Visual Notes

Why point estimates are dangerous for AI evals and how confidence intervals should be the standard for anyone measuring agentic system performance.

The first benchmark targeting long-horizon agentic memory. Spoiler: every existing memory system underperforms simply extending the context window.

New models ship weekly and MoE makes parameter counts misleading. A guide to knowing when to go small, go big, or route between both.

How dynamic programming finds the perfect speculation tree shape for 4–10x faster LLM inference.

How Genetic-Pareto optimization achieves 35x efficiency over RL by reading execution traces.

From quadratic self-attention to hybrid routing — visualizing the design space.