Changelog

This page records reader-visible changes to the book. Use it to decide whether to return to a chapter, review a correction, or explore new material. Each correction states what was wrong and what replaced it. Weeks without a reader-visible change are omitted, and release tags appear in the week in which they were published.

August 4 to 8, 2026

The English rewrite reached the Epilogue. The work proceeded one complete unit at a time, with each chapter rewritten, tested, rendered, and reviewed before the next chapter began.

Changed

• The English rewrite continued from Behavior Specifications and Preference Data through the remaining technical chapters, Practice Summary, and the Epilogue. Explanations now distinguish mechanism from claim, state the relevant scope, evidence, failure behavior, and operating handoff, and connect each chapter to the layers around it. Repetitive summaries, unsupported absolutes, and machine-like transitions were removed.
• Each completed unit gained chapter-specific regression tests for its central argument, stable headings, equations, diagrams, citations, and handoff. Every unit was also checked in the production reader on desktop and mobile before it was committed.

Corrected

• The rewrite exposed several accessibility faults in existing interactive material. The compute-budget and agent-training figures now name their axes and system dimensions; the KV-cache calculator, paged-attention visual, attention heatmaps, process steppers, Graphviz diagrams, and runnable editors now have usable labels. Stepper arrows no longer dangle when a row wraps, and their controls stay together at narrow widths.
• The quantization visual used a misleading comparison and was corrected. The Verification Frontier legend could overlap or become unreadable on a narrow screen; it now remains legible in the mobile reading column.

July 27 to August 3, 2026

The site made its open-source workflow visible to readers, repaired its URL handling, and began a complete English clarity pass.

New

• A new Contribute page explains which reports are useful, how to propose a change, and what the license permits. Every chapter now ends with one link for reporting a problem and another for editing that chapter's source file.

Changed

• The source repository is linked from both the header and the sidebar. The links above the table of contents remain visible when the Preface is open.
• The first stage of the English rewrite covered Orientation, Base Model Formation, Generative and Multimodal Architectures, and Supervised Fine-Tuning and PEFT. Each completed chapter received a focused content regression and a production render review before work moved forward.

Corrected

• Pages could render at both slash and non-slash forms even though relative images and chapter links worked at only one form. Every page now has one canonical URL, and the alternate form redirects to it before the page is served.

July 20 to 26, 2026

A primary-source audit produced 51 verified corrections. The eleven-chapter infrastructure section was also split into two parts. This work was released as v0.2.0.

New

• Law, Regulation, and Policy gained a section on frontier safety frameworks. It explains how a developer sets capability thresholds, which evaluations test them, which safeguards activate when a threshold is crossed, how published frameworks differed in 2026, and how Anthropic applied ASL-3 after a threshold was reached.
• RAG and Retrieval gained an authorization stage. It compares carrying access-control lists into the index with applying authorization filters at query time, accounts for the cost of both designs, and explains why filtering only after retrieval can leave authorized users with too few results.

Changed

• Part IX was divided into Infrastructure and Compute and Frontiers and Limits. Infrastructure and Compute now runs from Accelerators and Networking through The Machine That Breaks at Scale. Frontiers and Limits contains Where Learning Hits Limits, The Capability Horizon and Its Measurement, and The Verification Frontier. Ecosystem and Economics became Part XI, and Practice and Operations became Part XII. Redirects preserve the former addresses of the three moved chapters.
• Twelve technical terms regained a plain-language definition at first use. They include policy in its reinforcement-learning sense, chain of thought, held-out set, model judge, prompt injection, superposition, attention sink, and cross-entropy. Verifiers and Process Supervision and The Software Substrate now define symbols before displaying their equations. Reliability for Nondeterministic Systems owns the definition of the quality service-level indicator; Operating Contracts refers back to that definition instead of repeating it.
• The Chinese reader now localizes its sidebar links, renders part-opening quotations in their intended panels, and keeps the labels in the nested-loop figure from overlapping.

Corrected

• Tokenization defined fertility as characters per token. Fertility is the average number of tokens into which a word is split, so the former definition was its reciprocal and reversed the argument built on it.
• A caption in Beyond Dense Transformers described the KV cache and attention's per-step cost as quadratic in sequence length. The KV cache and attention's per-step cost are linear; only attention's total work over the complete sequence is quadratic.
• Faster Decoding reported an 80 percent acceptance rate for DeepSeek-V3 multi-token prediction. The technical report gives 85 to 90 percent.
• Choosing a Model still named GPT-5.5 as OpenAI's flagship and recommended it for agentic coding. The table and picker now use GPT-5.6 Sol, Terra, and Luna, with Sol identified as the flagship. The long-context price threshold was also corrected from 270K to 272K input tokens.
• The AI Value Chain described a fall from 87 percent to 79 percent as roughly four percentage points. The difference is eight percentage points, divided between infrastructure and applications in the table.
• Making the Silicon cited one TrendForce article for CoWoS capacity, lead time, and an elevenfold demand claim, although that article contained no capacity figure. Capacity statements now cite dated sources that contain the reported numbers. The unsupported lead-time claim was removed. NVIDIA's share is now described using reported 2026-27 bookings and its reported 2025 CoWoS-L share, rather than an uncited capacity estimate.
• Reference summaries stopped at the first percent sign, truncating 232 bibliography fields. Percent signs are now preserved. Six dead citation links were replaced, and two of those entries also received corrected author and year metadata.

July 6 to 12, 2026

The reader received a new visual system. The evaluation chapters and the sandbox credential model were also aligned with the model APIs and secret-handling mechanisms in use in mid-2026. This work was released as v0.1.0.

New

• Security and Authorization now treats egress substitution as a distinct credential mechanism. A sandbox holds an opaque placeholder, while a TLS-terminating proxy replaces it with the real secret only for an allowed destination. An interactive stepper compares three versions of the design, and a diagram shows the resulting data path.

Changed

• The reader now uses a floating header, translucent side panels and drawers, a separate article panel, and a circular reading-progress indicator. Sidebar entries respond to the pointer, and the article panel avoids an expensive blur effect while scrolling.
• The Harness, Evaluating Agents and Capabilities, Model Judges and Preference Rankings, Human Evaluation as a Measurement Protocol, and Operational Evaluation and Governance were reorganized around a shared rule: a score belongs to a complete evaluation setup, not to a model name. They now identify the served route, harness version, image tag, configured and observed effort, prompt hash, judge identity, and rubric hash. Small measured differences remain noise until an appropriate test resolves them.
• Evaluation and orchestration examples that referred to the retired agon product now describe Latere's Adversarial Review capability. A comments-service failure no longer blanks the chapter; the page remains usable and reports zero comments.

Corrected

• The book presented a caller-set budget_tokens field as the current way to request reasoning depth. The model APIs covered by the chapter no longer accepted that request shape. The text now describes adaptive reasoning under an effort ceiling and notes that an easy request can use few reasoning tokens even at the highest setting.
• Wiring the Stack claimed that all providers accepted the OpenAI request format. The chapter now distinguishes three incompatible dialects, treats a gateway as a potentially lossy translator, and states that OpenAI reasoning levels can be combined with tools only through the Responses API.
• The Tooling Ecosystem said that Cella injected a real secret into the sandbox when a command needed it. That mechanism had been retired. The secret now remains outside the sandbox and is substituted on egress. Related chapters also stopped assuming that every third-party credential can be replaced with a short-lived token; static API keys cannot always be minted per request.

June 29 to July 5, 2026

Eight new chapters and three new sections filled structural gaps, especially the software layer between model code and the accelerator. The rest of the book was checked against primary sources current to mid-2026.

New

• The Infrastructure Before This One explains what generative systems inherited from large-scale ranking, recommendation, and classical machine-learning systems, what changed, and where this book draws its boundary.
• The Software Substrate: Frameworks and Automatic Differentiation explains reverse-mode differentiation, eager execution, and graph capture. Compilers, Kernels, and the CUDA Moat follows a captured graph into device code, shows why data movement dominates many kernels, and explains how an accumulated software ecosystem affects accelerator competition.
• Confidential Inference: Trusted Execution and Private Serving separates a contractual promise not to train on prompts from controls enforced by hardware. It covers attestation, performance cost, and the limits of a trusted-execution boundary.
• Computer Use: the GUI as an Action Space examines visual grounding, action reliability, and the cost of screen interaction. Personalization Memory locates user state outside the base model and explains how that state affects product migration and user control.
• The Model as an Artifact treats weights as files that must be serialized, signed, distributed, verified, and loaded. It includes supply-chain attacks and weight-level backdoors. The Agent Economy covers identity, delegated authority, and payments when software holds and uses a credential.
• Law, Regulation, and Policy gained third-party assurance through ISO 42001, the NIST risk framework, evaluation providers, and insurance. Compute Markets and Unit Economics gained a capital layer covering depreciation, GPU-backed debt, and circular vendor deals. Serving Multimodal Models gained a cost model for diffusion and video generation.

Changed

• Every part was rechecked against primary sources. Benchmarks as Measurement Contracts now includes the disagreement between an outside audit of Humanity's Last Exam and the maintainers' review, as well as the dispute over undisclosed private leaderboard testing. Choosing a Model was updated to July 2026 and records the return of Fable 5 and Mythos 5 after a three-week suspension. The inference chapters now describe disaggregated prefill and decode as deployed practice rather than an unresolved proposal.
• Technical terms are explained at first use for software engineers who are new to machine learning. Scaling Laws and Diffusion and Flow Matching received gentler introductions. The “what remains contested” material in Agent Architectures once again renders as a callout rather than an ordinary heading.

Corrected

• The Whole Stack in One Pass drew a serving-volume crossover after which a mixture-of-experts model became cheaper than a dense model. No such crossover exists in the worked comparison: training and serving compute both follow activated parameters. The sparse model uses less compute at every shown volume but must keep all 671B parameters resident while activating 37B per token.
• The KV-cache calculation in Serving Multimodal Models omitted the grouped-query divisor and overstated the cache required for one image by about four times. The corrected calculation includes that divisor.
• The Compute Frontier interpreted GDPval's September 2025 result of 47.6 percent as near expert parity but not parity. By mid-2026, the leading model won or tied about seven in ten blind expert comparisons. The chapter now separates that graded result from evidence of value in deployed work.
• The Chinese Training and Fine-tuning in Practice chapter reversed the direction of the Tinker relationship. The user writes the training loop and rents distributed GPUs from the service, not the other way around.

June 22 to 28, 2026

Eighteen new chapters expanded reasoning, post-training, evaluation, infrastructure, and economics. Reader accounts also arrived, with comments, selected-passage discussions, private notes, and bookmarks.

New

• Reasoning and Test-Time Compute expanded from three chapters to seven. Structured Reasoning as Search frames reasoning as a search process. Programs, Solvers, and Symbolic Scaffolds separates successful execution from task correctness. Verifiers and Process Supervision covers intermediate checks, and Reasoning Data and Distillation explains how reasoning traces are produced and reused.
• Post-Training gained Behavior Specifications and Preference Data, Verifiable Rewards and Reasoning Transfer, and Safety Tuning and Instruction Hierarchy. Together they cover defining desired behavior before collecting comparisons, learning from machine-checkable outcomes, and resolving conflicts between instruction sources.
• Evaluation gained Statistical Reliability as a Decision Contract, Human Evaluation as a Measurement Protocol, Factuality, Grounding, and Evidence, and Operational Evaluation and Governance. These chapters connect benchmark uncertainty, human judgment, source support, and production release decisions.
• Ecosystem and Economics gained The AI Value Chain, Adoption and Productivity, and Data Rights and Compliance Economics. They examine where value accrues, how field evidence should support productivity claims, and how rights alter the cost of data.
• Four chapters closed gaps elsewhere: Mid-Training covers the stage between pre-training and post-training; The Verification Frontier examines claims that are costly for people to check; Human Interfaces and Oversight Loops places human authority inside a running system; and Operating Contracts defines the promises required when infrastructure is shared.

Changed

• Every part now ends with a summary. Material that had been attached to the part opener moved to the summary and was rewritten to hand the argument to the next part.
• Figures and interactive diagrams were repaired where labels overlapped, legends covered data, Chinese text clipped, or dark mode obscured diagram text. Text inside diagrams can now be selected and copied.
• Both editions received a wording and notation pass. Technical terms are defined where they first matter, and the probability that all n independent steps succeed is consistently written as p^n instead of alternating between p^t and p^n.

Corrected

• The front matter now discloses that Lux, Cella, Topos, Wallfacer, and Lectio are the author's products. Wiring the Stack stopped recommending a self-hosted Python proxy as the default gateway after a throughput test showed it saturating under agent load. It now recommends a compiled gateway for that workload and limits its latency claim to low load.
• Scaling Laws attributed L = E + A/N^α + B/D^β to Kaplan and colleagues, whose fitted form did not include the irreducible-loss term E. The attribution now points to Rosenfeld and colleagues and to Chinchilla. Unsupported explanations for a Chinchilla replication disagreement were removed.
• Non-Autoregressive and Diffusion Language Models printed the masked-diffusion weight with the wrong sign. The positive cross-entropy weight is -α'_t/(1-α_t), not α'_t/(1-α_t).
• Edge and On-Device Deployment said MobileLLM-350M was forty times smaller than LLaMA-2 7B; it is twenty times smaller by parameter count. Multi-Agent Systems attributed the PSPACE debate result to Brown-Cohen and colleagues; the result is due to Irving, Christiano, and Amodei.

June 19 to 21, 2026

The project began on June 19 and published v0.0.1 on June 21. In three days it established sixty-seven chapters across twelve parts in English and Chinese, together with the reader, initial figures, runnable code, and references.

New

• The first book structure followed a capability from compute and data through training, serving, orchestration, evaluation, governance, and operations. English and Chinese editions used the same chapter structure.
• Generative and Multimodal Architectures collected Diffusion and Flow Matching, Non-Autoregressive and Diffusion Language Models, Speech and Realtime Voice, and Multimodal Models: Fusion and Generation into a dedicated part.
• Borrowed Ideas traced concepts such as compression, reward signals, and thermodynamics to their source fields. Training Agents to Act covered learning to affect an environment. Embeddings and Representation Learning explained comparable vector representations. Edge and On-Device Deployment covered execution without a datacenter.
• The first release included twenty-six interactive figures and runnable Python cells, including an attention heatmap, a paged-attention allocator, an imbalanced router, step-by-step RLHF, DPO, and R1 processes, and a self-hosting cost calculator.
• The reader included a glossary of 121 terms linked from first use, full-text search under Cmd+K, and chapter-specific Further reading lists generated from local bibliographies with a short explanation for each source.

Changed

• The book was reorganized into eleven numbered parts plus Orientation, and numbers were removed from chapter URLs. A chapter can now move without losing its address; redirects cover addresses used before the reorganization.
• A purpose-built web reader replaced the earlier presentation. It provided a fixed header, collapsible navigation and table of contents, reading progress, light and dark themes, and settings for palette, type size, and body font. The book remained web-only because its runnable cells and interactive figures do not survive a static PDF or EPUB.
• The Chinese edition was rewritten as native Chinese rather than translated sentence by sentence from English.

Corrected

• The Chinese RLHF and Reward Modeling introduction promised three items but listed four. The introduction now matches the list.
• Training Models to Reason claimed that several verifier-based recipes reduced or removed the KL penalty used in RLHF. Only DAPO removes it. The chapter now limits the claim to DAPO and states which controls replace the penalty.
• Beyond Text assumed that readers already knew the data-wall argument from a chapter that the reorganization had moved seven parts later. The concept is now explained at first use. A Field Map still described nine parts after the same reorganization and was updated to match the actual structure.
• The reference system had three defects. An unencoded space broke the link to “The Era of Experience.” Several entries listed co-authors who could not be verified against the paper; those lists now retain verified names and use “and others” for the remainder. Six Chinese source notes in infrastructure chapters were unsupported translations or inventions and were removed.