143 lessons · 396 resources · updated 2026-10-19
The history, how models work, how to use them well and how to build with them. Work through this in order, ending with the math that explains why it all works.
From the 1950s to ChatGPT: the ideas, the winters and what finally worked.
0/7 lessons done · 1.6 h
Tokens, prediction, attention and training in plain language.
0/6 lessons done · 79 min
Vision, games, proteins, speech, code and open models.
0/6 lessons done · 78 min
Getting good answers, and controlling what the model sees.
0/8 lessons done · 1.6 h
Using AI for study and work, checking its answers, privacy, honesty and what is real.
0/5 lessons done · 62 min
A workflow for directing agents and verifying their work.
0/9 lessons done · 2.0 h
Calling models, tools, MCP, errors and multimodal input.
0/11 lessons done · 2.4 h
Embeddings, chunking, vector search and RAG, with how to measure each.
0/13 lessons done · 2.9 h
Loops, patterns, tools, memory and multi-agent systems.
0/10 lessons done · 2.2 h
How the craft changed, and which of the new words name something real.
0/4 lessons done · 54 min
Knowing whether it works, seeing what happened and staying safe.
0/11 lessons done · 2.2 h
A map of the layers and the tools in each.
0/4 lessons done · 48 min
Five build-along projects that put each part of the path to work.
0/5 lessons done · 25.0 h
AI at work in the real world: recommenders, forecasting, vision, voice, robots, industries and measuring return.
0/8 lessons done · 1.8 h
The math underneath everything above, with worked numbers and small demos. This is where using AI becomes understanding it.
0/11 lessons done · 2.6 h
Training, serving, production and staying current. Pick what you need.
Build a GPT, modern architecture, fine-tuning, RL and serving.
0/6 lessons done · 1.9 h
Advanced retrieval, cost, MLOps, protocols and human approval.
0/6 lessons done · 84 min
Video as code and deep research workflows.
0/2 lessons done · 27 min
Eight named techniques with worked examples, imported from the MIT-licensed DAIR.AI guide, with our own stories and quizzes.
0/8 lessons done · 37 min
Sources, lists and how to read benchmarks.
0/3 lessons done · 32 min
Recent releases and reads that changed this path. Last updated 2026-10-19.
A minor update to the open MIT-licensed V4.1 Flash with gains in reasoning, coding and agentic tasks. Read: Open-weight models
An 816B-parameter (42B active) open-weight multimodal model under the Muse Model License. Read: Open-weight models
Google's frontier model moves from trusted testers to the API, the Gemini app and Search. Read: Choosing a model
Altitude, depth and range as ways to ask models for more. Read: The agentic coding workflow
A much cheaper small model, suited to high-volume steps and subagent work. Read: Choosing a model
A 1.05T-parameter model with open weights promised by the end of October. Read: Open-weight models
The faster, lower-cost partner to Opus 5.5. Read: Choosing a model
Both labs shipped new model tiers on the same day. GPT-6 Astra had arrived on Sep 3. Read: ChatGPT and how companies kept improving
Agent = Model + Harness, and verification as the line between vibe coding and engineering. Read: Harness engineering: real or hype?
Planner, generator and evaluator agents with context resets. Read: Harness engineering: real or hype?
Scripts that orchestrate hundreds of subagents, with intermediate results kept out of context. Read: Model routing and dynamic workflows