mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-11-16 12:05:12 -06:00
2923 lines
89 KiB
JavaScript
2923 lines
89 KiB
JavaScript
// AUTO-GENERATED by socratiq/scripts/sync-taxonomy.mjs from taxonomy/framework-data.json
|
|
// Do not hand-edit without re-running: node scripts/sync-taxonomy.mjs
|
|
|
|
function attachLookupHelpers(record) {
|
|
Object.defineProperties(record, {
|
|
get: { value: (key) => record[key], enumerable: false },
|
|
has: { value: (key) => Object.prototype.hasOwnProperty.call(record, key), enumerable: false },
|
|
size: { get: () => Object.keys(record).length, enumerable: false },
|
|
});
|
|
return Object.freeze(record);
|
|
}
|
|
|
|
export const TAXONOMY_GROUPS = attachLookupHelpers({
|
|
"F": {
|
|
"id": "foundations",
|
|
"code": "F",
|
|
"name": "Foundations",
|
|
"question": "What does the system compute, what must it achieve, and how do we reason about its cost?",
|
|
"color": "#2563eb"
|
|
},
|
|
"S": {
|
|
"id": "stack",
|
|
"code": "S",
|
|
"name": "Stack",
|
|
"question": "What does the system run on, from silicon to the fleet?",
|
|
"color": "#7c3aed"
|
|
},
|
|
"P": {
|
|
"id": "lifecycle",
|
|
"code": "P",
|
|
"name": "Lifecycle",
|
|
"question": "What do engineers do to build, ship, and keep the system working?",
|
|
"color": "#059669"
|
|
},
|
|
"Q": {
|
|
"id": "qualities",
|
|
"code": "Q",
|
|
"name": "Qualities",
|
|
"question": "By which properties is the system judged, and how are they engineered?",
|
|
"color": "#d97706"
|
|
},
|
|
"T": {
|
|
"id": "system-types",
|
|
"code": "T",
|
|
"name": "System types",
|
|
"question": "What is distinctive when the system acts as an agent or in the physical world?",
|
|
"color": "#dc2626"
|
|
}
|
|
});
|
|
|
|
export const TAXONOMY_SUBJECTS = attachLookupHelpers({
|
|
"F1.1": {
|
|
"code": "F1.1",
|
|
"name": "Requirements & SLOs",
|
|
"group": "F",
|
|
"areaCode": "F1",
|
|
"areaName": "Workloads & requirements",
|
|
"summary": "Stating measurable targets for a system's behavior and service, and the baselines and budgets against which they are judged. (Requirements, SLOs & baselines) Includes: Acceptance criteria and baselines; SLIs, SLOs, error budgets; Latency budgets across a request path.",
|
|
"chapters": [
|
|
"vol1/01_introduction",
|
|
"vol1/03_ml_lifecycle",
|
|
"vol1/12_benchmarking",
|
|
"vol1/13_model_serving",
|
|
"vol2/12_ops_scale",
|
|
"vol2/10_inference"
|
|
]
|
|
},
|
|
"F1.2": {
|
|
"code": "F1.2",
|
|
"name": "Workload characterization",
|
|
"group": "F",
|
|
"areaCode": "F1",
|
|
"areaName": "Workloads & requirements",
|
|
"summary": "Describing a workload's compute, memory, data, and traffic demands, and where it runs, so it can be engineered and provisioned for. Includes: FLOP, byte, and memory accounting; Deployment tiers from cloud to TinyML; Workload traces.",
|
|
"chapters": [
|
|
"vol1/02_system_paradigms",
|
|
"vol1/05_nn_computation",
|
|
"vol1/06_nn_architectures",
|
|
"vol2/01_introduction",
|
|
"vol1/13_model_serving"
|
|
]
|
|
},
|
|
"F1.3": {
|
|
"code": "F1.3",
|
|
"name": "Objectives & decisions",
|
|
"group": "F",
|
|
"areaCode": "F1",
|
|
"areaName": "Workloads & requirements",
|
|
"summary": "How learned outputs become decisions, how optimized proxies relate to intended outcomes, and how competing objectives are traded. (Objectives, proxies & decision policies) Includes: Proxy objectives and reward hacking; Cost-sensitive thresholds; Multi-objective tradeoffs.",
|
|
"chapters": [
|
|
"vol1/03_ml_lifecycle",
|
|
"vol1/15_responsible_engr",
|
|
"vol3/17_rlvr",
|
|
"vol2/14_robust_ai"
|
|
]
|
|
},
|
|
"F2.1": {
|
|
"code": "F2.1",
|
|
"name": "Model families",
|
|
"group": "F",
|
|
"areaCode": "F2",
|
|
"areaName": "Models & architectures",
|
|
"summary": "The classical and neural model regimes, with their inductive assumptions, viewed as workloads with characteristic compute and memory profiles. Includes: Classical models and neural networks; Transformers, recommendation models, diffusion models.",
|
|
"chapters": [
|
|
"vol1/05_nn_computation",
|
|
"vol1/06_nn_architectures"
|
|
]
|
|
},
|
|
"F2.2": {
|
|
"code": "F2.2",
|
|
"name": "Building blocks",
|
|
"group": "F",
|
|
"areaCode": "F2",
|
|
"areaName": "Models & architectures",
|
|
"summary": "The layer-level operations from which neural models are composed and whose shapes determine compute, memory, and data movement. (Architecture building blocks) Includes: Embeddings, attention, position encoding, tokenization; Convolution, recurrence, normalization.",
|
|
"chapters": [
|
|
"vol1/06_nn_architectures",
|
|
"vol3/02_foundation_model",
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol3/05_kv_cache"
|
|
]
|
|
},
|
|
"F2.3": {
|
|
"code": "F2.3",
|
|
"name": "Conditional computation",
|
|
"group": "F",
|
|
"areaCode": "F2",
|
|
"areaName": "Models & architectures",
|
|
"summary": "Model designs in which the computation performed varies per input, such as mixture-of-experts routing and early exit. (Sparse & conditional computation) Includes: Mixture of experts, routing, capacity, load balancing; Active vs. total parameters; Early exit.",
|
|
"chapters": [
|
|
"vol1/06_nn_architectures",
|
|
"vol1/10_model_compression",
|
|
"vol2/06_distributed_training"
|
|
]
|
|
},
|
|
"F2.4": {
|
|
"code": "F2.4",
|
|
"name": "Long-context models",
|
|
"group": "F",
|
|
"areaCode": "F2",
|
|
"areaName": "Models & architectures",
|
|
"summary": "Sequence architectures that change how cost scales with context length, co-designed with the memory and compute they demand. (Long-context & efficient sequence models) Includes: Sparse, sliding-window, and linear attention; State-space models; grouped-query attention; Context-length extension.",
|
|
"chapters": [
|
|
"vol1/06_nn_architectures",
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol3/05_kv_cache",
|
|
"vol3/04_context_engineering"
|
|
]
|
|
},
|
|
"F2.5": {
|
|
"code": "F2.5",
|
|
"name": "Model artifacts",
|
|
"group": "F",
|
|
"areaCode": "F2",
|
|
"areaName": "Models & architectures",
|
|
"summary": "The weights, graphs, metadata, and interface contracts that make a trained model a portable, loadable, deployable artifact. (Model artifacts & formats) Includes: Weight serialization and exchange formats; Model export and signatures.",
|
|
"chapters": [
|
|
"vol1/07_frameworks",
|
|
"vol1/13_model_serving",
|
|
"vol1/14_ml_ops",
|
|
"vol2/04_data_storage",
|
|
"vol3/11_durable_execution",
|
|
"vol4/16_release"
|
|
]
|
|
},
|
|
"F3.1": {
|
|
"code": "F3.1",
|
|
"name": "Performance models",
|
|
"group": "F",
|
|
"areaCode": "F3",
|
|
"areaName": "System models",
|
|
"summary": "Bound-based reasoning about whether computation is limited by arithmetic or data movement, and how reuse changes that limit. (Roofline, performance models & simulation) Includes: Ridge point and compute- vs. memory-bound regimes; Energy cost of data movement.",
|
|
"chapters": [
|
|
"vol1/01_introduction",
|
|
"vol1/06_nn_architectures",
|
|
"vol1/11_hw_acceleration",
|
|
"vol1/02_system_paradigms",
|
|
"vol4/03_brain"
|
|
]
|
|
},
|
|
"F3.2": {
|
|
"code": "F3.2",
|
|
"name": "Speedup limits",
|
|
"group": "F",
|
|
"areaCode": "F3",
|
|
"areaName": "System models",
|
|
"summary": "Models of how much speedup added parallel or accelerated resources can deliver, given serial work and fixed or growing problem sizes. (Amdahl's law & scaling efficiency) Includes: Amdahl's law and serial fractions; Strong and weak scaling; Parallel efficiency.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol2/01_introduction",
|
|
"vol2/06_distributed_training",
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/05_collective_communication"
|
|
]
|
|
},
|
|
"F3.3": {
|
|
"code": "F3.3",
|
|
"name": "Queueing models",
|
|
"group": "F",
|
|
"areaCode": "F3",
|
|
"areaName": "System models",
|
|
"summary": "Models relating arrivals, service rates, utilization, and waiting, used to predict latency and capacity under load. (Queueing & capacity models) Includes: Little's law; Latency-throughput curves; Burstiness and variability.",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol3/05_kv_cache"
|
|
]
|
|
},
|
|
"F3.4": {
|
|
"code": "F3.4",
|
|
"name": "Scaling laws",
|
|
"group": "F",
|
|
"areaCode": "F3",
|
|
"areaName": "System models",
|
|
"summary": "Empirical relationships among parameters, data, and compute and observed quality, used to allocate training budgets. (Scaling laws & compute-optimal allocation) Includes: Loss scaling laws; Compute-optimal model/data allocation; Training compute estimates.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol1/09_data_selection",
|
|
"vol2/01_introduction"
|
|
]
|
|
},
|
|
"S1.1": {
|
|
"code": "S1.1",
|
|
"name": "Accelerator architecture",
|
|
"group": "S",
|
|
"areaCode": "S1",
|
|
"areaName": "Hardware & accelerators",
|
|
"summary": "The organization of processors specialized for ML: execution units, threading, dataflow, and the throughput they provide. Includes: GPUs, TPUs, NPUs; SIMT and matrix engines; Systolic arrays and dataflows; Domain-specific architectures.",
|
|
"chapters": [
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/02_compute_infrastructure",
|
|
"vol1/05_nn_computation"
|
|
]
|
|
},
|
|
"S1.2": {
|
|
"code": "S1.2",
|
|
"name": "Memory systems",
|
|
"group": "S",
|
|
"areaCode": "S1",
|
|
"areaName": "Hardware & accelerators",
|
|
"summary": "The storage levels of an ML processor, from registers and SRAM to HBM, and the capacity and bandwidth limits they impose. (Memory hierarchy & bandwidth) Includes: On-chip SRAM and scratchpads; HBM; The memory wall.",
|
|
"chapters": [
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/02_compute_infrastructure",
|
|
"vol1/02_system_paradigms",
|
|
"vol4/03_brain"
|
|
]
|
|
},
|
|
"S1.3": {
|
|
"code": "S1.3",
|
|
"name": "Node interconnect",
|
|
"group": "S",
|
|
"areaCode": "S1",
|
|
"areaName": "Hardware & accelerators",
|
|
"summary": "Links that move data between hosts and accelerators and among accelerators within a node or rack-scale domain. (Host-device & scale-up interconnect) Includes: PCIe host links; NVLink-class scale-up fabrics; CXL and peer-to-peer transfers.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol2/02_compute_infrastructure",
|
|
"vol2/03_network_fabrics",
|
|
"vol2/05_collective_communication"
|
|
]
|
|
},
|
|
"S1.4": {
|
|
"code": "S1.4",
|
|
"name": "Chip power & thermal",
|
|
"group": "S",
|
|
"areaCode": "S1",
|
|
"areaName": "Hardware & accelerators",
|
|
"summary": "Electrical and thermal limits at the chip and board level, and the runtime controls that keep execution within them. (Chip power & thermal limits) Includes: TDP and thermal throttling; DVFS and power capping; Dark silicon.",
|
|
"chapters": [
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/02_compute_infrastructure",
|
|
"vol4/02_body",
|
|
"vol4/13_placement",
|
|
"vol2/15_sustainable_ai"
|
|
]
|
|
},
|
|
"S1.5": {
|
|
"code": "S1.5",
|
|
"name": "Edge & embedded hardware",
|
|
"group": "S",
|
|
"areaCode": "S1",
|
|
"areaName": "Hardware & accelerators",
|
|
"summary": "Hardware for on-device ML, from mobile SoCs to microcontrollers, where memory and energy per inference bind. Includes: Microcontrollers and TinyML devices; Heterogeneous mobile SoCs and NPUs; Energy and memory budgets.",
|
|
"chapters": [
|
|
"vol1/02_system_paradigms",
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/11_edge_intelligence",
|
|
"vol4/13_placement",
|
|
"vol1/10_model_compression"
|
|
]
|
|
},
|
|
"S2.1": {
|
|
"code": "S2.1",
|
|
"name": "Programming models",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "The abstractions with which ML programs are written: tensors, computation graphs, execution modes, and automatic differentiation. (ML frameworks & programming models) Includes: Tensors and computation graphs; Eager vs. graph execution; tracing JIT; Automatic differentiation.",
|
|
"chapters": [
|
|
"vol1/07_frameworks"
|
|
]
|
|
},
|
|
"S2.2": {
|
|
"code": "S2.2",
|
|
"name": "Tensor operators",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "Implementations of core tensor operations, such as matrix multiply, convolution, and attention, and the layouts that govern their efficiency. (Tensor operators & memory layout) Includes: GEMM and convolution lowering; IO-aware attention kernels; Vendor kernel libraries.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol1/11_hw_acceleration",
|
|
"vol1/07_frameworks",
|
|
"vol1/05_nn_computation"
|
|
]
|
|
},
|
|
"S2.3": {
|
|
"code": "S2.3",
|
|
"name": "GPU programming",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "How kernels are written for parallel accelerators: the thread hierarchy, memory access patterns, occupancy, asynchronous data movement, and programming the matrix units. (Kernel & GPU programming model) Includes: Writing and tuning kernels in CUDA-class or tile-level languages such as Triton; Shared memory, synchronization, and latency hiding.",
|
|
"chapters": [
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/09_performance_engineering"
|
|
]
|
|
},
|
|
"S2.4": {
|
|
"code": "S2.4",
|
|
"name": "ML compilers",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "Transforming model programs through intermediate representations into optimized kernels for target hardware. (ML compilers & kernel optimization) Includes: Graph and loop IRs; lowering; Fusion, tiling, autotuning, kernel DSLs; Static memory planning.",
|
|
"chapters": [
|
|
"vol1/07_frameworks",
|
|
"vol1/11_hw_acceleration",
|
|
"vol2/09_performance_engineering"
|
|
]
|
|
},
|
|
"S2.5": {
|
|
"code": "S2.5",
|
|
"name": "Runtime & dispatch",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "How a runtime launches and orders device work and manages device memory during execution. (Runtime execution & dispatch) Includes: Launch overhead and graph replay; Streams and asynchronous execution; Caching allocators.",
|
|
"chapters": [
|
|
"vol1/07_frameworks",
|
|
"vol1/13_model_serving",
|
|
"vol2/09_performance_engineering",
|
|
"vol2/05_collective_communication"
|
|
]
|
|
},
|
|
"S2.6": {
|
|
"code": "S2.6",
|
|
"name": "Data loading",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "Delivering training and inference inputs to accelerators fast enough: host-side loading and preprocessing, prefetching, and sampling minibatches from large or irregular data such as graphs. (Data loading & input pipelines) Includes: Data-loader workers, decoding and augmentation on the host, pinned memory, and prefetching; Minibatch construction for irregular data, such as graph neighborhood sampling.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol2/04_data_storage",
|
|
"vol1/04_data_engineering"
|
|
]
|
|
},
|
|
"S2.7": {
|
|
"code": "S2.7",
|
|
"name": "Numerics",
|
|
"group": "S",
|
|
"areaCode": "S2",
|
|
"areaName": "Frameworks & compilers",
|
|
"summary": "Number formats used in ML computation and the techniques that keep training and inference accurate at reduced precision. (Numerical formats & precision) Includes: FP32/BF16/FP8/INT formats and block scaling; Mixed precision and loss scaling; Accumulation precision.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol1/10_model_compression",
|
|
"vol2/06_distributed_training"
|
|
]
|
|
},
|
|
"S3.1": {
|
|
"code": "S3.1",
|
|
"name": "Parallelism",
|
|
"group": "S",
|
|
"areaCode": "S3",
|
|
"areaName": "Parallelism & distribution",
|
|
"summary": "Ways of partitioning a model's computation and state across devices, and how those dimensions are combined. (Parallelism strategies) Includes: Data, tensor, pipeline, sequence, expert parallelism; ZeRO/FSDP state sharding; Hybrid 3D layouts and pipeline bubbles.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol2/06_distributed_training",
|
|
"vol2/10_inference",
|
|
"vol1/05_nn_computation",
|
|
"vol2/05_collective_communication"
|
|
]
|
|
},
|
|
"S3.2": {
|
|
"code": "S3.2",
|
|
"name": "Parallelism planning",
|
|
"group": "S",
|
|
"areaCode": "S3",
|
|
"areaName": "Parallelism & distribution",
|
|
"summary": "Choosing and changing how a model is partitioned across devices: describing devices as a mesh, searching or propagating sharding decisions, placing experts and stages on heterogeneous hardware, and resharding at run time. (Parallelization planning & placement) Includes: Automatic parallelization and sharding propagation; Device meshes, expert placement, and dynamic resharding.",
|
|
"chapters": [
|
|
"vol2/06_distributed_training",
|
|
"vol2/10_inference"
|
|
]
|
|
},
|
|
"S3.3": {
|
|
"code": "S3.3",
|
|
"name": "Collectives",
|
|
"group": "S",
|
|
"areaCode": "S3",
|
|
"areaName": "Parallelism & distribution",
|
|
"summary": "Group communication primitives and algorithms used by parallel ML, their cost models, and techniques to reduce or hide their cost. (Collective communication) Includes: All-reduce, all-gather, all-to-all; Ring and tree algorithms; alpha-beta cost model; Overlap and gradient compression.",
|
|
"chapters": [
|
|
"vol2/05_collective_communication",
|
|
"vol2/06_distributed_training",
|
|
"vol2/03_network_fabrics"
|
|
]
|
|
},
|
|
"S3.4": {
|
|
"code": "S3.4",
|
|
"name": "Coordination",
|
|
"group": "S",
|
|
"areaCode": "S3",
|
|
"areaName": "Parallelism & distribution",
|
|
"summary": "Mechanisms by which distributed components agree on authoritative control state and tolerate stale or failed participants. (Coordination & consensus) Includes: Consensus and leader election; Leases and fencing; Bounded-staleness consistency.",
|
|
"chapters": [
|
|
"vol2/06_distributed_training",
|
|
"vol2/07_fault_tolerance",
|
|
"vol3/11_durable_execution",
|
|
"vol3/18_multi_agent",
|
|
"vol2/04_data_storage",
|
|
"vol3/07_tool_calling",
|
|
"vol4/04_nervous"
|
|
]
|
|
},
|
|
"S3.5": {
|
|
"code": "S3.5",
|
|
"name": "Federated learning",
|
|
"group": "S",
|
|
"areaCode": "S3",
|
|
"areaName": "Parallelism & distribution",
|
|
"summary": "Training across devices or organizations that keep their data local, coordinated through a shared model. (Federated & decentralized learning) Includes: Federated averaging; Client heterogeneity and selection; Cross-device vs. cross-silo.",
|
|
"chapters": [
|
|
"vol1/15_responsible_engr",
|
|
"vol2/11_edge_intelligence",
|
|
"vol2/13_security_privacy"
|
|
]
|
|
},
|
|
"S4.1": {
|
|
"code": "S4.1",
|
|
"name": "Datacenter facilities",
|
|
"group": "S",
|
|
"areaCode": "S4",
|
|
"areaName": "Datacenter & fleet",
|
|
"summary": "Engineering of the buildings that house ML fleets: power provisioning, cooling, density, and facility efficiency. (Datacenter power, cooling & facilities) Includes: Facility power and oversubscription; Air and liquid cooling; rack density; PUE.",
|
|
"chapters": [
|
|
"vol2/02_compute_infrastructure",
|
|
"vol2/15_sustainable_ai",
|
|
"vol1/15_responsible_engr",
|
|
"vol4/02_body",
|
|
"vol4/13_placement"
|
|
]
|
|
},
|
|
"S4.2": {
|
|
"code": "S4.2",
|
|
"name": "Network fabrics",
|
|
"group": "S",
|
|
"areaCode": "S4",
|
|
"areaName": "Datacenter & fleet",
|
|
"summary": "Datacenter networks that connect accelerator nodes: transports, topologies, flow and congestion control, and in-network compute. Includes: RDMA over InfiniBand and RoCE; lossless flow control; Fat-tree, rail-optimized, torus topologies; Congestion control; in-network reduction.",
|
|
"chapters": [
|
|
"vol2/03_network_fabrics",
|
|
"vol2/05_collective_communication"
|
|
]
|
|
},
|
|
"S4.3": {
|
|
"code": "S4.3",
|
|
"name": "Storage & data delivery",
|
|
"group": "S",
|
|
"areaCode": "S4",
|
|
"areaName": "Datacenter & fleet",
|
|
"summary": "Storage systems that hold training data, checkpoints, and artifacts at scale and deliver them to accelerators fast enough. Includes: Parallel file systems and object stores; Input pipelines and data stalls; Checkpoint I/O and storage tiering.",
|
|
"chapters": [
|
|
"vol1/04_data_engineering",
|
|
"vol2/04_data_storage",
|
|
"vol3/11_durable_execution"
|
|
]
|
|
},
|
|
"S4.4": {
|
|
"code": "S4.4",
|
|
"name": "Cluster scheduling",
|
|
"group": "S",
|
|
"areaCode": "S4",
|
|
"areaName": "Datacenter & fleet",
|
|
"summary": "Allocating a shared cluster's accelerators to jobs over time: placement, gang allocation, fairness, preemption, and elasticity. (Cluster scheduling & orchestration) Includes: Gang scheduling and backfill; Fair sharing, preemption, elastic jobs; Topology-aware placement and bin packing.",
|
|
"chapters": [
|
|
"vol2/08_fleet_orchestration",
|
|
"vol1/13_model_serving",
|
|
"vol2/04_data_storage",
|
|
"vol3/11_durable_execution"
|
|
]
|
|
},
|
|
"S4.5": {
|
|
"code": "S4.5",
|
|
"name": "Capacity & multi-tenancy",
|
|
"group": "S",
|
|
"areaCode": "S4",
|
|
"areaName": "Datacenter & fleet",
|
|
"summary": "Provisioning fleet capacity and sharing it among tenants through quotas, accelerator sharing, and interference control. (Capacity planning & multi-tenancy) Includes: Capacity planning and quotas; Accelerator sharing and partitioning; Spot capacity and interference.",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol2/08_fleet_orchestration",
|
|
"vol2/10_inference",
|
|
"vol2/12_ops_scale",
|
|
"vol3/08_sandboxes",
|
|
"vol4/13_placement",
|
|
"vol3/19_agent_economics"
|
|
]
|
|
},
|
|
"P1.1": {
|
|
"code": "P1.1",
|
|
"name": "Collection & labeling",
|
|
"group": "P",
|
|
"areaCode": "P1",
|
|
"areaName": "Data engineering",
|
|
"summary": "Obtaining observations and supervisory targets whose sampling, coverage, and label quality are known and fit the task. (Data collection & labeling) Includes: Sampling design, human and programmatic labeling, labeling budgets.; Annotation quality and agreement..",
|
|
"chapters": [
|
|
"vol1/04_data_engineering",
|
|
"vol3/15_trajectory_curation",
|
|
"vol4/05_data"
|
|
]
|
|
},
|
|
"P1.2": {
|
|
"code": "P1.2",
|
|
"name": "Selection & curation",
|
|
"group": "P",
|
|
"areaCode": "P1",
|
|
"areaName": "Data engineering",
|
|
"summary": "Choosing, filtering, deduplicating, mixing, augmenting, and synthesizing training examples to maximize learning value per unit of compute and labeling cost. (Data selection & curation) Includes: Web-scale corpus filtering, deduplication, and domain mixtures.; Augmentation and synthetic or model-generated data.; Trajectory and demonstration curation for agents and robots..",
|
|
"chapters": [
|
|
"vol1/09_data_selection",
|
|
"vol3/15_trajectory_curation",
|
|
"vol4/05_data",
|
|
"vol2/01_introduction"
|
|
]
|
|
},
|
|
"P1.3": {
|
|
"code": "P1.3",
|
|
"name": "Validation & features",
|
|
"group": "P",
|
|
"areaCode": "P1",
|
|
"areaName": "Data engineering",
|
|
"summary": "Checking that recorded data is fit for use and transforming it into model inputs whose semantics and availability match training and serving. (Data validation & feature engineering) Includes: Schema and distribution checks before training and serving.; Feature construction and shared feature stores..",
|
|
"chapters": [
|
|
"vol1/04_data_engineering",
|
|
"vol2/12_ops_scale",
|
|
"vol1/03_ml_lifecycle"
|
|
]
|
|
},
|
|
"P1.4": {
|
|
"code": "P1.4",
|
|
"name": "Pipelines & formats",
|
|
"group": "P",
|
|
"areaCode": "P1",
|
|
"areaName": "Data engineering",
|
|
"summary": "Engineering ingestion, batch and stream transformation, event-time semantics, storage formats, and schemas for training and serving data. (Data pipelines & storage formats) Includes: ETL/ELT, batch and stream processing, columnar formats.; Data loaders that keep accelerators fed..",
|
|
"chapters": [
|
|
"vol1/04_data_engineering",
|
|
"vol2/04_data_storage",
|
|
"vol2/12_ops_scale"
|
|
]
|
|
},
|
|
"P1.5": {
|
|
"code": "P1.5",
|
|
"name": "Contracts & lineage",
|
|
"group": "P",
|
|
"areaCode": "P1",
|
|
"areaName": "Data engineering",
|
|
"summary": "Specifying, versioning, and tracing the datasets a system consumes so that dependencies, provenance, and changes are explicit. (Data contracts, versioning & lineage) Includes: Dataset versions, snapshots, and provenance records.; Producer-consumer contracts and dependency hygiene..",
|
|
"chapters": [
|
|
"vol1/04_data_engineering",
|
|
"vol2/12_ops_scale",
|
|
"vol4/05_data",
|
|
"vol3/06_long_term_memory",
|
|
"vol2/04_data_storage",
|
|
"vol4/16_release",
|
|
"vol1/03_ml_lifecycle"
|
|
]
|
|
},
|
|
"P1.6": {
|
|
"code": "P1.6",
|
|
"name": "Leakage & skew",
|
|
"group": "P",
|
|
"areaCode": "P1",
|
|
"areaName": "Data engineering",
|
|
"summary": "Ensuring examples use only information available at prediction time and that training and serving compute features identically and equally fresh. (Leakage & training-serving skew) Includes: Point-in-time joins and temporal leakage.; Feature staleness and skew between offline and online paths..",
|
|
"chapters": [
|
|
"vol1/04_data_engineering",
|
|
"vol2/12_ops_scale",
|
|
"vol3/15_trajectory_curation"
|
|
]
|
|
},
|
|
"P2.1": {
|
|
"code": "P2.1",
|
|
"name": "Learning paradigms",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "The ways supervision, self-generated targets, interaction, and demonstration produce learning signals, each with distinct data and compute demands. Includes: Supervised, self-supervised, unsupervised, reinforcement, and imitation learning..",
|
|
"chapters": [
|
|
"vol1/05_nn_computation",
|
|
"vol1/09_data_selection",
|
|
"vol4/06_training",
|
|
"vol2/11_edge_intelligence",
|
|
"vol3/16_fine_tuning",
|
|
"vol4/05_data"
|
|
]
|
|
},
|
|
"P2.2": {
|
|
"code": "P2.2",
|
|
"name": "Optimization & dynamics",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "The objectives, update rules, schedules, and batch regimes that determine whether and how efficiently training converges. (Optimization & training dynamics) Includes: Optimizers, learning-rate warmup and decay, large-batch scaling.; Loss spikes, divergence, and their mitigations..",
|
|
"chapters": [
|
|
"vol1/05_nn_computation",
|
|
"vol1/08_training",
|
|
"vol2/06_distributed_training"
|
|
]
|
|
},
|
|
"P2.3": {
|
|
"code": "P2.3",
|
|
"name": "Training memory",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "Accounting for and reducing the parameters, gradients, optimizer state, and activations a training step must hold. (Training memory & state) Includes: Activation recomputation, gradient accumulation, CPU/NVMe offload..",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol2/06_distributed_training",
|
|
"vol1/05_nn_computation",
|
|
"vol1/02_system_paradigms",
|
|
"vol2/02_compute_infrastructure",
|
|
"vol2/04_data_storage",
|
|
"vol3/11_durable_execution"
|
|
]
|
|
},
|
|
"P2.4": {
|
|
"code": "P2.4",
|
|
"name": "Experiments & tuning",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "Allocating a finite budget of training runs across configurations, and recording runs so their outcomes can be compared and reproduced. (Hyperparameter search & experiment management) Includes: Random, Bayesian, and multi-fidelity search; small-scale proxies.; Run metadata and experiment tracking..",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol1/14_ml_ops",
|
|
"vol2/01_introduction"
|
|
]
|
|
},
|
|
"P2.5": {
|
|
"code": "P2.5",
|
|
"name": "Adaptation & prompting",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "Adapting a pretrained model to a new task or behavior, by continued training of all or part of its parameters or by prompting without parameter updates. (Adaptation: fine-tuning, prompting & transfer) Includes: Instruction and supervised fine-tuning, adapters, LoRA, and model merging.; Prompt engineering and in-context learning without parameter updates..",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol2/11_edge_intelligence",
|
|
"vol3/16_fine_tuning",
|
|
"vol1/10_model_compression"
|
|
]
|
|
},
|
|
"P2.6": {
|
|
"code": "P2.6",
|
|
"name": "RL post-training",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "Training pretrained models against preference, reward-model, or verifiable-reward signals, and the rollout, scoring, and weight-synchronization systems that loop requires. (Preference & RL post-training) Includes: RLHF, DPO-style preference optimization, RL with verifiable rewards.; Rollout generation, reward scoring, colocated vs disaggregated actor/learner, weight sync..",
|
|
"chapters": [
|
|
"vol2/06_distributed_training",
|
|
"vol3/16_fine_tuning",
|
|
"vol3/17_rlvr",
|
|
"vol3/03_test_time_compute"
|
|
]
|
|
},
|
|
"P2.7": {
|
|
"code": "P2.7",
|
|
"name": "Continual learning",
|
|
"group": "P",
|
|
"areaCode": "P2",
|
|
"areaName": "Training & adaptation",
|
|
"summary": "Updating deployed models incrementally from new or streaming data while retaining prior competence. (Continual & online learning) Includes: Online and incremental updates, forgetting and replay.; Personalization and adaptation on device..",
|
|
"chapters": [
|
|
"vol2/11_edge_intelligence",
|
|
"vol1/14_ml_ops",
|
|
"vol2/14_robust_ai",
|
|
"vol4/05_data"
|
|
]
|
|
},
|
|
"P3.1": {
|
|
"code": "P3.1",
|
|
"name": "Evaluation design",
|
|
"group": "P",
|
|
"areaCode": "P3",
|
|
"areaName": "Evaluation & testing",
|
|
"summary": "Constructing offline splits and online experiments whose comparisons support the conclusions drawn from them. (Evaluation design & experiments) Includes: Holdout, temporal, and cross-validation splits.; A/B tests, sequential monitoring, interference between units..",
|
|
"chapters": [
|
|
"vol1/03_ml_lifecycle",
|
|
"vol2/12_ops_scale",
|
|
"vol4/07_evaluation",
|
|
"vol1/04_data_engineering",
|
|
"vol3/15_trajectory_curation",
|
|
"vol3/16_fine_tuning"
|
|
]
|
|
},
|
|
"P3.2": {
|
|
"code": "P3.2",
|
|
"name": "Metrics & uncertainty",
|
|
"group": "P",
|
|
"areaCode": "P3",
|
|
"areaName": "Evaluation & testing",
|
|
"summary": "Specifying what is measured and quantifying how far estimates and predicted probabilities can be trusted. (Metrics, calibration & uncertainty) Includes: Task metrics, calibration, confidence intervals, abstention.; Distribution-free prediction sets..",
|
|
"chapters": [
|
|
"vol1/12_benchmarking",
|
|
"vol2/14_robust_ai",
|
|
"vol3/13_evaluation",
|
|
"vol4/07_evaluation",
|
|
"vol4/10_intent"
|
|
]
|
|
},
|
|
"P3.3": {
|
|
"code": "P3.3",
|
|
"name": "Slices & robustness",
|
|
"group": "P",
|
|
"areaCode": "P3",
|
|
"areaName": "Evaluation & testing",
|
|
"summary": "Measuring behavior separately across populations and conditions, and under deliberate shift or perturbation, rather than only in aggregate. (Sliced & robustness evaluation) Includes: Subgroup and condition slices, shift and perturbation tests..",
|
|
"chapters": [
|
|
"vol2/14_robust_ai",
|
|
"vol2/16_responsible_ai",
|
|
"vol4/15_verification",
|
|
"vol1/14_ml_ops",
|
|
"vol4/05_data",
|
|
"vol1/15_responsible_engr",
|
|
"vol1/12_benchmarking"
|
|
]
|
|
},
|
|
"P3.4": {
|
|
"code": "P3.4",
|
|
"name": "Generative & agent eval",
|
|
"group": "P",
|
|
"areaCode": "P3",
|
|
"areaName": "Evaluation & testing",
|
|
"summary": "Assessing open-ended outputs and multi-step runs where no single reference answer exists, using graders, sampling-aware metrics, and contamination controls. (Evaluating generative & agentic systems) Includes: Model graders, pairwise and human preference, pass@k.; Contamination of benchmarks by training data..",
|
|
"chapters": [
|
|
"vol3/03_test_time_compute",
|
|
"vol3/13_evaluation"
|
|
]
|
|
},
|
|
"P3.5": {
|
|
"code": "P3.5",
|
|
"name": "ML testing",
|
|
"group": "P",
|
|
"areaCode": "P3",
|
|
"areaName": "Evaluation & testing",
|
|
"summary": "Testing data, models, pipelines, and their integration against stated expectations when exact expected outputs are often unknown. (ML testing & debugging) Includes: Oracles, metamorphic and behavioral tests, readiness rubrics.; Integration testing, training diagnostics, and error analysis..",
|
|
"chapters": [
|
|
"vol1/14_ml_ops",
|
|
"vol4/15_verification"
|
|
]
|
|
},
|
|
"P4.1": {
|
|
"code": "P4.1",
|
|
"name": "Serving architectures",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "The structure through which predictions are produced and delivered, from offline batch scoring to online and serverless model servers, multi-model and multi-adapter hosting, and streaming speech services. Includes: Batch vs online prediction, inference servers, serverless endpoints, and APIs.; Multi-model and multi-adapter hosting, ranking funnels, and streaming/voice pipelines..",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol3/19_agent_economics",
|
|
"vol2/14_robust_ai",
|
|
"vol3/06_long_term_memory"
|
|
]
|
|
},
|
|
"P4.2": {
|
|
"code": "P4.2",
|
|
"name": "Compound AI & routing",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Composing several models, retrievers, and tools into one application, and routing, cascading, caching, and gating requests across models to meet quality and cost targets. (Compound AI systems & model routing) Includes: Language-model programs and their optimization; Model routing, cascades, gateways, and semantic caches.",
|
|
"chapters": [
|
|
"vol2/10_inference",
|
|
"vol3/19_agent_economics",
|
|
"vol3/06_long_term_memory"
|
|
]
|
|
},
|
|
"P4.3": {
|
|
"code": "P4.3",
|
|
"name": "Latency & goodput",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Defining and meeting per-request latency and useful-throughput targets, including tail behavior and token-level latency for generative services. (Latency, throughput & goodput) Includes: Latency budgets, tail latency, TTFT and inter-token latency.; Tail-tolerant request techniques..",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol1/12_benchmarking",
|
|
"vol2/12_ops_scale"
|
|
]
|
|
},
|
|
"P4.4": {
|
|
"code": "P4.4",
|
|
"name": "Batching & scaling",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Policies that group, admit, distribute, and provision for requests so a service meets its targets under varying load. (Batching, admission & autoscaling) Includes: Dynamic and iteration-level batching, admission, load balancing.; Autoscaling and model cold start..",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference"
|
|
]
|
|
},
|
|
"P4.5": {
|
|
"code": "P4.5",
|
|
"name": "Generative inference",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Mechanisms specific to generative inference: managing attention state, separating prompt processing from token generation, accelerating autoregressive decoding, and reducing iterative denoising cost. (LLM & generative inference) Includes: KV cache management, prefix reuse, speculative decoding.; Chunked prefill and prefill/decode disaggregation.; Denoising-step reduction and feature caching for diffusion models..",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol3/05_kv_cache",
|
|
"vol1/06_nn_architectures",
|
|
"vol1/10_model_compression",
|
|
"vol2/06_distributed_training"
|
|
]
|
|
},
|
|
"P4.6": {
|
|
"code": "P4.6",
|
|
"name": "Decoding & structure",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Controlling how tokens are selected and constraining generated output to required syntax or schemas. (Decoding, sampling & structured output) Includes: Temperature and nucleus sampling, grammar-constrained decoding, schema checks..",
|
|
"chapters": [
|
|
"vol3/02_foundation_model"
|
|
]
|
|
},
|
|
"P4.7": {
|
|
"code": "P4.7",
|
|
"name": "Test-time compute",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Spending additional inference computation on reasoning, sampling, search, and verification to raise output quality under a budget. Includes: Reasoning budgets, parallel sampling and selection, verifiers, tree search.; Trading inference compute against model size..",
|
|
"chapters": [
|
|
"vol2/10_inference",
|
|
"vol3/03_test_time_compute",
|
|
"vol3/19_agent_economics",
|
|
"vol3/13_evaluation"
|
|
]
|
|
},
|
|
"P4.8": {
|
|
"code": "P4.8",
|
|
"name": "Retrieval & RAG",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Selecting relevant records from an external corpus by lexical, dense, or hybrid search and supplying them to a model at inference. (Retrieval & vector search) Includes: Retrieval-augmented generation, ANN indexes and vector stores, reranking..",
|
|
"chapters": [
|
|
"vol2/04_data_storage",
|
|
"vol2/10_inference",
|
|
"vol3/06_long_term_memory",
|
|
"vol3/04_context_engineering"
|
|
]
|
|
},
|
|
"P4.9": {
|
|
"code": "P4.9",
|
|
"name": "On-device deployment",
|
|
"group": "P",
|
|
"areaCode": "P4",
|
|
"areaName": "Deployment & serving",
|
|
"summary": "Converting, packaging, executing, and updating models on phones, embedded devices, and microcontrollers under memory, energy, and connectivity limits. (On-device & edge deployment) Includes: Export and conversion, on-device runtimes, device/cloud split.; Over-the-air model updates across heterogeneous fleets..",
|
|
"chapters": [
|
|
"vol1/02_system_paradigms",
|
|
"vol1/10_model_compression",
|
|
"vol1/14_ml_ops",
|
|
"vol2/11_edge_intelligence",
|
|
"vol4/03_brain",
|
|
"vol2/02_compute_infrastructure"
|
|
]
|
|
},
|
|
"P5.1": {
|
|
"code": "P5.1",
|
|
"name": "Release management",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Identifying, gating, and progressively exposing model releases, with the ability to return to a known-good configuration. (Versioning, rollout & rollback) Includes: Registries, CI/CD for models, release gates.; Shadow, canary, and staged rollout; rollback..",
|
|
"chapters": [
|
|
"vol1/14_ml_ops",
|
|
"vol2/12_ops_scale",
|
|
"vol3/14_agent_operations",
|
|
"vol4/16_release",
|
|
"vol2/04_data_storage",
|
|
"vol1/03_ml_lifecycle"
|
|
]
|
|
},
|
|
"P5.2": {
|
|
"code": "P5.2",
|
|
"name": "Monitoring",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Instrumenting deployed systems and detecting, from logged signals, changes in inputs, outputs, quality, and health. (Monitoring & observability) Includes: Prediction logging, drift statistics, alerting and alert fatigue.; Metrics, logs, traces for model services and agents..",
|
|
"chapters": [
|
|
"vol1/14_ml_ops",
|
|
"vol2/07_fault_tolerance",
|
|
"vol2/12_ops_scale",
|
|
"vol3/14_agent_operations",
|
|
"vol2/14_robust_ai",
|
|
"vol4/05_data"
|
|
]
|
|
},
|
|
"P5.3": {
|
|
"code": "P5.3",
|
|
"name": "Drift & retraining",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Understanding how deployment data and outcomes change, including changes the system itself causes, and deciding when and how to update the model. (Drift, feedback loops & retraining) Includes: Covariate and concept drift, delayed labels, hidden feedback loops.; Scheduled and triggered retraining, continuous training..",
|
|
"chapters": [
|
|
"vol1/14_ml_ops",
|
|
"vol2/14_robust_ai",
|
|
"vol4/05_data",
|
|
"vol1/15_responsible_engr"
|
|
]
|
|
},
|
|
"P5.4": {
|
|
"code": "P5.4",
|
|
"name": "Incident response",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Detecting, triaging, mitigating, and learning from production failures, including attributing them to data, model, or infrastructure causes. Includes: Triage, mitigation, blameless postmortems, on-call..",
|
|
"chapters": [
|
|
"vol1/14_ml_ops",
|
|
"vol2/12_ops_scale",
|
|
"vol3/14_agent_operations"
|
|
]
|
|
},
|
|
"P5.5": {
|
|
"code": "P5.5",
|
|
"name": "Pipelines & reproducibility",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Automating the steps that produce models and controlling their environments and configuration so runs can be repeated. (ML pipelines & reproducibility) Includes: Orchestrated pipelines, environment pinning, configuration as code..",
|
|
"chapters": [
|
|
"vol1/03_ml_lifecycle",
|
|
"vol1/14_ml_ops",
|
|
"vol2/12_ops_scale",
|
|
"vol3/06_long_term_memory"
|
|
]
|
|
},
|
|
"P5.6": {
|
|
"code": "P5.6",
|
|
"name": "Platforms & teams",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Shared internal platforms, team structures, and debt management that let an organization operate many models at sustainable cost. (ML platforms, teams & technical debt) Includes: Platform-as-product, maturity levels, ownership and roles.; ML-specific technical debt and anti-patterns..",
|
|
"chapters": [
|
|
"vol1/14_ml_ops",
|
|
"vol2/12_ops_scale"
|
|
]
|
|
},
|
|
"P5.7": {
|
|
"code": "P5.7",
|
|
"name": "Retirement",
|
|
"group": "P",
|
|
"areaCode": "P5",
|
|
"areaName": "MLOps & operations",
|
|
"summary": "Ending a model's service by migrating its consumers, archiving what must be kept, and deleting data and resources that are no longer justified. (Model retirement) Includes: Deprecation with consumer migration, archival for audit, retention and deletion..",
|
|
"chapters": [
|
|
"vol2/12_ops_scale",
|
|
"vol3/06_long_term_memory"
|
|
]
|
|
},
|
|
"Q1.1": {
|
|
"code": "Q1.1",
|
|
"name": "Profiling & utilization",
|
|
"group": "Q",
|
|
"areaCode": "Q1",
|
|
"areaName": "Performance & efficiency",
|
|
"summary": "Measuring where execution time and resources go under a stated workload, and expressing how much of the hardware's capacity does useful model work. Includes: Profilers and execution-time breakdowns; Utilization metrics such as MFU and training goodput; Bottleneck attribution across compute, memory, I/O, and communication.",
|
|
"chapters": [
|
|
"vol1/08_training",
|
|
"vol1/12_benchmarking",
|
|
"vol1/13_model_serving",
|
|
"vol2/09_performance_engineering",
|
|
"vol2/10_inference",
|
|
"vol1/11_hw_acceleration"
|
|
]
|
|
},
|
|
"Q1.2": {
|
|
"code": "Q1.2",
|
|
"name": "System benchmarking",
|
|
"group": "Q",
|
|
"areaCode": "Q1",
|
|
"areaName": "Performance & efficiency",
|
|
"summary": "Measuring training and inference system performance under specified workloads, quality targets, load scenarios, and run rules so results are comparable. Includes: MLPerf-style time-to-quality and scenario-based inference metrics..",
|
|
"chapters": [
|
|
"vol1/12_benchmarking",
|
|
"vol2/09_performance_engineering",
|
|
"vol2/10_inference",
|
|
"vol2/15_sustainable_ai",
|
|
"vol1/08_training"
|
|
]
|
|
},
|
|
"Q1.3": {
|
|
"code": "Q1.3",
|
|
"name": "Model compression",
|
|
"group": "Q",
|
|
"areaCode": "Q1",
|
|
"areaName": "Performance & efficiency",
|
|
"summary": "Techniques that shrink a trained or training model's parameters, precision, or rank to reduce memory, compute, and latency at a measured cost in quality. Includes: Quantization (post-training and quantization-aware); Pruning (structured and unstructured) and sparsity exploitation; Knowledge distillation and low-rank factorization.",
|
|
"chapters": [
|
|
"vol1/10_model_compression",
|
|
"vol2/10_inference"
|
|
]
|
|
},
|
|
"Q1.4": {
|
|
"code": "Q1.4",
|
|
"name": "Hardware-aware design",
|
|
"group": "Q",
|
|
"areaCode": "Q1",
|
|
"areaName": "Performance & efficiency",
|
|
"summary": "Choosing or searching model architectures against a target device's measured cost, so that quality and efficiency are optimized together rather than compressed after the fact. (Hardware-aware model design) Includes: Neural architecture search, including hardware-aware objectives; Efficient building blocks and the accuracy-efficiency frontier; Joint model and hardware co-design.",
|
|
"chapters": [
|
|
"vol1/10_model_compression",
|
|
"vol1/11_hw_acceleration"
|
|
]
|
|
},
|
|
"Q2.1": {
|
|
"code": "Q2.1",
|
|
"name": "Failures & degradation",
|
|
"group": "Q",
|
|
"areaCode": "Q2",
|
|
"areaName": "Reliability",
|
|
"summary": "Declared failure assumptions and the service-level mechanisms (retries, idempotency, replication, circuit breaking, degradation) that keep requests correct and available when components fail. (Failure models, retries & degradation) Includes: Failure models and recovery semantics (idempotency, compensation); Retries, circuit breakers, replication, and failover; Graceful degradation of quality or features under failure.",
|
|
"chapters": [
|
|
"vol2/07_fault_tolerance",
|
|
"vol2/10_inference",
|
|
"vol3/12_failure_recovery",
|
|
"vol3/07_tool_calling",
|
|
"vol4/12_enforcement",
|
|
"vol4/01_boundary"
|
|
]
|
|
},
|
|
"Q2.2": {
|
|
"code": "Q2.2",
|
|
"name": "Checkpointing & recovery",
|
|
"group": "Q",
|
|
"areaCode": "Q2",
|
|
"areaName": "Reliability",
|
|
"summary": "Persisting recoverable state for long-running jobs and choosing checkpoint frequency and recovery strategy to minimize lost work and restart cost under a given failure rate. (Checkpointing & elastic recovery) Includes: Checkpoints and consistent cuts for distributed state; Optimal checkpoint interval (Young-Daly) and asynchronous checkpointing; Elastic recovery on a changed set of workers.",
|
|
"chapters": [
|
|
"vol2/04_data_storage",
|
|
"vol2/07_fault_tolerance",
|
|
"vol1/05_nn_computation",
|
|
"vol3/11_durable_execution",
|
|
"vol2/09_performance_engineering"
|
|
]
|
|
},
|
|
"Q2.3": {
|
|
"code": "Q2.3",
|
|
"name": "Hardware faults & SDC",
|
|
"group": "Q",
|
|
"areaCode": "Q2",
|
|
"areaName": "Reliability",
|
|
"summary": "How hardware fails in a fleet, from fail-stop crashes to gray and silent faults, how often it fails, and how such faults are detected and contained. (Hardware faults & silent data corruption) Includes: Fail-stop, gray, and silent failures and their detection; Fleet failure rates (MTBF composition) and availability; Correlated failures and failure domains.",
|
|
"chapters": [
|
|
"vol2/07_fault_tolerance",
|
|
"vol2/14_robust_ai",
|
|
"vol4/13_placement"
|
|
]
|
|
},
|
|
"Q2.4": {
|
|
"code": "Q2.4",
|
|
"name": "Fault injection & testing",
|
|
"group": "Q",
|
|
"areaCode": "Q2",
|
|
"areaName": "Reliability",
|
|
"summary": "Deliberately introducing software, infrastructure, or hardware faults under controlled scope to verify that detection, recovery, and degradation work as designed. (Fault injection & resilience testing) Includes: Chaos engineering and failure injection in production-like settings; Game days and recovery rehearsals; Hardware fault injection to estimate model error resilience.",
|
|
"chapters": [
|
|
"vol2/07_fault_tolerance",
|
|
"vol4/15_verification"
|
|
]
|
|
},
|
|
"Q3.1": {
|
|
"code": "Q3.1",
|
|
"name": "Threats & attacks",
|
|
"group": "Q",
|
|
"areaCode": "Q3",
|
|
"areaName": "Security & privacy",
|
|
"summary": "Specifying adversaries against learned systems and the attacks they mount on training data, model inputs, and instruction boundaries, with the defenses aimed at each. (Threat models & adversarial attacks) Includes: Threat models for ML assets and interfaces; Poisoning, backdoors, evasion, jailbreaks, and prompt injection; Defenses such as adversarial training.",
|
|
"chapters": [
|
|
"vol2/13_security_privacy",
|
|
"vol2/14_robust_ai",
|
|
"vol3/09_security"
|
|
]
|
|
},
|
|
"Q3.2": {
|
|
"code": "Q3.2",
|
|
"name": "Model & supply chain",
|
|
"group": "Q",
|
|
"areaCode": "Q3",
|
|
"areaName": "Security & privacy",
|
|
"summary": "Protecting models as valuable artifacts from theft and tampering, and protecting the chain of datasets, pretrained weights, libraries, and builds that produce them. (Model & supply-chain security) Includes: Model extraction and ownership evidence (watermarking); Artifact signing and provenance; Malicious weights, serialization formats, and dependencies.",
|
|
"chapters": [
|
|
"vol2/13_security_privacy",
|
|
"vol3/06_long_term_memory"
|
|
]
|
|
},
|
|
"Q3.3": {
|
|
"code": "Q3.3",
|
|
"name": "Privacy & DP",
|
|
"group": "Q",
|
|
"areaCode": "Q3",
|
|
"areaName": "Security & privacy",
|
|
"summary": "Privacy requirements for data about people, the attacks that extract it from models and outputs, and formal or architectural mechanisms that bound or remove that leakage. (Privacy leakage & differential privacy) Includes: Membership inference and model inversion; Differential privacy and DP-SGD; Machine unlearning and secure aggregation.",
|
|
"chapters": [
|
|
"vol1/15_responsible_engr",
|
|
"vol2/11_edge_intelligence",
|
|
"vol2/13_security_privacy",
|
|
"vol2/16_responsible_ai"
|
|
]
|
|
},
|
|
"Q3.4": {
|
|
"code": "Q3.4",
|
|
"name": "Access control & isolation",
|
|
"group": "Q",
|
|
"areaCode": "Q3",
|
|
"areaName": "Security & privacy",
|
|
"summary": "Constraining who may act, what code may reach, and which parties can observe data in use, through authorization, least privilege, isolation, and hardware-rooted trust. Includes: Authorization, least privilege, and sandboxed execution; Confidential computing and remote attestation; Side channels across shared accelerators and caches.",
|
|
"chapters": [
|
|
"vol2/13_security_privacy",
|
|
"vol3/08_sandboxes",
|
|
"vol3/09_security",
|
|
"vol4/13_placement"
|
|
]
|
|
},
|
|
"Q4.1": {
|
|
"code": "Q4.1",
|
|
"name": "Hazards & operating domains",
|
|
"group": "Q",
|
|
"areaCode": "Q4",
|
|
"areaName": "Safety & assurance",
|
|
"summary": "Identifying how a system can contribute to harm, including hazards from functional insufficiency rather than faults, and fixing the operating conditions within which safety claims are made. (Hazard analysis & operating domains) Includes: Hazard identification and analysis (STPA, FMEA, fault trees) applied to learned components.; Operational design domains and triggering conditions for functional insufficiency (SOTIF)..",
|
|
"chapters": [
|
|
"vol4/01_boundary",
|
|
"vol4/05_data",
|
|
"vol4/15_verification",
|
|
"vol1/14_ml_ops",
|
|
"vol2/14_robust_ai"
|
|
]
|
|
},
|
|
"Q4.2": {
|
|
"code": "Q4.2",
|
|
"name": "Runtime assurance & filters",
|
|
"group": "Q",
|
|
"areaCode": "Q4",
|
|
"areaName": "Safety & assurance",
|
|
"summary": "Mechanisms that monitor a learned component at run time and override, constrain, or replace its outputs so the system stays within a verified safe set. (Runtime assurance & safety filters) Includes: Safety filters, control barrier functions, Simplex switching, and fallback to a minimal-risk condition.; Input and output guardrails that block or rewrite unsafe generated content or actions..",
|
|
"chapters": [
|
|
"vol4/12_enforcement",
|
|
"vol4/13_placement",
|
|
"vol2/16_responsible_ai",
|
|
"vol2/14_robust_ai",
|
|
"vol4/11_planning",
|
|
"vol3/09_security",
|
|
"vol4/01_boundary"
|
|
]
|
|
},
|
|
"Q4.3": {
|
|
"code": "Q4.3",
|
|
"name": "Safety V&V",
|
|
"group": "Q",
|
|
"areaCode": "Q4",
|
|
"areaName": "Safety & assurance",
|
|
"summary": "Producing evidence that a system meets its safety requirements across its operating domain, through scenario testing, falsification, formal proof, and statistical validation of rare failures. (Safety verification & validation) Includes: Scenario coverage, simulation-based falsification, and software/hardware-in-the-loop qualification.; Formal and statistical guarantees, including verification of neural networks and rare-event estimation..",
|
|
"chapters": [
|
|
"vol4/07_evaluation",
|
|
"vol4/15_verification",
|
|
"vol4/06_training"
|
|
]
|
|
},
|
|
"Q4.4": {
|
|
"code": "Q4.4",
|
|
"name": "Safety cases & standards",
|
|
"group": "Q",
|
|
"areaCode": "Q4",
|
|
"areaName": "Safety & assurance",
|
|
"summary": "Structured arguments that a system is acceptably safe for a defined use, the standards that set what evidence is required, and the field indicators that keep the argument valid after release. (Safety cases, standards & certification) Includes: Safety-case structure and notation (claims, arguments, evidence; GSN).; Functional-safety and autonomy standards (ISO 26262, IEC 61508, ISO 21448, UL 4600) and integrity levels..",
|
|
"chapters": [
|
|
"vol4/16_release",
|
|
"vol4/18_frontier",
|
|
"vol3/14_agent_operations",
|
|
"vol2/16_responsible_ai",
|
|
"vol2/04_data_storage",
|
|
"vol2/12_ops_scale",
|
|
"vol4/05_data",
|
|
"vol4/12_enforcement"
|
|
]
|
|
},
|
|
"Q5.1": {
|
|
"code": "Q5.1",
|
|
"name": "Fairness & bias",
|
|
"group": "Q",
|
|
"areaCode": "Q5",
|
|
"areaName": "Responsible AI",
|
|
"summary": "Specifying, measuring, and mitigating how a system's outcomes and errors differ across people and groups. Includes: Group and individual fairness criteria and the metrics that measure them.; Pre-, in-, and post-processing mitigation and its accuracy cost..",
|
|
"chapters": [
|
|
"vol1/15_responsible_engr",
|
|
"vol2/16_responsible_ai",
|
|
"vol4/17_responsibility",
|
|
"vol1/12_benchmarking"
|
|
]
|
|
},
|
|
"Q5.2": {
|
|
"code": "Q5.2",
|
|
"name": "Explainability & docs",
|
|
"group": "Q",
|
|
"areaCode": "Q5",
|
|
"areaName": "Responsible AI",
|
|
"summary": "Making model behavior understandable to the people who build, deploy, and are affected by it, through explanations, interpretable models, and standardized documentation of models and data. (Explainability & documentation) Includes: Post-hoc explanations (feature attribution, counterfactuals) and inherently interpretable models.; Model cards, datasheets, and the serving cost of producing explanations..",
|
|
"chapters": [
|
|
"vol1/15_responsible_engr",
|
|
"vol2/16_responsible_ai",
|
|
"vol3/06_long_term_memory"
|
|
]
|
|
},
|
|
"Q5.3": {
|
|
"code": "Q5.3",
|
|
"name": "Oversight & contestability",
|
|
"group": "Q",
|
|
"areaCode": "Q5",
|
|
"areaName": "Responsible AI",
|
|
"summary": "Arrangements that let people supervise, intervene in, take over from, and challenge consequential system decisions, and the human-factors limits on doing so. (Human oversight & contestability) Includes: Human-in-the-loop and on-the-loop control, recourse and appeal, and control handover.; Automation bias, complacency, and the ironies of automation..",
|
|
"chapters": [
|
|
"vol2/16_responsible_ai",
|
|
"vol4/14_intervention",
|
|
"vol4/17_responsibility",
|
|
"vol2/14_robust_ai",
|
|
"vol4/01_boundary",
|
|
"vol3/09_security"
|
|
]
|
|
},
|
|
"Q5.4": {
|
|
"code": "Q5.4",
|
|
"name": "Governance & regulation",
|
|
"group": "Q",
|
|
"areaCode": "Q5",
|
|
"areaName": "Responsible AI",
|
|
"summary": "Organizational and legal mechanisms that identify, assess, and control AI risk across a system's life, including risk-management frameworks, impact assessments, audits, and regulation. (AI governance, audit & regulation) Includes: NIST AI RMF and ISO/IEC 42001 management systems; internal and external audits.; Risk-tiered regulation (EU AI Act) and frontier-model capability thresholds..",
|
|
"chapters": [
|
|
"vol1/15_responsible_engr",
|
|
"vol2/16_responsible_ai",
|
|
"vol4/16_release",
|
|
"vol4/17_responsibility"
|
|
]
|
|
},
|
|
"Q6.1": {
|
|
"code": "Q6.1",
|
|
"name": "Energy & carbon",
|
|
"group": "Q",
|
|
"areaCode": "Q6",
|
|
"areaName": "Sustainability & cost",
|
|
"summary": "Measuring the energy a workload consumes and converting it to operational emissions under declared boundaries, and scheduling work to reduce both. (Energy & carbon accounting) Includes: Energy measurement and energy efficiency (work per joule); Grid carbon intensity and operational carbon; Carbon-aware placement and scheduling.",
|
|
"chapters": [
|
|
"vol1/12_benchmarking",
|
|
"vol1/15_responsible_engr",
|
|
"vol2/15_sustainable_ai",
|
|
"vol1/02_system_paradigms",
|
|
"vol4/02_body"
|
|
]
|
|
},
|
|
"Q6.2": {
|
|
"code": "Q6.2",
|
|
"name": "Embodied & lifecycle",
|
|
"group": "Q",
|
|
"areaCode": "Q6",
|
|
"areaName": "Sustainability & cost",
|
|
"summary": "Environmental impacts beyond operating energy, from manufacturing emissions and water to hardware lifetime and disposal, assessed over the full hardware lifecycle. (Embodied carbon & lifecycle impact) Includes: Embodied carbon and lifecycle assessment; Water consumption of training and serving; Hardware lifetime, refresh, and e-waste.",
|
|
"chapters": [
|
|
"vol2/15_sustainable_ai",
|
|
"vol1/15_responsible_engr"
|
|
]
|
|
},
|
|
"Q6.3": {
|
|
"code": "Q6.3",
|
|
"name": "Cost & unit economics",
|
|
"group": "Q",
|
|
"areaCode": "Q6",
|
|
"areaName": "Sustainability & cost",
|
|
"summary": "Monetary accounting for ML systems, from total cost of ownership to cost per token or accepted outcome, and the architectural choices that lower cost at fixed quality. (Cost, TCO & unit economics) Includes: Total cost of ownership and cost attribution (FinOps); Cost per token and cost per accepted outcome; Model routing and cascades chosen for cost.",
|
|
"chapters": [
|
|
"vol1/13_model_serving",
|
|
"vol1/15_responsible_engr",
|
|
"vol2/08_fleet_orchestration",
|
|
"vol2/09_performance_engineering",
|
|
"vol2/12_ops_scale",
|
|
"vol3/19_agent_economics",
|
|
"vol3/05_kv_cache"
|
|
]
|
|
},
|
|
"T1.1": {
|
|
"code": "T1.1",
|
|
"name": "Context & agent memory",
|
|
"group": "T",
|
|
"areaCode": "T1",
|
|
"areaName": "Agentic systems",
|
|
"summary": "Deciding what information enters each model call over a long task, and maintaining persistent memory the agent can write, consolidate, and retrieve across calls and sessions. (Context engineering & agent memory) Includes: Working-set selection, ordering for prefix reuse, compaction, and long-context degradation.; Long-term memory write paths, consolidation, expiry, and invalidation..",
|
|
"chapters": [
|
|
"vol3/04_context_engineering",
|
|
"vol3/06_long_term_memory",
|
|
"vol1/13_model_serving",
|
|
"vol2/10_inference",
|
|
"vol3/05_kv_cache",
|
|
"vol3/09_security"
|
|
]
|
|
},
|
|
"T1.2": {
|
|
"code": "T1.2",
|
|
"name": "Tool use & actions",
|
|
"group": "T",
|
|
"areaCode": "T1",
|
|
"areaName": "Agentic systems",
|
|
"summary": "The interfaces through which a model acts on the world, including structured tool calls, tool discovery protocols, computer use, and the effects and authority those actions carry. (Tool use & action interfaces) Includes: Tool schemas, preconditions, function calling, and MCP-style discovery.; Computer-use agents and the handling of irreversible effects..",
|
|
"chapters": [
|
|
"vol3/02_foundation_model",
|
|
"vol3/07_tool_calling",
|
|
"vol3/08_sandboxes",
|
|
"vol4/13_placement",
|
|
"vol3/09_security"
|
|
]
|
|
},
|
|
"T1.3": {
|
|
"code": "T1.3",
|
|
"name": "Agent harness & loop",
|
|
"group": "T",
|
|
"areaCode": "T1",
|
|
"areaName": "Agentic systems",
|
|
"summary": "The runtime that owns an agent's loop of model calls and tool actions, enforcing budgets, stopping rules, and approval gates and recording the trajectory. (Agent harness & control loop) Includes: Reason-act loops, workflows versus model-directed loops, and plan revision.; Turn, token, time, and spend budgets; stopping rules; human approval gates..",
|
|
"chapters": [
|
|
"vol3/01_introduction",
|
|
"vol3/10_agent_harness",
|
|
"vol4/17_responsibility",
|
|
"vol3/19_agent_economics",
|
|
"vol3/13_evaluation",
|
|
"vol3/09_security",
|
|
"vol3/12_failure_recovery"
|
|
]
|
|
},
|
|
"T1.4": {
|
|
"code": "T1.4",
|
|
"name": "Durable execution",
|
|
"group": "T",
|
|
"areaCode": "T1",
|
|
"areaName": "Agentic systems",
|
|
"summary": "Persisting an agent's progress so long-running trajectories survive crashes, pauses, and redeployments without repeating completed effects, and can be replayed for debugging. (Durable execution & agent state) Includes: Event-sourced trajectory logs, resume protocols, and record-and-replay.; Pausing for human or external events and resuming safely..",
|
|
"chapters": [
|
|
"vol3/11_durable_execution",
|
|
"vol3/12_failure_recovery",
|
|
"vol2/04_data_storage",
|
|
"vol3/07_tool_calling",
|
|
"vol2/07_fault_tolerance"
|
|
]
|
|
},
|
|
"T1.5": {
|
|
"code": "T1.5",
|
|
"name": "Multi-agent orchestration",
|
|
"group": "T",
|
|
"areaCode": "T1",
|
|
"areaName": "Agentic systems",
|
|
"summary": "Decomposing a task across several model-driven agents and coordinating their delegation, shared state, and results, including when coordination costs outweigh the benefit. Includes: Orchestrator-worker and peer topologies, typed task envelopes, and delegation.; Coordination failures and correlated errors among agents sharing a base model..",
|
|
"chapters": [
|
|
"vol3/18_multi_agent",
|
|
"vol3/13_evaluation",
|
|
"vol2/07_fault_tolerance"
|
|
]
|
|
},
|
|
"T2.1": {
|
|
"code": "T2.1",
|
|
"name": "Sensing & perception",
|
|
"group": "T",
|
|
"areaCode": "T2",
|
|
"areaName": "Embodied & physical",
|
|
"summary": "Turning time-stamped measurements from physical sensors into task-relevant estimates of the scene, including calibration, synchronization, and fusion across modalities. (Sensing, perception & fusion) Includes: Camera, LiDAR, radar, event, and tactile sensing; calibration chains.; Detection, segmentation, learned depth, and multi-sensor fusion..",
|
|
"chapters": [
|
|
"vol4/08_perception",
|
|
"vol1/14_ml_ops",
|
|
"vol2/14_robust_ai",
|
|
"vol4/05_data"
|
|
]
|
|
},
|
|
"T2.2": {
|
|
"code": "T2.2",
|
|
"name": "State estimation & mapping",
|
|
"group": "T",
|
|
"areaCode": "T2",
|
|
"areaName": "Embodied & physical",
|
|
"summary": "Inferring a robot's and its environment's latent state and uncertainty over time, through recursive filtering, localization and mapping, and learned predictive world models. (State estimation, localization & mapping) Includes: Kalman and particle filters, tracking and data association, SLAM, metric and semantic maps.; Learned world models used for prediction and planning..",
|
|
"chapters": [
|
|
"vol4/09_memory",
|
|
"vol4/18_frontier"
|
|
]
|
|
},
|
|
"T2.3": {
|
|
"code": "T2.3",
|
|
"name": "Motion planning & control",
|
|
"group": "T",
|
|
"areaCode": "T2",
|
|
"areaName": "Embodied & physical",
|
|
"summary": "Generating feasible, collision-free motions and closing feedback loops that drive a physical system along them under dynamics, contact, and actuation limits. Includes: Sampling- and optimization-based motion planning, task and motion planning.; Feedback control, MPC, and contact-rich manipulation..",
|
|
"chapters": [
|
|
"vol4/02_body",
|
|
"vol4/10_intent",
|
|
"vol4/11_planning",
|
|
"vol4/04_nervous",
|
|
"vol4/12_enforcement",
|
|
"vol4/01_boundary"
|
|
]
|
|
},
|
|
"T2.4": {
|
|
"code": "T2.4",
|
|
"name": "Real-time & criticality",
|
|
"group": "T",
|
|
"areaCode": "T2",
|
|
"areaName": "Embodied & physical",
|
|
"summary": "Guaranteeing that sensing, inference, and control complete within deadlines on shared embedded hardware, and isolating safety-critical functions from less trusted ones. (Real-time & mixed-criticality computing) Includes: Multi-rate scheduling, response-time analysis, WCET, and real-time buses.; Mixed-criticality partitioning and hypervisors on shared silicon..",
|
|
"chapters": [
|
|
"vol4/04_nervous",
|
|
"vol4/13_placement",
|
|
"vol3/08_sandboxes",
|
|
"vol4/08_perception"
|
|
]
|
|
},
|
|
"T2.5": {
|
|
"code": "T2.5",
|
|
"name": "Robot learning & sim-to-real",
|
|
"group": "T",
|
|
"areaCode": "T2",
|
|
"areaName": "Embodied & physical",
|
|
"summary": "Learning robot policies from teleoperated demonstrations, simulation, and pretrained vision-language backbones, and transferring them to physical hardware. Includes: Vision-language-action models, action chunking, and cross-embodiment pretraining.; Teleoperation data collection, domain randomization, and sim-to-real transfer..",
|
|
"chapters": [
|
|
"vol4/03_brain",
|
|
"vol4/05_data",
|
|
"vol4/06_training",
|
|
"vol1/14_ml_ops",
|
|
"vol2/14_robust_ai"
|
|
]
|
|
}
|
|
});
|
|
|
|
export const CHAPTER_TAXONOMY_INDEX = attachLookupHelpers({
|
|
"vol1/01_introduction": {
|
|
"slug": "vol1/01_introduction",
|
|
"volume": "vol1",
|
|
"title": "Introduction",
|
|
"primaryGroup": "F",
|
|
"groups": [
|
|
"F"
|
|
],
|
|
"subjects": [
|
|
"F1.1",
|
|
"F3.1"
|
|
]
|
|
},
|
|
"vol1/02_system_paradigms": {
|
|
"slug": "vol1/02_system_paradigms",
|
|
"volume": "vol1",
|
|
"title": "System Paradigms",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F1.2",
|
|
"S1.5",
|
|
"P4.9",
|
|
"F3.1",
|
|
"S1.2",
|
|
"P2.3",
|
|
"Q6.1"
|
|
]
|
|
},
|
|
"vol1/03_ml_lifecycle": {
|
|
"slug": "vol1/03_ml_lifecycle",
|
|
"volume": "vol1",
|
|
"title": "ML Lifecycle",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"F1.1",
|
|
"F1.3",
|
|
"P3.1",
|
|
"P5.5",
|
|
"P1.3",
|
|
"P1.5",
|
|
"P5.1"
|
|
]
|
|
},
|
|
"vol1/04_data_engineering": {
|
|
"slug": "vol1/04_data_engineering",
|
|
"volume": "vol1",
|
|
"title": "Data Engineering",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"S",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"S4.3",
|
|
"P1.1",
|
|
"P1.3",
|
|
"P1.4",
|
|
"P1.5",
|
|
"P1.6",
|
|
"S2.6",
|
|
"P3.1"
|
|
]
|
|
},
|
|
"vol1/05_nn_computation": {
|
|
"slug": "vol1/05_nn_computation",
|
|
"volume": "vol1",
|
|
"title": "Neural Computation",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F1.2",
|
|
"F2.1",
|
|
"P2.1",
|
|
"P2.2",
|
|
"S1.1",
|
|
"S2.2",
|
|
"S3.1",
|
|
"P2.3",
|
|
"Q2.2"
|
|
]
|
|
},
|
|
"vol1/06_nn_architectures": {
|
|
"slug": "vol1/06_nn_architectures",
|
|
"volume": "vol1",
|
|
"title": "Network Architectures",
|
|
"primaryGroup": "F",
|
|
"groups": [
|
|
"F",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"F1.2",
|
|
"F2.1",
|
|
"F2.2",
|
|
"F2.3",
|
|
"F2.4",
|
|
"F3.1",
|
|
"P4.5"
|
|
]
|
|
},
|
|
"vol1/07_frameworks": {
|
|
"slug": "vol1/07_frameworks",
|
|
"volume": "vol1",
|
|
"title": "ML Frameworks",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S"
|
|
],
|
|
"subjects": [
|
|
"F2.5",
|
|
"S2.1",
|
|
"S2.4",
|
|
"S2.5",
|
|
"S2.2"
|
|
]
|
|
},
|
|
"vol1/08_training": {
|
|
"slug": "vol1/08_training",
|
|
"volume": "vol1",
|
|
"title": "Model Training",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F3.2",
|
|
"F3.4",
|
|
"S1.3",
|
|
"S2.2",
|
|
"S2.6",
|
|
"S2.7",
|
|
"S3.1",
|
|
"P2.2",
|
|
"P2.3",
|
|
"P2.4",
|
|
"P2.5",
|
|
"Q1.1",
|
|
"Q1.2"
|
|
]
|
|
},
|
|
"vol1/09_data_selection": {
|
|
"slug": "vol1/09_data_selection",
|
|
"volume": "vol1",
|
|
"title": "Data Selection",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"F3.4",
|
|
"P1.2",
|
|
"P2.1"
|
|
]
|
|
},
|
|
"vol1/10_model_compression": {
|
|
"slug": "vol1/10_model_compression",
|
|
"volume": "vol1",
|
|
"title": "Model Compression",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F2.3",
|
|
"S2.7",
|
|
"P4.9",
|
|
"Q1.3",
|
|
"Q1.4",
|
|
"S1.5",
|
|
"P2.5",
|
|
"P4.5"
|
|
]
|
|
},
|
|
"vol1/11_hw_acceleration": {
|
|
"slug": "vol1/11_hw_acceleration",
|
|
"volume": "vol1",
|
|
"title": "Hardware Acceleration",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F3.1",
|
|
"S1.1",
|
|
"S1.2",
|
|
"S1.4",
|
|
"S1.5",
|
|
"S2.2",
|
|
"S2.3",
|
|
"S2.4",
|
|
"F3.2",
|
|
"Q1.1",
|
|
"Q1.4"
|
|
]
|
|
},
|
|
"vol1/12_benchmarking": {
|
|
"slug": "vol1/12_benchmarking",
|
|
"volume": "vol1",
|
|
"title": "Benchmarking",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F1.1",
|
|
"P3.2",
|
|
"Q1.1",
|
|
"Q1.2",
|
|
"Q6.1",
|
|
"P3.3",
|
|
"P4.3",
|
|
"Q5.1"
|
|
]
|
|
},
|
|
"vol1/13_model_serving": {
|
|
"slug": "vol1/13_model_serving",
|
|
"volume": "vol1",
|
|
"title": "Model Serving",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"F1.1",
|
|
"F2.5",
|
|
"F3.3",
|
|
"S2.5",
|
|
"S4.5",
|
|
"P4.1",
|
|
"P4.3",
|
|
"P4.4",
|
|
"P4.5",
|
|
"Q1.1",
|
|
"Q6.3",
|
|
"F1.2",
|
|
"F2.2",
|
|
"F2.4",
|
|
"S4.4",
|
|
"T1.1"
|
|
]
|
|
},
|
|
"vol1/14_ml_ops": {
|
|
"slug": "vol1/14_ml_ops",
|
|
"volume": "vol1",
|
|
"title": "ML Operations",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"F2.5",
|
|
"P2.4",
|
|
"P3.5",
|
|
"P4.9",
|
|
"P5.1",
|
|
"P5.2",
|
|
"P5.3",
|
|
"P5.4",
|
|
"P5.5",
|
|
"P5.6",
|
|
"P2.7",
|
|
"P3.3",
|
|
"Q4.1",
|
|
"T2.1",
|
|
"T2.5"
|
|
]
|
|
},
|
|
"vol1/15_responsible_engr": {
|
|
"slug": "vol1/15_responsible_engr",
|
|
"volume": "vol1",
|
|
"title": "Responsible Engineering",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"F1.3",
|
|
"S3.5",
|
|
"Q3.3",
|
|
"Q5.1",
|
|
"Q5.2",
|
|
"Q5.4",
|
|
"Q6.1",
|
|
"Q6.3",
|
|
"S4.1",
|
|
"P3.3",
|
|
"P5.3",
|
|
"Q6.2"
|
|
]
|
|
},
|
|
"vol2/01_introduction": {
|
|
"slug": "vol2/01_introduction",
|
|
"volume": "vol2",
|
|
"title": "Introduction",
|
|
"primaryGroup": "F",
|
|
"groups": [
|
|
"F",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"F1.2",
|
|
"F3.2",
|
|
"F3.4",
|
|
"P1.2",
|
|
"P2.4"
|
|
]
|
|
},
|
|
"vol2/02_compute_infrastructure": {
|
|
"slug": "vol2/02_compute_infrastructure",
|
|
"volume": "vol2",
|
|
"title": "Compute Infrastructure",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"S",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"S1.1",
|
|
"S1.2",
|
|
"S1.3",
|
|
"S1.4",
|
|
"S4.1",
|
|
"P2.3",
|
|
"P4.9"
|
|
]
|
|
},
|
|
"vol2/03_network_fabrics": {
|
|
"slug": "vol2/03_network_fabrics",
|
|
"volume": "vol2",
|
|
"title": "Network Fabrics",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"S"
|
|
],
|
|
"subjects": [
|
|
"S1.3",
|
|
"S4.2",
|
|
"S3.3"
|
|
]
|
|
},
|
|
"vol2/04_data_storage": {
|
|
"slug": "vol2/04_data_storage",
|
|
"volume": "vol2",
|
|
"title": "Data Storage",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"S2.6",
|
|
"S4.3",
|
|
"P1.4",
|
|
"P4.8",
|
|
"Q2.2",
|
|
"F2.5",
|
|
"S3.4",
|
|
"S4.4",
|
|
"P1.5",
|
|
"P2.3",
|
|
"P5.1",
|
|
"Q4.4",
|
|
"T1.4"
|
|
]
|
|
},
|
|
"vol2/05_collective_communication": {
|
|
"slug": "vol2/05_collective_communication",
|
|
"volume": "vol2",
|
|
"title": "Collective Communication",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S"
|
|
],
|
|
"subjects": [
|
|
"S3.3",
|
|
"F3.2",
|
|
"S1.3",
|
|
"S2.5",
|
|
"S3.1",
|
|
"S4.2"
|
|
]
|
|
},
|
|
"vol2/06_distributed_training": {
|
|
"slug": "vol2/06_distributed_training",
|
|
"volume": "vol2",
|
|
"title": "Distributed Training",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"F2.3",
|
|
"F3.2",
|
|
"S2.7",
|
|
"S3.1",
|
|
"S3.2",
|
|
"S3.3",
|
|
"S3.4",
|
|
"P2.2",
|
|
"P2.3",
|
|
"P2.6",
|
|
"P4.5"
|
|
]
|
|
},
|
|
"vol2/07_fault_tolerance": {
|
|
"slug": "vol2/07_fault_tolerance",
|
|
"volume": "vol2",
|
|
"title": "Fault Tolerance",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"S",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"S3.4",
|
|
"P5.2",
|
|
"Q2.1",
|
|
"Q2.2",
|
|
"Q2.3",
|
|
"Q2.4",
|
|
"T1.4",
|
|
"T1.5"
|
|
]
|
|
},
|
|
"vol2/08_fleet_orchestration": {
|
|
"slug": "vol2/08_fleet_orchestration",
|
|
"volume": "vol2",
|
|
"title": "Fleet Orchestration",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"S",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"S4.4",
|
|
"S4.5",
|
|
"Q6.3"
|
|
]
|
|
},
|
|
"vol2/09_performance_engineering": {
|
|
"slug": "vol2/09_performance_engineering",
|
|
"volume": "vol2",
|
|
"title": "Fleet Efficiency",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"S",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"S2.3",
|
|
"S2.4",
|
|
"S2.5",
|
|
"Q1.1",
|
|
"Q1.2",
|
|
"Q6.3",
|
|
"Q2.2"
|
|
]
|
|
},
|
|
"vol2/10_inference": {
|
|
"slug": "vol2/10_inference",
|
|
"volume": "vol2",
|
|
"title": "Serving at Scale",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"F3.3",
|
|
"S3.1",
|
|
"S3.2",
|
|
"S4.5",
|
|
"P4.1",
|
|
"P4.2",
|
|
"P4.3",
|
|
"P4.4",
|
|
"P4.5",
|
|
"P4.7",
|
|
"P4.8",
|
|
"Q1.3",
|
|
"Q2.1",
|
|
"F1.1",
|
|
"F2.2",
|
|
"F2.4",
|
|
"Q1.1",
|
|
"Q1.2",
|
|
"T1.1"
|
|
]
|
|
},
|
|
"vol2/11_edge_intelligence": {
|
|
"slug": "vol2/11_edge_intelligence",
|
|
"volume": "vol2",
|
|
"title": "Edge Intelligence",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"S1.5",
|
|
"S3.5",
|
|
"P2.5",
|
|
"P2.7",
|
|
"P4.9",
|
|
"Q3.3",
|
|
"P2.1"
|
|
]
|
|
},
|
|
"vol2/12_ops_scale": {
|
|
"slug": "vol2/12_ops_scale",
|
|
"volume": "vol2",
|
|
"title": "ML Operations at Scale",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"S4.5",
|
|
"P1.3",
|
|
"P1.4",
|
|
"P1.5",
|
|
"P1.6",
|
|
"P3.1",
|
|
"P5.1",
|
|
"P5.2",
|
|
"P5.4",
|
|
"P5.5",
|
|
"P5.6",
|
|
"P5.7",
|
|
"Q6.3",
|
|
"F1.1",
|
|
"P4.3",
|
|
"Q4.4"
|
|
]
|
|
},
|
|
"vol2/13_security_privacy": {
|
|
"slug": "vol2/13_security_privacy",
|
|
"volume": "vol2",
|
|
"title": "Security & Privacy",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"S",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"Q3.1",
|
|
"Q3.2",
|
|
"Q3.3",
|
|
"Q3.4",
|
|
"S3.5"
|
|
]
|
|
},
|
|
"vol2/14_robust_ai": {
|
|
"slug": "vol2/14_robust_ai",
|
|
"volume": "vol2",
|
|
"title": "Robust AI",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P3.2",
|
|
"P3.3",
|
|
"P5.3",
|
|
"Q2.3",
|
|
"Q3.1",
|
|
"F1.3",
|
|
"P2.7",
|
|
"P4.1",
|
|
"P5.2",
|
|
"Q4.1",
|
|
"Q4.2",
|
|
"Q5.3",
|
|
"T2.1",
|
|
"T2.5"
|
|
]
|
|
},
|
|
"vol2/15_sustainable_ai": {
|
|
"slug": "vol2/15_sustainable_ai",
|
|
"volume": "vol2",
|
|
"title": "Sustainable AI",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"S",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"S4.1",
|
|
"Q6.1",
|
|
"Q6.2",
|
|
"S1.4",
|
|
"Q1.2"
|
|
]
|
|
},
|
|
"vol2/16_responsible_ai": {
|
|
"slug": "vol2/16_responsible_ai",
|
|
"volume": "vol2",
|
|
"title": "Responsible AI",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"P3.3",
|
|
"Q3.3",
|
|
"Q4.2",
|
|
"Q4.4",
|
|
"Q5.1",
|
|
"Q5.2",
|
|
"Q5.3",
|
|
"Q5.4"
|
|
]
|
|
},
|
|
"vol3/01_introduction": {
|
|
"slug": "vol3/01_introduction",
|
|
"volume": "vol3",
|
|
"title": "Introduction",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T1.3"
|
|
]
|
|
},
|
|
"vol3/02_foundation_model": {
|
|
"slug": "vol3/02_foundation_model",
|
|
"volume": "vol3",
|
|
"title": "The Foundation Model",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"F2.2",
|
|
"P4.6",
|
|
"T1.2"
|
|
]
|
|
},
|
|
"vol3/03_test_time_compute": {
|
|
"slug": "vol3/03_test_time_compute",
|
|
"volume": "vol3",
|
|
"title": "Test-Time Compute",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"P3.4",
|
|
"P4.7",
|
|
"P2.6"
|
|
]
|
|
},
|
|
"vol3/04_context_engineering": {
|
|
"slug": "vol3/04_context_engineering",
|
|
"volume": "vol3",
|
|
"title": "Context Engineering",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T1.1",
|
|
"F2.4",
|
|
"P4.8"
|
|
]
|
|
},
|
|
"vol3/05_kv_cache": {
|
|
"slug": "vol3/05_kv_cache",
|
|
"volume": "vol3",
|
|
"title": "KV Cache Management",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"F3.3",
|
|
"P4.5",
|
|
"F2.2",
|
|
"F2.4",
|
|
"Q6.3",
|
|
"T1.1"
|
|
]
|
|
},
|
|
"vol3/06_long_term_memory": {
|
|
"slug": "vol3/06_long_term_memory",
|
|
"volume": "vol3",
|
|
"title": "Long-Term Memory",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P4.8",
|
|
"T1.1",
|
|
"P1.5",
|
|
"P4.1",
|
|
"P4.2",
|
|
"P5.5",
|
|
"P5.7",
|
|
"Q3.2",
|
|
"Q5.2"
|
|
]
|
|
},
|
|
"vol3/07_tool_calling": {
|
|
"slug": "vol3/07_tool_calling",
|
|
"volume": "vol3",
|
|
"title": "Tool Calling",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"S",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T1.2",
|
|
"S3.4",
|
|
"Q2.1",
|
|
"T1.4"
|
|
]
|
|
},
|
|
"vol3/08_sandboxes": {
|
|
"slug": "vol3/08_sandboxes",
|
|
"volume": "vol3",
|
|
"title": "Agent Sandboxes",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"S",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q3.4",
|
|
"S4.5",
|
|
"T1.2",
|
|
"T2.4"
|
|
]
|
|
},
|
|
"vol3/09_security": {
|
|
"slug": "vol3/09_security",
|
|
"volume": "vol3",
|
|
"title": "Agent Security",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q3.1",
|
|
"Q3.4",
|
|
"Q4.2",
|
|
"Q5.3",
|
|
"T1.1",
|
|
"T1.2",
|
|
"T1.3"
|
|
]
|
|
},
|
|
"vol3/10_agent_harness": {
|
|
"slug": "vol3/10_agent_harness",
|
|
"volume": "vol3",
|
|
"title": "The Agent Harness",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T1.3"
|
|
]
|
|
},
|
|
"vol3/11_durable_execution": {
|
|
"slug": "vol3/11_durable_execution",
|
|
"volume": "vol3",
|
|
"title": "Durable Execution",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"S3.4",
|
|
"T1.4",
|
|
"F2.5",
|
|
"S4.3",
|
|
"S4.4",
|
|
"P2.3",
|
|
"Q2.2"
|
|
]
|
|
},
|
|
"vol3/12_failure_recovery": {
|
|
"slug": "vol3/12_failure_recovery",
|
|
"volume": "vol3",
|
|
"title": "Failure Recovery",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q2.1",
|
|
"T1.4",
|
|
"T1.3"
|
|
]
|
|
},
|
|
"vol3/13_evaluation": {
|
|
"slug": "vol3/13_evaluation",
|
|
"volume": "vol3",
|
|
"title": "Agent Evaluation",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P3.2",
|
|
"P3.4",
|
|
"P4.7",
|
|
"T1.3",
|
|
"T1.5"
|
|
]
|
|
},
|
|
"vol3/14_agent_operations": {
|
|
"slug": "vol3/14_agent_operations",
|
|
"volume": "vol3",
|
|
"title": "Operating Agents",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"P5.1",
|
|
"P5.2",
|
|
"P5.4",
|
|
"Q4.4"
|
|
]
|
|
},
|
|
"vol3/15_trajectory_curation": {
|
|
"slug": "vol3/15_trajectory_curation",
|
|
"volume": "vol3",
|
|
"title": "Trajectory Curation",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"P1.1",
|
|
"P1.2",
|
|
"P1.6",
|
|
"P3.1"
|
|
]
|
|
},
|
|
"vol3/16_fine_tuning": {
|
|
"slug": "vol3/16_fine_tuning",
|
|
"volume": "vol3",
|
|
"title": "Trajectory Fine-Tuning",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"P2.5",
|
|
"P2.6",
|
|
"P2.1",
|
|
"P3.1"
|
|
]
|
|
},
|
|
"vol3/17_rlvr": {
|
|
"slug": "vol3/17_rlvr",
|
|
"volume": "vol3",
|
|
"title": "Reinforcement Learning from Verifiable Rewards",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"F",
|
|
"P"
|
|
],
|
|
"subjects": [
|
|
"F1.3",
|
|
"P2.6"
|
|
]
|
|
},
|
|
"vol3/18_multi_agent": {
|
|
"slug": "vol3/18_multi_agent",
|
|
"volume": "vol3",
|
|
"title": "Multi-Agent Coordination",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"S",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"S3.4",
|
|
"T1.5"
|
|
]
|
|
},
|
|
"vol3/19_agent_economics": {
|
|
"slug": "vol3/19_agent_economics",
|
|
"volume": "vol3",
|
|
"title": "Agent Economics",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"S",
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P4.1",
|
|
"P4.2",
|
|
"Q6.3",
|
|
"S4.5",
|
|
"P4.7",
|
|
"T1.3"
|
|
]
|
|
},
|
|
"vol4/01_boundary": {
|
|
"slug": "vol4/01_boundary",
|
|
"volume": "vol4",
|
|
"title": "The Causal Boundary",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q4.1",
|
|
"Q2.1",
|
|
"Q4.2",
|
|
"Q5.3",
|
|
"T2.3"
|
|
]
|
|
},
|
|
"vol4/02_body": {
|
|
"slug": "vol4/02_body",
|
|
"volume": "vol4",
|
|
"title": "The Physical Body",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"S",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"S1.4",
|
|
"T2.3",
|
|
"S4.1",
|
|
"Q6.1"
|
|
]
|
|
},
|
|
"vol4/03_brain": {
|
|
"slug": "vol4/03_brain",
|
|
"volume": "vol4",
|
|
"title": "The Cognitive Brain",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"F",
|
|
"S",
|
|
"P",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P4.9",
|
|
"T2.5",
|
|
"F3.1",
|
|
"S1.2"
|
|
]
|
|
},
|
|
"vol4/04_nervous": {
|
|
"slug": "vol4/04_nervous",
|
|
"volume": "vol4",
|
|
"title": "The Nervous System",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"S",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T2.4",
|
|
"S3.4",
|
|
"T2.3"
|
|
]
|
|
},
|
|
"vol4/05_data": {
|
|
"slug": "vol4/05_data",
|
|
"volume": "vol4",
|
|
"title": "Physical Data",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P1.1",
|
|
"P1.2",
|
|
"P1.5",
|
|
"P5.3",
|
|
"Q4.1",
|
|
"T2.5",
|
|
"P2.1",
|
|
"P2.7",
|
|
"P3.3",
|
|
"P5.2",
|
|
"Q4.4",
|
|
"T2.1"
|
|
]
|
|
},
|
|
"vol4/06_training": {
|
|
"slug": "vol4/06_training",
|
|
"volume": "vol4",
|
|
"title": "Policy Training",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"P",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P2.1",
|
|
"T2.5",
|
|
"Q4.3"
|
|
]
|
|
},
|
|
"vol4/07_evaluation": {
|
|
"slug": "vol4/07_evaluation",
|
|
"volume": "vol4",
|
|
"title": "Closed-Loop Evaluation",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"P3.1",
|
|
"P3.2",
|
|
"Q4.3"
|
|
]
|
|
},
|
|
"vol4/08_perception": {
|
|
"slug": "vol4/08_perception",
|
|
"volume": "vol4",
|
|
"title": "Sensor Perception",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T2.1",
|
|
"T2.4"
|
|
]
|
|
},
|
|
"vol4/09_memory": {
|
|
"slug": "vol4/09_memory",
|
|
"volume": "vol4",
|
|
"title": "Spatial Memory",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T2.2"
|
|
]
|
|
},
|
|
"vol4/10_intent": {
|
|
"slug": "vol4/10_intent",
|
|
"volume": "vol4",
|
|
"title": "Grounded Intent",
|
|
"primaryGroup": "P",
|
|
"groups": [
|
|
"P",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"P3.2",
|
|
"T2.3"
|
|
]
|
|
},
|
|
"vol4/11_planning": {
|
|
"slug": "vol4/11_planning",
|
|
"volume": "vol4",
|
|
"title": "Trajectory Planning",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"T2.3",
|
|
"Q4.2"
|
|
]
|
|
},
|
|
"vol4/12_enforcement": {
|
|
"slug": "vol4/12_enforcement",
|
|
"volume": "vol4",
|
|
"title": "Safety Enforcement",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q4.2",
|
|
"Q2.1",
|
|
"Q4.4",
|
|
"T2.3"
|
|
]
|
|
},
|
|
"vol4/13_placement": {
|
|
"slug": "vol4/13_placement",
|
|
"volume": "vol4",
|
|
"title": "Workload Placement",
|
|
"primaryGroup": "S",
|
|
"groups": [
|
|
"S",
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"S1.4",
|
|
"S1.5",
|
|
"Q4.2",
|
|
"T2.4",
|
|
"S4.1",
|
|
"S4.5",
|
|
"Q2.3",
|
|
"Q3.4",
|
|
"T1.2"
|
|
]
|
|
},
|
|
"vol4/14_intervention": {
|
|
"slug": "vol4/14_intervention",
|
|
"volume": "vol4",
|
|
"title": "Supervisory Intervention",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"Q5.3"
|
|
]
|
|
},
|
|
"vol4/15_verification": {
|
|
"slug": "vol4/15_verification",
|
|
"volume": "vol4",
|
|
"title": "Adversarial Verification",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"P3.3",
|
|
"P3.5",
|
|
"Q2.4",
|
|
"Q4.1",
|
|
"Q4.3"
|
|
]
|
|
},
|
|
"vol4/16_release": {
|
|
"slug": "vol4/16_release",
|
|
"volume": "vol4",
|
|
"title": "Deployment Release",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"F",
|
|
"P",
|
|
"Q"
|
|
],
|
|
"subjects": [
|
|
"P5.1",
|
|
"Q4.4",
|
|
"Q5.4",
|
|
"F2.5",
|
|
"P1.5"
|
|
]
|
|
},
|
|
"vol4/17_responsibility": {
|
|
"slug": "vol4/17_responsibility",
|
|
"volume": "vol4",
|
|
"title": "Responsible Physical AI",
|
|
"primaryGroup": "Q",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q5.1",
|
|
"Q5.3",
|
|
"Q5.4",
|
|
"T1.3"
|
|
]
|
|
},
|
|
"vol4/18_frontier": {
|
|
"slug": "vol4/18_frontier",
|
|
"volume": "vol4",
|
|
"title": "Engineering Physical AI",
|
|
"primaryGroup": "T",
|
|
"groups": [
|
|
"Q",
|
|
"T"
|
|
],
|
|
"subjects": [
|
|
"Q4.4",
|
|
"T2.2"
|
|
]
|
|
}
|
|
});
|
|
|
|
/**
|
|
* Resolves any chapter URL, pathname, or slug (including `/vol1/01_introduction/`,
|
|
* `/vol1/chapters/01_introduction/01_introduction.html`, or `vol1/01_introduction`)
|
|
* to its canonical chapter taxonomy entry with subject details.
|
|
*/
|
|
export function resolveChapterTaxonomy(urlOrPath = '') {
|
|
if (typeof urlOrPath !== 'string' || !urlOrPath.trim()) return null;
|
|
const raw = urlOrPath.trim();
|
|
|
|
// Direct slug lookup first
|
|
if (CHAPTER_TAXONOMY_INDEX[raw]) {
|
|
return enrichChapterEntry(CHAPTER_TAXONOMY_INDEX[raw]);
|
|
}
|
|
|
|
let pathname = raw;
|
|
try {
|
|
if (/^https?:\/\//i.test(raw)) {
|
|
pathname = new URL(raw).pathname;
|
|
}
|
|
} catch {
|
|
pathname = raw;
|
|
}
|
|
|
|
const normalized = pathname.replace(/\\/g, '/');
|
|
|
|
// Match /volN/chapters/NN_slug or /volN/NN_slug
|
|
const match = normalized.match(/(?:^|\b|\/)(vol[1-4])\/(?:chapters\/)?([0-9]{2}_[a-z0-9_]+)(?:\/|\.html|$|#|\?)/i);
|
|
if (match) {
|
|
const slug = `${match[1].toLowerCase()}/${match[2].toLowerCase()}`;
|
|
if (CHAPTER_TAXONOMY_INDEX[slug]) {
|
|
return enrichChapterEntry(CHAPTER_TAXONOMY_INDEX[slug]);
|
|
}
|
|
}
|
|
|
|
// Fallback: check if any indexed chapter slug or unique chapter directory appears in the path
|
|
for (const [slug, entry] of Object.entries(CHAPTER_TAXONOMY_INDEX)) {
|
|
const [, folder] = slug.split('/');
|
|
if (normalized.includes(slug)) {
|
|
return enrichChapterEntry(entry);
|
|
}
|
|
if (folder && normalized.includes(`/${entry.volume}/`) && normalized.includes(folder)) {
|
|
return enrichChapterEntry(entry);
|
|
}
|
|
}
|
|
|
|
return null;
|
|
}
|
|
|
|
function enrichChapterEntry(entry) {
|
|
if (!entry) return null;
|
|
const primaryGroupInfo = TAXONOMY_GROUPS[entry.primaryGroup] || null;
|
|
const subjectDetails = (entry.subjects || [])
|
|
.map((code) => TAXONOMY_SUBJECTS[code])
|
|
.filter(Boolean);
|
|
return {
|
|
...entry,
|
|
groupTitle: primaryGroupInfo?.name || entry.primaryGroup,
|
|
primaryGroupInfo,
|
|
subjectDetails,
|
|
};
|
|
}
|
|
|
|
/**
|
|
* Returns a concise (~180-240 char) grounding line when a chapter matches,
|
|
* or a general 1-line 5-Group overview when on a non-chapter page.
|
|
*/
|
|
export function formatChapterTaxonomyPrompt(urlOrPath = '') {
|
|
const chapter = resolveChapterTaxonomy(urlOrPath);
|
|
if (chapter) {
|
|
const shortCode = chapter.slug.replace(/^((?:vol[1-4])\/[0-9]{2})_.*$/, '$1');
|
|
const topSubjects = chapter.subjectDetails
|
|
.slice(0, 3)
|
|
.map((s) => `${s.code} ${s.name}`)
|
|
.join(', ');
|
|
return `Active chapter ML Systems Taxonomy (${shortCode} ${chapter.title}): Groups ${chapter.groups.join(', ')} | Subjects ${topSubjects}. Situate explanations and trade-offs across Foundations (F), Stack (S), Lifecycle (P), Qualities (Q), and System Types (T).`;
|
|
}
|
|
return 'ML Systems Taxonomy lens: situate mechanisms and trade-offs across the 5 Groups — Foundations (F), Stack (S), Lifecycle (P), Qualities (Q), and System Types (T).';
|
|
}
|
|
|
|
/**
|
|
* Returns matching taxonomy items (starting with a '🧭 Current chapter taxonomy overview'
|
|
* item and the active chapter's subjects, followed by matching subjects from TAXONOMY_SUBJECTS)
|
|
* for use in + Context and @ mentions.
|
|
*/
|
|
export function searchTaxonomyItems(query = '', currentUrl = '') {
|
|
const normalizedQuery = String(query || '').trim().toLowerCase();
|
|
const chapter = resolveChapterTaxonomy(currentUrl);
|
|
const results = [];
|
|
const seenKeys = new Set();
|
|
|
|
const overviewTitle = 'Current chapter taxonomy overview';
|
|
const overviewContent = chapter
|
|
? `${formatChapterTaxonomyPrompt(currentUrl)}\n\nPrimary Group: ${chapter.primaryGroup} (${TAXONOMY_GROUPS[chapter.primaryGroup]?.name || chapter.primaryGroup})\nSubjects: ${chapter.subjectDetails.map((s) => `${s.code} ${s.name}`).join(', ')}`
|
|
: 'ML Systems Taxonomy Overview: Foundations (F: workloads, requirements, models, performance models), Stack (S: hardware, compilers, parallelism, fleet), Lifecycle (P: data, training, evaluation, serving, MLOps), Qualities (Q: performance, reliability, security, safety, responsible AI, sustainability), System Types (T: agentic and embodied/physical systems).';
|
|
|
|
const matchesOverview = !normalizedQuery
|
|
|| 'taxonomy'.includes(normalizedQuery)
|
|
|| overviewTitle.toLowerCase().includes(normalizedQuery)
|
|
|| (chapter && (
|
|
chapter.slug.toLowerCase().includes(normalizedQuery)
|
|
|| chapter.title.toLowerCase().includes(normalizedQuery)
|
|
));
|
|
|
|
if (matchesOverview) {
|
|
results.push({
|
|
type: 'taxonomy',
|
|
icon: '🧭',
|
|
code: 'overview',
|
|
key: overviewTitle,
|
|
title: overviewTitle,
|
|
content: overviewContent,
|
|
value: {
|
|
title: overviewTitle,
|
|
content: overviewContent,
|
|
text: overviewContent,
|
|
},
|
|
});
|
|
seenKeys.add(overviewTitle);
|
|
}
|
|
|
|
function makeSubjectItem(subject) {
|
|
const title = `${subject.code} ${subject.name}`;
|
|
const groupInfo = TAXONOMY_GROUPS[subject.group];
|
|
const content = `${subject.code} ${subject.name} [Group ${subject.group} ${groupInfo ? groupInfo.name : ''} · ${subject.areaCode} ${subject.areaName}]: ${subject.summary}`;
|
|
return {
|
|
type: 'taxonomy',
|
|
icon: '🧭',
|
|
code: subject.code,
|
|
group: subject.group,
|
|
key: title,
|
|
title,
|
|
content,
|
|
value: {
|
|
title,
|
|
content,
|
|
text: content,
|
|
},
|
|
};
|
|
}
|
|
|
|
function subjectMatches(subject) {
|
|
if (!normalizedQuery || 'taxonomy'.includes(normalizedQuery)) return true;
|
|
const haystack = `${subject.code} ${subject.name} ${subject.areaCode} ${subject.areaName} ${subject.summary} ${TAXONOMY_GROUPS[subject.group]?.name || ''}`.toLowerCase();
|
|
return haystack.includes(normalizedQuery);
|
|
}
|
|
|
|
if (chapter) {
|
|
for (const subject of chapter.subjectDetails) {
|
|
const item = makeSubjectItem(subject);
|
|
if (!seenKeys.has(item.key) && subjectMatches(subject)) {
|
|
results.push(item);
|
|
seenKeys.add(item.key);
|
|
}
|
|
}
|
|
}
|
|
|
|
for (const subject of Object.values(TAXONOMY_SUBJECTS)) {
|
|
const item = makeSubjectItem(subject);
|
|
if (!seenKeys.has(item.key) && subjectMatches(subject)) {
|
|
results.push(item);
|
|
seenKeys.add(item.key);
|
|
}
|
|
}
|
|
|
|
return results;
|
|
}
|