Escolha um formato:
| zip | Baixar grátis |
|---|---|
| rar | Baixar grátis |
| Baixar grátis | |
| mp3 | Baixar grátis |
| audiobook | Baixar grátis |
| kindle | Baixar grátis |
Descrição do livro
Eery year, a new model claims to change everything. Benchmarks are shattered, demo videos circulate, and for a week or two it seems as though the discipline of building software has been quietly retired, replaced by the discipline of prompting. Then the demo has to become a product, the product has to survive a thousand concurrent users, a hostile input, a provider outage, a compliance auditor, and a finance team asking why the inference bill tripled — and the model, however good, turns out to have been the easy part. This book is not about models. It is about the load-bearing structure you build around them: the seams where a request enters the system, the layer that decides which model answers it, the cache that keeps you solvent, the retrieval pipeline that keeps the model honest, the guardrail that keeps a single bad output from becoming an incident report, and the observability layer that lets you find out what actually happened at three in the morning when a customer says the agent refunded them for something they never bought. I have spent the better part of two decades doing exactly this kind of work — first with ordinary distributed systems, then, more recently, with systems that have a genuinely new property ordinary distributed systems never had to fully reckon with: a component in the critical path that does not behave the same way twice. A payment service either processes a payment or it doesn't, deterministically, given the same inputs. A large language model given the identical prompt twice, at the identical temperature, can still hand you two different answers, one of which happens to be wrong in a way that sounds exactly as confident as the one that was right. Traditional software architecture has excellent tools for handling systems that fail loudly. It has far fewer tools, historically, for systems that fail fluently — that hand you a plausible, well-formatted, entirely incorrect answer and give you no exception to catch. That single property — plausible failure — is the quiet organizing obsession behind nearly every pattern in this book. Once you take it seriously, a great many architectural decisions that look optional in a tutorial become load-bearing in production: the retrieval layer that grounds a model's answer in something checkable, the evaluation harness that catches a regression before your users do, the circuit breaker that knows the difference between a slow model and a wrong one, the audit trail that lets you reconstruct, after the fact, exactly which version of which prompt against which retrieved documents produced the sentence that is currently sitting in a legal team's inbox. This book will not make you a better prompt engineer. There are other, better resources for that, and prompting itself moves fast enough that anything I wrote about it here would be stale before the ink dried. What I can offer instead is something that changes more slowly: a way of thinking about the shape of an AI system — where its seams should go, what should be swappable, what must be observable, what must degrade gracefully rather than catastrophically — that will still be true when the model behind the seam has been replaced three more times. Keep this book close to the actual system you are building. Dog-ear the chapters that describe failures you recognize. Skip, guiltlessly, the ones describing problems you don't yet have — that section on durable execution for long-running agent workflows can wait until you actually have a long-running agent workflow, and reading it now will only convince you that you need infrastructure you don't. Architecture is the discipline of deciding, in advance, what you will regret not having built. I hope this codex helps you regret less.
Autores populares
Icon Group International (289) Publishing lmgdaw (229) Nikolay Krechet (123) Kennie B Journals (118) Charles Nehme (107) webmarbook design (97) Riddler Books (94) Ondrej Sarek (87) Jillo Trukudo (77) dkdesigns Publishing (76) Mint and Cherry Notebooks (74) Maria Emmrich (73) Just Visualize It (72) Various (69) Moonpeak Library (67) Reynhard Boegl (67) Bettina Schipp (66) matlasor Sticker's (66) Talor Press (66) Joseph Alexander (65)
