Changelog

The book is revised continuously. This page records what changed each week, so a reader who finished a chapter last month can see whether it has moved and whether anything new is worth returning for. Corrections are listed as plainly as additions: if a claim was wrong, it says what it said and what it says now.

Weeks with no reader-visible change are not listed. Version tags are noted where they fall.

July 20 to 26, 2026

The whole book was checked against primary sources, which produced 51 verified corrections, and the infrastructure part was split in two after growing to eleven chapters. Released as v0.2.0.

New

• "The framework the developer writes for itself", a section in Law, Regulation, and Policy, on the gate that decides whether a frontier model may be deployed at all: how a capability threshold is set, which evaluations test for it, what safeguards attach when it trips, where the published frameworks diverged during 2026, and what Anthropic's ASL-3 looked like when a threshold actually tripped.
• "The stage the funnel leaves out: who may read this", a section in RAG and Retrieval, answering whether the person asking is allowed to see what the retriever found: access lists carried into the index against a filter applied at query time, what each costs, and why filtering after retrieval quietly shortens the result list for the users with the narrowest permissions.

Changed

• Part IX became two parts. Part IX is now Infrastructure and Compute, eight chapters running from Accelerators and Networking to The Machine That Breaks at Scale, ending on the open question of supply. A new Part X, Frontiers and Limits, holds Where Learning Hits Limits, The Capability Horizon and Its Measurement, and The Verification Frontier. Ecosystem and Economics is now Part XI and Practice and Operations is Part XII. The three moved chapters have new addresses; saved links to the old ones still work and land on the new page in one hop.
• Twelve terms got their plain-language definition back at the point where they first carry weight, among them policy in its reinforcement-learning sense in RLHF and Reward Modeling, chain of thought, held-out set, LLM-as-judge, prompt injection, superposition, attention sink and cross-entropy. Earlier editing passes had cut those definitions as filler. Two formulas, in Verifiers and Process Supervision and in The Software Substrate: Frameworks and Autodiff, now name their symbols in the prose before the math arrives. The quality service-level indicator had been defined from scratch in two chapters of the practice part; the definition now lives in Reliability for Nondeterministic Systems, and Operating Contracts: SLOs, Cost, Incidents, and Tenancy recalls it in a clause and keeps to what it can add.
• On the Chinese pages, sidebar links now appear in Chinese, the quotation that opens each part renders as a pinned panel again, and the labels on the nested-loop figure no longer sit on top of each other.

Corrected

• Tokenization defined fertility as the number of characters carried per token. Fertility is the average number of tokens a word is split into, the reciprocal of what the chapter said, so the argument built on it ran backwards.
• A figure caption in Beyond Dense Transformers: MoE, SSMs, Hybrids called the KV cache and attention's per-step cost quadratic in sequence length. Both are linear. Only the total cost of mixing across a whole sequence is quadratic, which is what the chapter says elsewhere.
• Faster Decoding put DeepSeek-V3's multi-token prediction acceptance rate at 80 percent. The technical report gives 85 to 90 percent.
• Choosing a Model still listed GPT-5.5 as OpenAI's flagship. The table now carries GPT-5.6 Sol, Terra and Luna, with Sol as the flagship, public since 2026-07-09. The long-context price threshold is 272K input tokens, and the chapter said 270K in one place. The interactive picker on the same page was recommending GPT-5.5 for agentic coding, against the chapter's own decision diagram.
• The AI Value Chain said the silicon layer gave up about four points of value-chain share. The table eight lines above it shows 87 percent falling to 79, which is eight points, split between infrastructure and applications.
• Making the Silicon: Packaging, HBM, and the Geopolitics of Compute hung its CoWoS wafers-per-month capacity figures, a lead time, and an elevenfold demand figure on a single TrendForce piece that states no capacity number at all. Each figure now cites a source that carries it, with its date. The lead-time claim is gone, because no primary source states one. NVIDIA's share of CoWoS capacity, previously asserted with no citation, is now the reported share of 2026-27 bookings and of CoWoS-L in 2025.
• Reference summaries were cut off at the first percent sign, so one read as "reduce inter-GPU communication by up to 95" and stopped there. 232 bibliography fields were affected and now display in full. Six citations pointed at URLs that returned 404 and now point at live pages, and two of those carried the wrong author and year.

July 6 to 12, 2026

The reading site changed appearance, with the header and side panels now floating over the text instead of boxing it in. On the content side, the evaluation chapters and the passages about credentials inside sandboxes were reconciled with how model APIs and secret handling actually work in mid-2026. Released as v0.1.0.

New

• Security and Authorization gains egress substitution as a credential mechanism in its own right: the sandbox holds an opaque placeholder, and a TLS-terminating proxy swaps in the real secret on the wire, only toward the hosts that secret is scoped to. The section carries an interactive stepper through three generations of the design and a diagram of the data path.

Changed

• The reader was rebuilt. The header is a floating pill, panels and drawers are translucent over the page, the chapter text sits on its own panel, reading progress is a ring with a percentage where it used to be a bar along the top edge, and sidebar entries light up under the cursor. Scrolling a long chapter is smoother than before.
• The chapters The Harness, Evaluating Agents and Capabilities, Judging and Holistic Evaluation, Human Evaluation and Rubrics, and Operational Evaluation and Governance were reworked around one claim: a score belongs to a whole setup, not to a model. They now spell out what has to be pinned and recorded (served route, harness and its exact version, image tag, the effort configured, sent and observed, prompt hash), why a judge is identified by model id plus rubric hash, and why a small measured difference counts as noise until a test says otherwise.
• The shipped example in the evaluation and orchestration chapters was named agon. That product was retired, and the passages now describe Latere's Adversarial Review capability.
• A failing comments service used to blank the whole page. A chapter now renders with zero comments instead.

Corrected

• The book described per-request thinking budgets as the current way to buy reasoning depth, with a caller-set budget_tokens. Current models reject that shape. Thinking is adaptive under an effort level the model regulates itself, and that level is a ceiling rather than a purchase: at the highest effort an easy question can still spend almost no reasoning tokens. A reader who wired the old field is sending calls that fail.
• Wiring a 2026 Stack said everything speaks the OpenAI request shape. Three incompatible dialects are in use, a gateway between them is a translator that loses information, and OpenAI reasoning tiers are tool-drivable only on the Responses API.
• The Tooling Ecosystem described Latere Cella as injecting the real secret into the sandbox at the moment a command needs it. That mechanism was retired and the real secret no longer enters the box. Agents, Frameworks, and Sandboxes together with Wiring a 2026 Stack also said secrets reach the agent as short-lived scoped tokens, which assumed every credential can be minted per call; a static third-party API key cannot.

June 29 to July 5, 2026

Eight new chapters and three new sections, plus a pass over every part of the book checking its facts against primary sources current to mid-2026. The additions fill gaps the book had been carrying, the largest being the software layer between model code and the accelerator.

New

• "The Infrastructure Before This One" in the orientation part. Machine learning ran the world's feeds, search results, and ad auctions at planetary scale for fifteen years before a language model wrote its first paragraph. The chapter asks what the generative stack inherited from that incumbent, what it could not reuse, and where the two are converging, and it marks the boundary of what this book covers.
• "The Software Substrate: Frameworks and Autodiff" and "Compilers, Kernels, and the CUDA Moat", two chapters on the layer the book had been taking for granted. The first explains why the gradient of a program costs only a small multiple of the program itself, and how the decade-long argument between graph compilers and eager execution ended. The second follows the captured graph down to device code, where moving bytes costs more than doing arithmetic on them, and asks what nineteen years of accumulated software mean for anyone competing with NVIDIA.
• "Confidential Inference: Trusted Execution and Private Serving" in the safety part. When a company sends prompts to a hosted model, "we don't train on your data" is a promise written in a contract and enforced by nothing in the machine. The chapter covers what hardware attestation can enforce instead, what it costs, and what the hardware boundary still cannot hold.
• Two chapters in the agent part. "Computer Use: the GUI as an Action Space" explains why an agent that reasons well still misses a Submit button, and what a screen action costs next to an API call. "Personalization and User Memory" covers where a product's memory of its user actually lives, why every shipped system chose the same place, and why memory rather than model quality is becoming the switching cost of the assistant market.
• Two chapters in the ecosystem part. "The Model as an Artifact: Formats, Distribution, and the Supply Chain" treats a model as a file that has to be serialized, signed, downloaded, and loaded, and covers the attacks along that chain, including backdoors that live in the weights where no scanner will find them. "The Agent Economy: Identity, Delegation, and Machine Payments" covers what happens when the party holding the credential is software.
• Three new sections inside existing chapters: third-party assurance (ISO 42001, the NIST risk framework, evaluation vendors, AI insurance) in "Law, Regulation, and Policy"; the capital layer (the depreciation-schedule dispute, GPU-backed debt, circular vendor deals) in "Compute Markets and Unit Economics"; and the cost model for serving diffusion and video in "Serving Multimodal Models".

Changed

• Every part was rechecked against primary sources current to mid-2026, so numbers, model names, and live debates now reflect this year rather than last. "Benchmarks and Their Discontents" gained an outside audit's finding that roughly 29% of Humanity's Last Exam's text-only chemistry and biology answers are contradicted by peer-reviewed literature, with the maintainers' own review putting the figure at 18%, along with the dispute over undisclosed private leaderboard testing. The provider table in "Choosing a Model" now runs to July 2026 and records that Fable 5 and Mythos 5 returned after their three-week suspension. Debates the book had left open have closed where the evidence closed them: separating prefill from decode onto different hardware is now the default rather than a live argument.
• Terms a software engineer new to machine learning would not know are now explained in the prose at first use, not only in the glossary, across most of the book in both languages. The scaling-law and diffusion chapters were rebuilt with a gentler on-ramp. The "what's contested" box in "Agent Architectures" had lost its formatting and was rendering as an ordinary heading; it is a box again.

Corrected

• "The Whole Stack in One Pass" charged the dense and sparse designs their lifetime compute and drew a crossover point past which the sparse design became cheaper. There is no crossover. Both the training bill and the serving bill scale with activated parameters, so the mixture-of-experts design, DeepSeek-V3 in the worked example, is cheaper at every serving volume. What it pays instead is memory: all 671B parameters stay resident so that each token can activate its 37B.
• The KV-cache calculation in "Serving Multimodal Models" left out the grouped-query divisor, and so overstated the cache one image costs by roughly a factor of four.
• "The Compute Frontier: Bandwidth, Not FLOPs" read GDPval's September 2025 result, 47.6%, as approaching expert parity without reaching it. That reading did not hold: by mid-2026 the best model wins or ties the blind expert comparison on roughly seven tasks in ten. The open question is no longer whether models reach graded parity, but what a won comparison is worth outside the grading room.
• In the Chinese edition, "Training and Fine-tuning in Practice" described Tinker as the case where the distributed GPUs are rented out. The direction was backwards: the training loop is still yours to write, and the GPUs are rented from the service.

June 22 to 28, 2026

Eighteen new chapters landed this week, filling out the parts on reasoning, post-training, evaluation, and the ecosystem. The site also gained accounts: a reader can sign in, comment on a page or on a passage they select, keep private notes, and bookmark a chapter.

New

• Reasoning and Test-Time Compute went from three chapters to seven. "Structured Reasoning as Search" treats a chain of thought as a search problem, "Programs, Solvers, and Symbolic Work" covers handing a step to a tool that cannot get it wrong, "Verifiers and Process Supervision" asks who checks the intermediate steps, and "Reasoning Data and Distillation" covers where reasoning traces come from.
• Post-Training gained "Behavior Specifications and Preference Data" on writing down what a model should be before collecting comparisons, "Verifiable Rewards and Reasoning Transfer" on training against answers a machine can check, and "Safety Tuning and Instruction Hierarchy" on which instruction wins when a system prompt and a user disagree.
• Evaluation gained "Statistical Reliability" on when a benchmark difference is large enough to believe, plus "Human Evaluation and Rubrics", "Factuality, Grounding, and Citations", and "Operational Evaluation and Governance" on what to measure once a system is live.
• Ecosystem and Economics gained "The AI Value Chain" on where the profit in the stack sits, "Adoption and Productivity" on what the deployment evidence shows, and "Data Rights and Compliance Economics" on what training data costs once rights are priced.
• Four chapters arrived on their own: "Mid-Training: Annealing, Domain Bridges, and Context Extension" on the pass between pre-training and post-training, "The Verification Frontier: Proof, Oversight, and Trust After Capability" on trusting output that outruns human checking, "Human Interfaces and Oversight Loops" on where a person sits inside a running system, and "Operating Contracts: SLOs, Cost, Incidents, and Tenancy" on what a stack promises once it is shared.

Changed

• Every part now ends with a summary page. The reflection that used to sit at the bottom of the part opener moved there and was rewritten as prose that hands off to the next part.
• Figures, diagrams, and interactive charts were redrawn wherever they were hard to read: labels and legends stacked on top of curves were moved clear, diagram text can be selected and copied, Chinese labels no longer clip out of their cells, and dark mode no longer renders diagram text in near-invisible ink.
• Both editions went through a wording pass. Terms are glossed in plain language where they first appear instead of only in the glossary, and symbols were reconciled across chapters, so a task of n independent steps now succeeds with probability p^n everywhere rather than p^t in one chapter and p^n in the next.

Corrected

• The book now says which of its worked examples are the author's own products. A note in the front matter names the latere.ai stack (Lux, Cella, Topos, Wallfacer, Lectio) as the author's own work, and "Wiring a 2026 Stack" no longer recommends a self-hosted Python proxy as the default gateway: a throughput test showed it saturating under agent load, so the default is a compiled gateway, with the neutral option and the author's own tied on that benchmark and the latency figure now stated only at low load.
• "Scaling Laws and Compute Allocation" credited the fitted loss form L = E + A/N^α + B/D^β to Kaplan et al., whose fit carried no irreducible-loss term E at all. The form belongs to Rosenfeld et al. and to Chinchilla. The same chapter listed specific causes for the Chinchilla replication's disagreement, averaging the loss instead of summing it, a rounded exponent, a fitting window chosen by eye, none of which the replication claims; that passage is gone.
• In "Non-Autoregressive and Diffusion Language Models" the masked diffusion training objective carried the wrong sign. As printed, the weight α'_t/(1-α_t) is negative, while the prose called it a positive weighted average of cross-entropies. The weight is -α'_t/(1-α_t).
• Two smaller repairs: "Edge and On-Device Deployment" said MobileLLM-350M is forty times smaller than LLaMA-2 7B, and it is twenty; "Multi-Agent Systems" credited the result that debate can answer PSPACE questions to Brown-Cohen et al., and it belongs to Irving, Christiano, and Amodei.

June 19 to 21, 2026

The book started this week and reached its first tagged version, v0.0.1, on June 21. Three days put up the whole spine: sixty-seven chapters across twelve parts, written in English and Chinese, plus the reader they are served in and a first pass of figures, runnable code, and references.

New

• The book itself, twelve parts running from Orientation to Practice and Operations. It follows one arc, the lifecycle of a capability from raw compute to a deployed and governed behavior, and at every layer asks why the piece has the shape it has rather than only what it does. The English and Chinese versions carry the same chapters.
• Generative and Multimodal Architectures, a part covering what the field builds outside next-token prediction: Diffusion and Flow Matching, Non-Autoregressive and Diffusion Language Models, Speech and Realtime Voice, and Multimodal Models: Fusion and Generation.
• Four more chapters where the spine had a hole. Borrowed Ideas: What AI Took From Other Sciences traces compression, reward signals, and thermodynamics back to the fields they came from. Training Agents to Act covers training a model to take actions instead of answering questions. Embeddings and Representation Learning covers how text becomes a vector worth comparing. Edge and On-Device Deployment covers running a model where there is no datacenter.
• Twenty-six interactive figures and runnable Python cells that execute in the page, among them an attention heatmap, a paged-attention allocator, a router that collapses under imbalance, step-throughs of the RLHF, DPO, and R1 recipes, and a calculator for where self-hosting overtakes an API.
• A glossary of 121 terms, linked from the first place each term is used in a chapter, full-text search over the whole book under Cmd+K, and a Further reading list at the end of each chapter built from that chapter's own bibliography, with a note on why each work is there.

Changed

• The book was reorganized into eleven numbered parts plus Orientation, and chapter addresses lost their numbers. A chapter now keeps its link when the reading order changes. Addresses from before the move redirect to the new ones.
• The site moved onto a reader built for this book: a fixed header with foldable navigation and table of contents, reading progress, light and dark themes, and a settings panel for palette, text size, and body font. The book is web-only, since runnable cells and interactive figures do not survive a static file, so there is no PDF or EPUB.
• The Chinese side was rewritten as Chinese rather than carried over from the English sentence by sentence, which changes how most chapters read on that side.

Corrected

• The Chinese RLHF and Reward Modeling chapter announced three things and then listed four. The opening now matches what follows.
• Training Models to Reason said that several verifier-based recipes shrink or drop the KL penalty that RLHF uses as a leash. That overstated the evidence. Only DAPO drops it, and the text now says so and names what it relies on instead.
• Beyond Text: Multimodal, World Models, and Embodiment opened as though the reader had already met the data wall in a chapter that, after the reorganization, comes seven parts later. The data wall is now explained where it is first used. A Field Map and How to Read This Book still described a nine-part book after the same move, and now describes the one that exists.
• Three faults in the references. The link to "The Era of Experience" was broken by an unencoded space and never opened. Several cited works listed co-authors that could not be checked against the paper, and those lists are now trimmed to the names that could, with the rest marked as others. Six Chinese notes in the infrastructure chapters had been invented or translated rather than carried from the source, and are gone.