Benedict Brady

Notes from a talk on vibe coding: giving agents tools, speeding up the bottleneck, critics, memory, skills, integration tests, and the agent native stack.

Fine-tuning π0.5, the open-source VLA from Physical Intelligence, to play chess on a 5 DOF arm, and ablating how much it really understands about the board, colors, and pieces.

Working up the robotic control stack from the lowest level: brushless motors, the torque loop, the velocity loop, and the position loop.

A cross-disciplinary survey of foundation models, motivated at a high level of abstraction: VLMs, realtime interaction models, VLAs, image and video generation, world models, and biological foundation models.

Revisiting private frontier benchmarks (philosophy, negotiation, chess) for the GPT 5.5 release.

Revisiting private frontier benchmarks (chess perception, chess bench, negotiation, Optimization Arena) for the Opus 4.7 release.

Analyzing how different input modalities (FEN, PGN, board images) impact frontier LLM chess ability.

Testing how frontier LLMs negotiate by pitting them against each other in a resource-splitting game across thinking levels, model sizes, and providers.

Three lessons from building and running Optimization Arena: robust verification, bottleneck identification, and inner vs. outer loop optimization.

A comprehensive chess benchmark testing frontier LLMs on endgames, puzzles, and full games against Stockfish.

How coding agents are changing software development, from feature requests to autonomous deployment with minimal human intervention.

A walkthrough of building a multi-DEX router using Claude.

Designing and deploying a proportional AMM on Solana using Claude Code and Opus 4.5.

24 predictions covering crypto assets, DeFi applications, infrastructure, Solana ecosystem, and AI developments.

In-depth analysis of how applications monetize user orderflow on Solana through fees, transaction tips, and market maker access.