New chapters from the English Web Edition.https://imbrooklyn.dev/books/understanding-llms/en/enhttps://imbrooklyn.dev/books/understanding-llms/en/ch-40/https://imbrooklyn.dev/books/understanding-llms/en/ch-40/Build and test the actual input, its evidence boundaries and its budget.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-41/https://imbrooklyn.dev/books/understanding-llms/en/ch-41/Read JSON, validate a schema, and keep structural success separate from truth.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-42/https://imbrooklyn.dev/books/understanding-llms/en/ch-42/Trace a read-only order proposal through validation, authorization, execution and return.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-43/https://imbrooklyn.dev/books/understanding-llms/en/ch-43/Locate MCP responsibilities and inspect a version-pinned read-only exchange.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-44/https://imbrooklyn.dev/books/understanding-llms/en/ch-44/Compare complete lexical, dense, hybrid and reranked results on the retained FAQ.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-45/https://imbrooklyn.dev/books/understanding-llms/en/ch-45/Follow versioned source lines through retrieval to a cited answer or abstention.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-46/https://imbrooklyn.dev/books/understanding-llms/en/ch-46/Locate pipeline failures and assess one paired improvement with explicit uncertainty.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-47/https://imbrooklyn.dev/books/understanding-llms/en/ch-47/Follow one travel question through explicit states, one bounded model choice, and inspectable stopping rules.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-48/https://imbrooklyn.dev/books/understanding-llms/en/ch-48/Package source verification with a clear contract, selective loading, executable checks, and reviewed provenance.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-49/https://imbrooklyn.dev/books/understanding-llms/en/ch-49/Follow a lost cancellation response through exact approval, durable receipts, bounded retries, and recovery.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-50/https://imbrooklyn.dev/books/understanding-llms/en/ch-50/Locate failures across the whole assistant and calibrate a local Judge against preserved human labels.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-51/https://imbrooklyn.dev/books/understanding-llms/en/ch-51/Follow one policy answer through its data lifecycle and assign checkable responsibilities.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-52/https://imbrooklyn.dev/books/understanding-llms/en/ch-52/Trace local attacks across instruction, evidence, tool and display boundaries, then retest specific controls.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-53/https://imbrooklyn.dev/books/understanding-llms/en/ch-53/Bind identity, approved actions and reviewed dependencies to observable security boundaries.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-54/https://imbrooklyn.dev/books/understanding-llms/en/ch-54/Follow a cited answer through an authenticated local service, cancellation races, failures and an accessible user experience.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-55/https://imbrooklyn.dev/books/understanding-llms/en/ch-55/Identify every dependency of a request, reject a bad local candidate and distinguish replay from recomputation and recovery.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-56/https://imbrooklyn.dev/books/understanding-llms/en/ch-56/Work through cache boundaries, overload, a failing dependency and a counted service promise before turning feedback into reviewed regressions.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-57/https://imbrooklyn.dev/books/understanding-llms/en/ch-57/Accept a bounded knowledge assistant by following real requests, failed cases, permissions, evaluations, and recovery evidence.Tue, 15 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-05/https://imbrooklyn.dev/books/understanding-llms/en/ch-05/Track every axis from embedding rows to batched matrix multiplication, calculate dot products and cosine, and diagnose padding and semantic mismatches.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-06/https://imbrooklyn.dev/books/understanding-llms/en/ch-06/Calculate softmax, target loss, mean loss, perplexity and directional KL, with explicit denominators, units, numerical stability and task boundaries.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-07/https://imbrooklyn.dev/books/understanding-llms/en/ch-07/Follow one complete gradient update, then diagnose a measured overfit on fixed two-dimensional data while keeping training, validation and test roles separate.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-08/https://imbrooklyn.dev/books/understanding-llms/en/ch-08/Distinguish displayed text, Unicode code points, UTF-8 bytes and tokens, then test an explicit bilingual search-key contract and its irreversible losses.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-09/https://imbrooklyn.dev/books/understanding-llms/en/ch-09/Build KA-0 on one frozen bilingual collection, calculate complete count and TF-IDF representations, compare three lexical rankings, and retain their failures.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-10/https://imbrooklyn.dev/books/understanding-llms/en/ch-10/Motivate learned static representations, trace a complete fixed-window neural LM, compare matched retrieval and continuation tasks, and expose missing context.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-11/https://imbrooklyn.dev/books/understanding-llms/en/ch-11/Compute three recurrent steps, follow shared parameters and changing states, and separate an absent clue from an attenuated clue.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-12/https://imbrooklyn.dev/books/understanding-llms/en/ch-12/Trace an error backward through time, compute a shared weight gradient, compare gated state paths, and locate ELMo among contextual representations.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-13/https://imbrooklyn.dev/books/understanding-llms/en/ch-13/Follow an encoder, shift a translation target for training, trace self-fed generation, and separate prefix mismatch from a fixed-vector collision.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-14/https://imbrooklyn.dev/books/understanding-llms/en/ch-14/Calculate two cross-attention reads, inspect source deletion with and without re-encoding, and identify the recurrent dependencies that remain.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-15/https://imbrooklyn.dev/books/understanding-llms/en/ch-15/Follow three positions through learned projections, scaled scores, row probabilities, and a complete value mixture.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-16/https://imbrooklyn.dev/books/understanding-llms/en/ch-16/Specify who may see whom, expose a future-copy failure, and calculate two heads through their output projection.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-17/https://imbrooklyn.dev/books/understanding-llms/en/ch-17/Follow a position through position addition, attention, residual paths, LayerNorm and a shared nonlinear network.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-18/https://imbrooklyn.dev/books/understanding-llms/en/ch-18/Compare visible information and prediction targets on the same sentence, keeping architecture, objective and training stage separate.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-19/https://imbrooklyn.dev/books/understanding-llms/en/ch-19/Train a tiny vocabulary by explicit merges, test text round trips, and freeze the Mini GPT tokenizer interface.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-20/https://imbrooklyn.dev/books/understanding-llms/en/ch-20/Connect raw records, frozen token IDs, every block tensor, vocabulary logits and a checked CPU loss in MG-1.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-21/https://imbrooklyn.dev/books/understanding-llms/en/ch-21/Read an actual CPU training curve, resume the same run, and trace both successful and failed generation in MG-2.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-22/https://imbrooklyn.dev/books/understanding-llms/en/ch-22/Use Mini GPT and a model card to separate scale, reuse, training stage, reasoning behavior and the interface that serves a model.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-23/https://imbrooklyn.dev/books/understanding-llms/en/ch-23/Trace a small corpus through parsing, filtering, deduplication, redaction, mixing and contamination checks while preserving two project Data Cards.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-24/https://imbrooklyn.dev/books/understanding-llms/en/ch-24/Build a transparent training budget, interpret an empirical power law and distinguish compute-optimal training from lifetime deployment cost.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-25/https://imbrooklyn.dev/books/understanding-llms/en/ch-25/Work through a rotation, a normalization and a gated feed-forward calculation, then place all three changes inside one residual block.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-26/https://imbrooklyn.dev/books/understanding-llms/en/ch-26/Locate changes to stored context, token interactions and parameter use without treating them as one upgrade ladder.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-27/https://imbrooklyn.dev/books/understanding-llms/en/ch-27/Account for weights, gradients, optimizer state and activations before deciding how precision and parallelism can help.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-28/https://imbrooklyn.dev/books/understanding-llms/en/ch-28/Use matched interventions to distinguish information present in an activation from information the model uses.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-29/https://imbrooklyn.dev/books/understanding-llms/en/ch-29/Follow one candidate list through score changes, filtering, renormalization, sampling and stopping.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-30/https://imbrooklyn.dev/books/understanding-llms/en/ch-30/Follow a request through its first token, later tokens and growing cache, then count the cost of storing approximate numbers.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-31/https://imbrooklyn.dev/books/understanding-llms/en/ch-31/Compare four different optimization targets on an explicit workload, keeping first-token waits, token pace and throughput separate.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-32/https://imbrooklyn.dev/books/understanding-llms/en/ch-32/Turn a user scenario into fixed cases, scoring rules, baselines and rejection conditions before comparing candidates.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-33/https://imbrooklyn.dev/books/understanding-llms/en/ch-33/Complete KA-1 with a dated, reproducible comparison that can reject both candidates and separate measurements from assumptions.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-34/https://imbrooklyn.dev/books/understanding-llms/en/ch-34/Turn a dialogue into next-token targets and separate instruction training from parameter-efficient updates.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-35/https://imbrooklyn.dev/books/understanding-llms/en/ch-35/Audit preference pairs, calculate a reference-policy penalty, and trace what RLHF and DPO actually update.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-36/https://imbrooklyn.dev/books/understanding-llms/en/ch-36/Use a checkable reimbursement problem to separate generating solutions, checking them, and training from feedback.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-37/https://imbrooklyn.dev/books/understanding-llms/en/ch-37/Compare fixed-parameter answer selection under a declared budget, including rejected candidates, verification and stopping.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-38/https://imbrooklyn.dev/books/understanding-llms/en/ch-38/Use KA-1's retained failures and a real paired runtime comparison to choose which layer to change.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-39/https://imbrooklyn.dev/books/understanding-llms/en/ch-39/Separate training data and objectives from trainable parameters, then calculate a low-rank update and its limits.Sun, 13 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-01/https://imbrooklyn.dev/books/understanding-llms/en/ch-01/Establish task criteria, analyze five controlled conversation cases, and distinguish a language model from the system that supplies evidence and executes actions.Thu, 10 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-02/https://imbrooklyn.dev/books/understanding-llms/en/ch-02/Trace a token through a causal text generator, distinguish the two feedback loops of training and generation, and analyze six static changes to model and system state.Thu, 10 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-03/https://imbrooklyn.dev/books/understanding-llms/en/ch-03/Derive a finite text-encoding interface, execute bilingual tokenization and lookup, analyze information loss and version compatibility, and calculate vector similarity.Thu, 10 Sep 2026 00:00:00 GMThttps://imbrooklyn.dev/books/understanding-llms/en/ch-04/https://imbrooklyn.dev/books/understanding-llms/en/ch-04/Build and analyze count-based language models, derive complete-sequence probabilities, trace reproducible generation, and examine smoothing, context length, and sparse evidence.Thu, 10 Sep 2026 00:00:00 GMT