Skill Distillation

2026-05-28 · Source: Tomasz Tunguz · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Advanced, extended

Summary

Y Combinator has developed an internal AI infrastructure, exemplified by a "personal agent" based on Pi, which automates various workflows like inbox management and research. This system employs "skill distillation," where frontier models such as Opus 4.7, GPT-5.1, or Gemini 3 Pro author and refine SKILL.md files, which are then executed by smaller local models like Qwen 35B or Gemma 26B. This method transfers procedural knowledge via inspectable markdown, differing from classical knowledge distillation or instruction tuning. YC's architecture integrates a QMD markdown knowledge base, atomic skill files, and an Agent Loop that interacts with over 350 internal Rust APIs and MCP integrations. The system leverages a unified Postgres database for comprehensive context and features self-improving "dream cycles" that enhance skills by analyzing historical agent conversations and meeting transcripts, enabling non-technical staff to manage complex tasks and fostering a "superintelligent organization" approach.

Key takeaway

For Directors of AI/ML aiming to build a "superintelligent organization," prioritize creating a unified data context and a shared internal tool registry. Implement skill distillation to transfer procedural knowledge from large models to smaller, cost-effective local agents, making workflows inspectable and self-improving. Embrace transparency by defaulting agent conversations to be viewable, accelerating team ramp-up and fostering collective skill development, allowing your organization to operate in a future state today.

Key insights

Skill distillation transfers procedural knowledge from large models to smaller, local models via inspectable markdown files.

Principles

• AI acts as a foundational building layer.
• Centralize all organizational context.
• Autonomous loops continuously refine agent skills.

Method

A frontier model authors and refines SKILL.md files, which a smaller local model then executes. This process transfers procedural knowledge through markdown, making skills inspectable and hot-swappable.

In practice

• Create a shared internal tool registry.
• Use meeting transcripts for skill refinement.
• Default agent conversations to public viewing.

Topics

• Skill Distillation
• Agentic AI Systems
• Internal Tool Registry
• Procedural Knowledge Transfer
• Organizational AI
• Local LLMs

Code references

• earendil-works/pi
• tobi/qmd

Best for: AI Architect, CTO, VP of Engineering/Data, AI Engineer, Director of AI/ML, MLOps Engineer

Related on AIssential

• AI tools for hearing difficulties — helpful or harmful for language learning? — AI assistive technology presents a dilemma for language learners with hearing difficulties, balancing aid with skill development.
• How Enterprise Marketing Teams Are Actually Using AI Agents in 2026 — Enterprise marketing is shifting from human-assisted AI tools to autonomous AI agents that redesign entire workflows.
• [AINews] Microsoft Build: MAI-Thinking-1 and MAI Family models — Microsoft's MAI models showcase a transparent, full-stack AI strategy emphasizing clean data and local agent execution.
• Are we getting better at explaining things because of our interaction with LLMs? — Crafting clear LLM prompts may enhance human explanatory and communication abilities.
• OpenAI expands Codex with role-specific plugins to build a general-purpose app for non-developers — OpenAI is transforming Codex into a general-purpose, role-specific AI work app, expanding beyond coding for non-developers.
• How to Use Claude Managed Agents? — Anthropic's Claude Managed Agents simplifies AI agent deployment by providing a hosted, stateful, and secure execution environment.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Tomasz Tunguz.