almanack

AI

Thoughts and references related to AI.

Learning

• NanoChat: ChatGPT trained fromm scratch.
• Smol Training Playbook
  • FineWeb: Building datasets at scale.
  • Ultra Scale Playbook: Orchestrating GPUs for a training run
  • Evaluation Guidebook: Choose the right evaluations

Coding agents

• Goose
• Factory
• Writing a good Claude.md
• Claude-mem: Persist memory of claude code between sessions
• Claude code setup: Walk through of the setup from the Claude team for claude code.

Projects

• KumoRFM: A foundation model for business data. Predictions over structured data.
• Liquid Nanos: Frontier-grade performance on small models.
  • LFM-1B-Math: Detailed description of how it was installed.
  • HF models
• TOON: Text object-oriented notation to save tokens when interacting with LLMs
• NoF1.ai: Model crypto trading agents and portfolio comparison.
• Alpamayo and self-driving ecosystem from nvidia

Infrastructure

• Ray: Ray is a unified framework for scaling AI and Python applications.

Agent environments

Current dev tools target human developers, we need infrastructure targeting agents

• Sketch.dev: Coding agents that run as a standalone process with an end-to-end code commit workflow.
• exe.dev: Isolated containers and development flow for coding agents

Bookmarks

• RAG search and GraphRAG for financial data
• LM Arena Leaderboard

Papers

• Agentic context engineering
• Less is more: Recursive reasoning with tiny networks

Evaluations

• Evalkit: Typescript LLM evaluations library.
• Opik: LLM evaluation platform
• Evaluating deep agents: Some learnings from LangChain.

Trends

High-level trends and ideas

• Environments are a key asset to train AIs through RL. The closer the training environment is to the real one, the better it will end up operating. In the coming years I expect the programming of RL environments for AI to be as important as wiring the agents themselves (esp. for robotics and real world –analogous and flawed environments–). See prime intellect’s RL environments as the seed for this trend.
• Small language models outperform LLMs for agentic behavior:
  • See paper: https://arxiv.org/pdf/2510.03847

Tutorials

• Claude code with local models
• GLM local deployment tutorial

Agent sandboxes

• Sprites
• Exe.dev

This site is open source. Improve this page.