forge — a true-3D n-body gravity sandbox: fling planets, capture moons
• Iron Fists, Turbo Seven, Grimhold 3D, Isle Keep — a 2.5D tier: IK-animated fighters, Mode-7 racing, a raycaster, isometric sieges