My Big Beautiful Ambitious Learning And Curiosity Plan For 2026
Goal: One new AI topic learned and one new piece of software built, every week in 2026.
| Week | Date | AI Topic | Build | Link |
|---|---|---|---|---|
| 01 | Jan 05 | Beads — coding agent memory system | Bloom filter | Week 01 |
| 02 | Jan 12 | Claude Code | TODO MCP server | Week 02 |
| 03 | Jan 19 | Strands Agents + Ollama | Jarvis chatbot | Week 03 |
| 04 | Jan 26 | AG-UI protocol | Jarvis TODO integration + two-phase commit | Week 04 |
| 05 | Feb 02 | A2A protocol | Jarvis A2A server | Week 05 |
| 06 | Feb 09 | Agent Client Protocol (ACP) | Symmetric encryption | Week 06 |
| 07 | Feb 16 | Embedding models + vector similarity | Vector database | Week 07 |
| 08 | Feb 23 | Ralph autonomous agent system | TCP three-way handshake | Week 08 |
| 09 | Mar 02 | OpenClaw — sandboxed AI coding | CRDT collaborative editor | Week 09 |
| 10 | Mar 09 | Language modeling — bigram model (makemore) | Bigram language model | Week 10 |
| 11 | Mar 16 | Language modeling — neural net framework (makemore) | Bigram model in PyTorch | Week 11 |
| 12 | Mar 23 | LLM Wiki (Karpathy) | LLM Music Producer | Week 12 |
| 13 | Mar 30 | Feature engineering + representation learning | (no build) | Week 13 |
| 14 | Apr 06 | — | — | (not started) |
| 15 | Apr 13 | — | — | (not started) |
Best builds:
- LLM Music Producer — two complete pipelines for LLM-driven audio composition: Base95 frame payloads and MIDI-as-text; the MIDI approach produces genuinely musical output
- Jarvis — a personal AI assistant grown incrementally across 3 weeks, ending up with MCP tool access and an A2A server; a complete end-to-end agent system
- CRDT Collaborative Editor — collision-free concurrent edits without coordination, elegant data structure design
- Bigram Neural Net — the moment count-based statistics and neural networks converge to the same answer
Most interesting AI topics:
- LLM Wiki (Karpathy) — a dense practitioner-oriented map of the entire LLM stack: architecture, training, inference, and emergent capabilities
- Language Modeling fundamentals — building a language model from scratch reveals how the entire LLM stack is constructed
- Agent Protocols — three layers (MCP, A2A, ACP/AG-UI) converging into a standard agent communication stack
Running themes:
- Agent infrastructure — 7 of 13 weeks touched agent frameworks, protocols, or tooling
- Build-what-you-learn — several AI topics were immediately applied as hands-on builds (embeddings → vector DB, language modeling → bigram model, MCP → TODO server)
- Incremental systems — Jarvis and the makemore series both show how complex systems grow from simple foundations
- Representation matters — week 12 showed that MIDI (semantic) beats Base95 (statistical) for LLM audio generation; the choice of representation is the most important design decision
- Wiki Index — full internal catalog: tools, builds, concepts, synthesis
- Tools — one page per AI topic or tool learned
- Builds — one page per thing built
- Concepts — cross-cutting ideas that emerged across multiple weeks
- Code — all builds organized by domain
Updated by LLM on every ingest. See wiki/index.md for the full internal catalog.
The markdown-based knowledge base in this repo is an implementation of the "LLM Wiki" idea from Andrej Karpathy.