Sources worth spending time with.

Language-model training, inference, and the distinction between pretraining and further training.

Suggested depth: Technical

An intuitive introduction to numerical representations and attention. The original encoder–decoder architecture is historical; modern models vary.

Suggested depth: Intermediate

Tokenization, numerical IDs, and subword representations.

Suggested depth: Intermediate

Selecting and maintaining context, rather than merely accumulating conversation and tool outputs.

Suggested depth: Intermediate

Position-sensitive performance in the specific long-context tasks and models studied; not a current benchmark for all models.

Suggested depth: Technical

Limitations of pretrained models and the persistence of bias.

Suggested depth: Beginner

Clear instructions, examples, and separating instructions from context. Consult the current linked guidance for model-specific details.

Suggested depth: Practical

Recording changes, comparing versions, and restoring prior states.

Suggested depth: Beginner

Explicit progress records and recoverable artifacts across sessions.

Suggested depth: Intermediate

The original research architecture combining retrieval with generation. Modern application pipelines may differ.

Suggested depth: Technical

Text extraction, reading-order limitations, layout, and tables.

Suggested depth: Practical

Distinguishing workflows from agents and decomposing tasks into controlled stages. Architectural concepts are more durable than its tool examples.

Suggested depth: Intermediate

Review Table functionality and changes. Feature availability depends on the account and current product.

Suggested depth: Practical

Speech recognition as a separate model task from subsequent rewriting.

Suggested depth: Intermediate

Packaging reusable instructions and resources, with information loaded as needed.

Suggested depth: Intermediate

Direct and indirect prompt injection and layered mitigation concepts.

Suggested depth: Intermediate

OCR commands, rotation, deskew, sidecar limitations, and precautions around image processing and signed PDFs.

Suggested depth: Practical

Embedding-based passage retrieval and similarity search.

Suggested depth: Intermediate

A candidate-retrieval stage followed by a more focused relevance-ranking stage.

Suggested depth: Intermediate

Retrieval in application and agent architectures; inspect the implementation actually used.

Suggested depth: Technical

Referential relationships and enabling foreign-key enforcement.

Suggested depth: Technical

Numerical arrays, tensor operations, and device placement.

Suggested depth: Technical

Lower-precision numerical representations and memory tradeoffs.

Suggested depth: Intermediate

Configured context and its relationship to memory requirements.

Suggested depth: Practical

Installation entry point and basic local inference workflow.

Suggested depth: Practical

Running, listing, and inspecting local models.

Suggested depth: Practical

ollama ps and the PROCESSOR column, model storage locations, and disabling cloud features.

Suggested depth: Technical

Local inference implementation, formats, and supported hardware backends.

Suggested depth: Technical

Filesystem paths, file traversal, and file metadata.

Suggested depth: Practical

CSV writing and quoting values correctly.

Suggested depth: Practical

Further study of paths and file operations through practical programming examples.

Suggested depth: Practical

Client–server requests and responses.

Suggested depth: Beginner

Encoding and decoding JSON data.

Suggested depth: Practical

Constructing and sending HTTP requests from Python.

Suggested depth: Practical

Local API usage; the linked API documentation covers endpoint payloads and streaming.

Suggested depth: Practical

In-memory databases, parameterized statements, and Python's SQLite interface.

Suggested depth: Practical

Parameter-efficient adaptation using trainable update matrices with frozen original weights.

Suggested depth: Technical

Containers and their role in packaging and running applications.

Suggested depth: Beginner

Persistence beyond a container's lifecycle; backup and restore require deliberate design.

Suggested depth: Practical

Defined tasks, graders, repeated trials, traces, and evaluating actual outcomes.

Suggested depth: Intermediate

Entry point to current professional technology and generative-AI guidance. Linked resource listing located 15 October 2026; the page blocked automated full-text access.

Suggested depth: Practical

Loss, gradients, and optimization in a training loop.

Suggested depth: Technical

Decoding, sampling, and generation settings.

Suggested depth: Intermediate

Persistent project instructions and their distinction from enforced configuration. Reviewed 15 October 2026.

Suggested depth: Intermediate

How the Rules of Professional Conduct apply to generative-AI-assisted work: competence, confidentiality, supervision, and candour. The paper is guidance, not a rule.

Suggested depth: Practical

The 1 December 2024 amendments adding certification of the authenticity of authorities for factums (r. 4.06.1) and expert reports (r. 53.03), and the presumption for authorities from recognized sources. Confirm the current rule text before filing.

Suggested depth: Practical

A survey of certification and disclosure requirements across Canadian courts and of early decisions on fabricated citations, including Zhang v. Chen, 2024 BCSC 285. Requirements change; verify the current notice for the relevant court.

Suggested depth: Practical

A summary of Ko v. Li, 2025 ONSC 2766, and of the Federal Court notice on the use of AI in court proceedings (updated May 2024). Read the decision itself before relying on any proposition from it.

Suggested depth: Practical

Summaries of the Canadian Judicial Council Guidelines for the Use of Artificial Intelligence in Canadian Courts (24 October 2024) and the Federal Court interim principles on its own use of AI (September 2025).

Suggested depth: Practical

A compilation of Canadian decisions in which courts addressed fabricated or unverified AI-generated material. Useful as a catalogue of fact patterns; read each decision before citing it.

Suggested depth: Practical

Install routes, the llama cli and llama serve quick start, supported backends, and CPU+GPU hybrid inference.

Suggested depth: Technical

winget, Homebrew, and conda-forge install commands.

Suggested depth: Technical

CMake build commands for CPU, CUDA, Vulkan, HIP, and Metal, and the Metal --n-gpu-layers 0 note.

Suggested depth: Technical

Flags including -hf, -ngl (auto or all), --list-devices, --offline, --cpu-moe, and KV cache types.

Suggested depth: Technical

llama-server defaults (host 127.0.0.1, port 9931) and the LLAMA_CACHE model cache.

Suggested depth: Technical

Benchmark syntax: prompt-processing (pp) and text-generation (tg) tests.

Suggested depth: Technical

Pre-built binaries for Windows CUDA, Vulkan, ROCm, and CPU, macOS, and Linux.

Suggested depth: Technical

Install commands for macOS, Linux, and Windows and the ollama run command.

Suggested depth: Technical

Windows requirements: NVIDIA driver 551.61+, AMD ROCm v7 or Vulkan; OLLAMA_MODELS.

Suggested depth: Technical

Supported NVIDIA compute capabilities and AMD Radeon models by platform.

Suggested depth: Technical

Apple Silicon and macOS 14+; Windows x64 with AVX2 or ARM; 16 GB RAM and 4 GB VRAM recommended.

Suggested depth: Technical

lms ls, lms ps, lms load --gpu, and lms server start.

Suggested depth: Technical

The ~/.lmstudio/models directory structure.

Suggested depth: Technical

pip install mlx-lm, mlx_lm.generate and mlx_lm.chat on Apple silicon.

Suggested depth: Technical

pip install open-webui (Python 3.11) and open-webui serve on localhost:8080.

Suggested depth: Technical

Desktop application for Mac, Windows, and Linux; built-in local engine; connects to Ollama and LM Studio.

Suggested depth: Technical

uv pip install vllm; serving-oriented engine.

Suggested depth: Technical

Creating outbound block rules by program in Windows Firewall with PowerShell.

Suggested depth: Technical

Memory size, bus width, and memory bandwidth for the GeForce RTX 50 series.

Suggested depth: Technical

575 W total graphics power, 1,000 W required system power, power connector options.

Suggested depth: Technical

32 GB GDDR6, 256-bit, 640 GB/s, 300 W, 750 W PSU recommendation.

Suggested depth: Technical

Unified memory options and memory bandwidth for M5 Max and M5 Ultra.

Suggested depth: Technical

Unified memory options and memory bandwidth for M6 and M5 Pro.

Suggested depth: Technical

256-bit LPDDR5x-8000 memory, up to 128 GB, Radeon 8060S graphics.

Suggested depth: Technical

64 or 128 GB unified LPDDR5x, 256-bit, 273 GB/s.

Suggested depth: Technical

40 layers, 8 KV heads, head dimension 128; GGUF Q4_K_M 14.33 GB and Q8_0 25.05 GB in the lmstudio-community repository.

Suggested depth: Technical

35B total, 3B activated, 256 experts with 8 routed and 1 shared; Apache 2.0.

Suggested depth: Technical

gpt-oss-20b: 21B parameters with 3.6B active; gpt-oss-120b: 117B with 5.1B active; Apache 2.0.

Suggested depth: Technical

Dense and mixture-of-experts sizes; 26B A4B with 3.8B active; Apache 2.0.

Suggested depth: Technical

The small model used for the measured benchmark in lesson 9.9.

Suggested depth: Technical

AI oversight and verification; paragraph125 reproduces factum certification rule. Updated March17 2026; civil scope excludes Divisional and Small Claims proceedings. Primary page checked16 October2026.

Suggested depth: Practical

Declaration for directly AI-generated litigation content; suggestions/critique distinction and separate treatment of tribunal records/expert reports. Primary notice checked16 October2026; recheck before filing.

Suggested depth: Practical

Factum authenticity certification and listed-source presumption. Direct e-Laws access was blocked in this review; relevant rule wording checked in official SCJ practice direction para125.

Suggested depth: Practical

Court-specific policy, distinct from the notice to litigants. Do not treat it as permission for another court or as legal advice.

Suggested depth: Practical

Conceptual foundations and starting with a bounded agent. Webpage dated January 15, 2026. Legal examples and illustrative traces are independent Case & Code additions, not executed SDK examples.

Suggested depth: Technical

Runner loop, tool results, handoffs, final output, and maximum turns. A terminated run does not establish task correctness. Reviewed October 16, 2026.

Suggested depth: Technical

Configured approval interruptions before tool execution; not a guarantee that every platform or tool requests approval. Reviewed October 16, 2026.

Suggested depth: Technical

Transfer of control, configurable history, and distinction from agents used as tools. Reviewed October 16, 2026.

Suggested depth: Technical

§3, Figure 3, Appendix B.1 and §5. Selective clarification in older-model question-answering experiments; does not validate the legal-drafting adaptation.

Suggested depth: Technical

§§2.2–2.3 and §4.2. Behavioral test construction and NLP studies; not a measured result for model-generated legal test cases.

Suggested depth: Technical

§§3.2–3.4, Tables 1–4 and §8. Context-separated verification in the authors’ experiments; does not validate the source-grounded legal adaptation.

Suggested depth: Technical

Base material adapted and rewritten for these lessons: LLM basics, history, hidden prompts, tokens and limits, prompt hacking, give-a-bot-a-fish and teach-a-bot-to-fish, data formats, citations, chain of thought, fine-tuning. Copyright (c) 2023 Brex; used under the MIT License. Model names, limits and prices in it date from 2023.

Suggested depth: General

What an n-gram language model is and its sparsity problem.

Suggested depth: Technical

Layer-wise training method that helped make deeper networks trainable.

Suggested depth: Technical

The LSTM architecture.

Suggested depth: Technical

Character-level recurrent networks generating text and code.

Suggested depth: General

The transformer architecture and attention.

Suggested depth: Technical

Unsupervised pretraining followed by task fine-tuning.

Suggested depth: Technical

GPT-3; zero-, one- and few-shot task performance from prompts alone.

Suggested depth: Technical

The human-preference learning technique underlying RLHF.

Suggested depth: Technical

InstructGPT; RLHF applied to instruction following.

Suggested depth: Technical

ChatGPT release and its RLHF training.

Suggested depth: General

GPT-4 release.

Suggested depth: Technical

Meta LLaMA models released to researchers.

Suggested depth: Technical

The emergent-abilities claim.

Suggested depth: Technical

Argument that some apparent emergence reflects the choice of metric.

Suggested depth: Technical

Interactive token counting; rule of thumb of about four characters per English token.

Suggested depth: Technical

Open-source tokenizer library; used with the o200k_base encoding for the lesson 1.2 figure (run 17 October 2026).

Suggested depth: Technical

Why subword tokenization is used; how tokenizers differ.

Suggested depth: Technical

Definition of context window; 1M-token windows and 128k output for current Claude models, 200k for older ones; context rot; thinking tokens count toward the window.

Suggested depth: Technical

XML tags; long documents at the top and query at the end (up to 30% improvement in Anthropic tests); quote extraction for long documents.

Suggested depth: Technical

Keep prompts simple and direct; "think step by step" unnecessary for reasoning models; use delimiters; try zero-shot first.

Suggested depth: Technical

Few-shot chain-of-thought prompting.

Suggested depth: Technical

"Let's think step by step" zero-shot chain of thought.

Suggested depth: Technical

Sampling several reasoning paths and taking the majority answer.

Suggested depth: Technical

Reasoning before answering; thinking tokens and budgets.

Suggested depth: Technical

Visible chain of thought often omitted hints that influenced answers.

Suggested depth: Technical

Thought–action–observation loop.

Suggested depth: Technical

How tool use (function calling) works; client and server tools.

Suggested depth: Technical

Defining functions with JSON schemas and handling model calls.

Suggested depth: Technical

Schema-compliant responses through constrained decoding; strict tool use.

Suggested depth: Technical

Structured Outputs versus JSON mode: only Structured Outputs ensures schema adherence.

Suggested depth: Technical

Open standard for connecting AI applications to tools and data.

Suggested depth: Technical

Searching with the embedding of a model-written hypothetical document.

Suggested depth: Technical

Supported methods (SFT, DPO, RFT, vision); notice that the fine-tuning platform is winding down and closed to new users.

Suggested depth: Technical

System prompts are not secrets or security controls; keep credentials out; enforce controls outside the model.

Suggested depth: Technical

Indirect prompt injection against real LLM-integrated applications.

Suggested depth: Technical

Prompt-level defenses such as instruction repetition and separating inputs.

Suggested depth: General

Client–server requests and responses.

Suggested depth: Technical

Lower-precision numerical representations and memory tradeoffs.

Suggested depth: Technical

AlexaTM 20B outperforming GPT-3 on some few-shot tasks with far fewer parameters.

Suggested depth: Technical

Worked examples and guides for building with models.

Suggested depth: Technical

Survey of prompting techniques with links to research.

Suggested depth: General