AI as an Infrastructure

June 21, 2026

I have been writing a book called AI as an Infrastructure. It is not a tour of this week's model releases, and it is not another recipe for getting a demo to run. It asks a slower question: why does the modern AI stack have this particular shape? The book now reads end to end at aaai.latere.ai.

AI is becoming a layer other software stands on. Search, generation, planning, coding, automation: more and more product behavior is being pushed into systems that call models. That means AI has to be treated less like a clever feature and more like infrastructure. It has to stay available, remain controllable, show what happened inside it, and cost an amount someone can explain. Yet much of the writing around AI chases either news or recipes. Which model passed which benchmark? How do you fine-tune the latest release? Both are useful, and both go stale. More importantly, neither explains why tokenization exists, why attention became the center of the model, why KV caches reshape serving, or why a model that can continue text still has to be trained to act like an assistant.

This book follows that causal thread. Each chapter starts with pressure: what problem has to be solved, and what constraints close around it. Then it follows the design that answered those constraints, what it replaced, what it improved, and what it gave up. Implementation comes after the shape makes sense. Read this way, tokenization, attention, scaling laws, RLHF, inference serving, and agent harnesses stop looking like a glossary and start looking like a sequence of engineering choices. I call these links constraint arrows: the places where physics, data, algorithms, and cost below force the behavior above.

To keep the stack concrete, the book keeps returning to one small request. A user asks an assistant to find a bug in a repository, fix it, and explain what happened. That ordinary request cuts through the whole system. Web text becomes tokens. Tokens, data, and compute become model weights. Alignment teaches the model to do the task instead of merely continuing text. Serving turns weights into a low-latency, scalable, observable service. Then an agent gets tools, reads files, runs tests, sees the failure, edits code, and verifies the result. One everyday action pulls on everything from power and chips to distributed training, sandboxes, and audit logs.

The arc of the book is the lifecycle of a capability. It emerges from compute and data, is compressed into weights during pretraining, stretched through multimodal models and test-time compute, shaped by fine-tuning and alignment, exposed through serving systems, and disciplined by orchestration, evaluation, safety, and interpretability. Hardware, power, and cost are not an appendix; they feed back into nearly every design choice. The final part comes back to practice: how to wire existing tools into an AI stack that can actually be deployed, observed, and operated.

Two threads run through the book. The first is three loops: training, inference, and agents. They live at different layers, but all spend compute to reduce some kind of error. The second is the closing lens of each chapter: capability, efficiency, and trust. Can the system do the task? At what cost? On what grounds should anyone trust it? Where the field is still unsettled, I do not smooth over the disagreement. The book uses "what's contested" boxes to spell out who disagrees, why, what the evidence settles, and what it does not. The book is now whole, but not frozen. The infrastructure is still moving, and so is the text.

This is not a beginner's introduction. You should be comfortable in at least one programming language and willing to use some probability and linear algebra. It is written for engineers building and operating AI systems, and for researchers who want to see what their ideas become under production pressure. My hope is that after a chapter, you leave with more than a technique. You should have a way to judge the next model, framework, or deployment pattern by asking what constraints pushed it into that form.

Read it at aaai.latere.ai. Bring counterexamples.

AI as an Infrastructure is written at Latere by Dr. Changkun Ou, under CC BY-NC-ND 4.0.