Where should you begin?
This site has 120+ long-form articles across more than a dozen series. Pick the track that fits you and read in order — each step builds on the last, and every series article ends with a link to the next one.
Interested in the hardware, energy and money behind AI? 🖥️ See the AI Compute Stack at a glance →
Want to see how a frontier model is actually made? 🏭 See Producing Intelligence at a glance →
Curious who actually runs these machines? 🧑🔧 See Who Runs the AI Factory at a glance →
Want to compare the world's biggest AI factories? 🏗️ See AI Factories at a glance →
Want an AI factory of your own? 🧰 See AI Factory as a Service at a glance →
Curious about the software inside? 🧠 See Inside the ML Layer at a glance →
Where does training data come from? 🗃️ See Inside the Data Layer at a glance →
How are models tested before they ship? 🎯 See Evaluation at a glance → · What lets agents act safely? 🤖 See the Agent Layer at a glance →
See the whole machine at once? 🏭 Open The AI Factory, Cut Open → · 🛰️ From Above → · 🏗️ In 3D → · 🧬 Software stack, live → · 🧬 Stack Innovation Map →
🌱 Understand how it works
For curious readers, students and anyone switching in. About 6 hours of reading.
- 1Build a GPT from scratch, explained plainlyThe interactive course in Layman mode: what a language model actually is, how it learns, and how it's trained — no jargon.
- 20The complete stack (interactive)Everything from transformer to production agent on one animated map — follow a request through every layer.
- 2How a transformer works, end to endFollow one sentence through a GPT-style model step by step, with interactive demos and code you can run — including how training works.
- 3Reasoning models & test-time computeWhy the newest models "think" before answering, and why that became the second scaling law.
- 4Multimodality & the great consolidationHow text, image, audio and video models are merging into one.
- 5AI agents & the enterprise agent economyWhat changes when models stop answering questions and start doing tasks.
- 6AI safety, alignment & interpretabilityThe open problems that decide whether any of this goes well.
- 7The GPT lineage, GPT-1 to GPT-6One model family's full history — the clearest way to see how the field actually progressed. Other labs' lineages follow via the Next links.
- 8The research frontier mapWhat every major lab is betting on right now, side by side.
🔬 Build the skills to do the research
For students and engineers aiming at research or research-engineering roles. Weeks, not hours — work the exercises.
- 1Self-assessment: find your gaps firstCheck what you already know so you can skip it — then fill only the gaps.
- 2The mathematical foundationsThe linear algebra, probability and optimization you'll actually use.
- 3Core deep learning conceptsThen continue through the Foundations curriculum — RL, interpretability, alignment, world models, systems — using the Next links.
- 4Build a language model from scratch (Stanford CS336)A lecture-by-lecture companion with runnable code: tokenizer, resource accounting, architecture, and on through training.
- 5How pretraining works, end to endFrom raw web text to a foundation model: data filtering, deduplication, the training recipe and the systems that scale it — with a pipeline you can run.
- 6How fine-tuning works: full fine-tuning vs LoRATurning a base model into an assistant or specialist, with a runnable experiment comparing the two main methods.
- 7How post-training works: SFT, RLHF, DPO and RLAIFHow a base model becomes an aligned assistant, with an experiment comparing SFT, DPO and RLHF on the same model.
- 8How inference works: serving a model at scalePrefill and decode, the KV cache, batching, quantization and speculative decoding, with a mini engine you can run.
- 9Scaling laws, GPUs and distributed trainingWhy bigger models work, what GPUs are fast at, and how one model is split across thousands of them, with a mini scaling study you can run.
- 10Production LLM system designEverything around the model in a ChatGPT- or Claude-style assistant: context building, memory, tools, safety layers, caching and streaming, with a runtime you can run.
- 11Context engineering and memoryWhat goes into the context window and how it is chosen, compressed and remembered, with five memory strategies measured side by side.
- 12RAG: retrieval-augmented generationHow assistants find and cite the right documents — chunking, embeddings, hybrid search, reranking and evaluation, measured on real data.
- 13Tool and function callingHow models go from answering to acting: schemas, the call loop, permissions, retries and prompt-injection defence, with an orchestrator you can run.
- 14Agentic systemsThe plan-act-observe loop, subagents, human approval and stopping rules, and why reliability decides how long a task an agent can finish.
- 15Multi-agent systemsWhen a team of agents beats one agent — and when it doesn't: roles, routing, shared state, parallelism and cost, measured.
- 16Agent memory and planning patternsHow long-running agents keep state, survive crashes, learn from experience and replan, with each effect measured.
- 17MCP and tool ecosystemsThe Model Context Protocol on the wire: how agents discover and use tools, resources and prompts, with a server you can run.
- 18Production agent architectureHow everything comes together into a reliable platform: control and execution planes, durable execution, routing, tracing and scaling.
- 19Evaluating agentsHow to tell whether an agent really works: trajectories, environment checks, judges, pass^k reliability and release gates.
- 20Harness engineeringWhy the system around a model decides what it can do — measured with one model inside five harnesses.
- 21How to read a paper like a researcherA repeatable method for getting through the literature without drowning in it.
- 22Day one at a frontier labThe engineering practicum: production training code, distributed debugging, kernels, on-call.
- 23Recursive self-improvement: a primerOne live research frontier, end to end — then the CS329A self-improving agents handbook.
- 24The researcher's field guide: getting hiredPapers, labs, people and the hiring process — plus the study stack and PhD programs guides.
🚀 Find the opportunity and build in it
For founders, operators and investors. About 6 hours, then the 12-part startup series.
- 1The six real levers of frontier AIThe framework the rest of this track uses: what actually moves capability forward.
- 2The AGI capability stack, mappedWhere today's frontier models stand on each layer, and what closing each gap takes.
- 3The levers nobody's pulled yetWhere a new lab could realistically win against incumbents.
- 4Compute frontier: why clusters keep growingThe cost structure underneath everything: training runs, mega-clusters, hardware.
- 5Frontier model pricingWhat the models cost to use, and what that means for product margins.
- 6AGI startups to watchWho's already building in the open lanes.
- 7The three startup lanesFrontier scale-first, the applied & agentic layer, and the narrow research bet — strategy and team for each.
- 8Founding an AGI startup: the 12-part seriesFrom finding the idea through legal, team, seed, product, go-to-market and Series A.