Writing

What modern RL post-training implies for the next generation of AI.

What enterprises should do in the next two years.

Wiring NVIDIA NIM into the Pi coding agent for free model inference.

Paper on episode capture, replay, distillation, and four-mode evals.

Implementing provably near-optimal 7× KV-cache compression inside a Rust inference server.

Engineering software for mission-critical systems — 60 years of NASA's hardest-won lessons.

The case for open-ended AI.

How to teach language models to actually explore.

How production AI systems orchestrate retrieval, memory, and context windows.

A field-tested synthesis of RAG, agents, and the PoC-to-production gap.

Why protocol-first (MCP, A2A) wins over framework lock-in.

Program synthesis is the dark horse that might get us to world models first.

Research survey across 600+ AI papers: continual RL benchmarking, plasticity crisis, GVFs as proto-world-models.

How continual-RL benchmarks have evolved and where they fall short.

Why deep nets lose the ability to learn — and what recent work suggests as a fix.

What the leading labs are thinking about agentic systems.

Wiring Agave's TransactionBatchProcessor into an optimistic rollup.

A production-grade Rust MCP server for AI agents in crypto and DeFi.

Practical patterns for typed tool surfaces in DeFi-facing agents.

A survey of stablecoin design over the past cycle.

From wallets to validators: how Solana actually includes a transaction.

On Solana's DeFi summer and what Bitcoin and Cardano are building meanwhile.