AI, Machine Learning & LLM Foundations

Outcomes and course terminology rules

Learning path

Block 1

1. AI, ML, deep learning, and generative AI

Artificial Intelligence
├── rule-based, search, planning, knowledge, optimization, robotics …
└── Machine Learning
    ├── classical statistical and algorithmic ML
    └── Deep Learning
        ├── discriminative models
        └── generative models
            ├── diffusion and other families
            └── foundation models
                └── large language models

These are overlapping levels, not interchangeable labels. AI describes the broad capability and field. ML learns behavior from data or experience. Deep learning is ML using multilayer neural networks. Generative AI produces new content. A foundation model is broadly reusable. An LLM is a large language model, usually Transformer-based.

Block 2

2. How machine learning learns

Collect and document data
  → clean, transform, label, and split
  → select architecture and objective
  → initialize parameters
  → train on batches
  → validate and tune hyperparameters
  → evaluate once on held-out test conditions
  → release a specific artifact
  → monitor inputs, behavior, outcomes, cost, and drift
  → retrain, roll back, replace, or retire

Parameters versus hyperparameters

Parameters are learned values such as weights. Hyperparameters are selected configuration such as learning rate, batch size, regularization, architecture depth, and decoding settings. The distinction matters for provenance and incident reconstruction.

Generalization versus memorization

Training loss measures the optimization objective on training examples. Validation supports development choices. A test set estimates behavior on held-out data only if it has not become part of iterative tuning. Strong benchmark performance can coexist with poor deployment performance, privacy leakage, or adversarial fragility.

Block 3

3. Neural networks and optimization

input tensor
  → weighted transformation
  → activation function
  → hidden representations through multiple layers
  → output logits or values
  → loss against target
  → backpropagation computes gradients
  → optimizer updates parameters

A neuron or unit combines numeric inputs with learned weights and usually a bias, then applies an activation. Layers compose these transformations. A forward pass computes output. A loss function quantifies training error. Backpropagation applies the chain rule through the computation graph. An optimizer such as Adam uses gradients to update weights.

Block 4

4. Transformers, tokenization, and LLM generation

From text to tokens

A tokenizer converts raw text into token IDs using a versioned vocabulary and algorithm. Tokens may be words, fragments, bytes, punctuation, or special control markers. Tokenization changes length, cost, truncation, multilingual behavior, and attack representation.

From tokens to representations

Token IDs are mapped to embeddings. Positional information represents order. Transformer layers combine self-attention, feed-forward transformations, residual connections, and normalization to build contextual representations.

messages and application context
  → model-specific chat template
  → tokenizer → token IDs
  → token and position representations
  → repeated Transformer layers
      self-attention: query–key relevance weights value information
      feed-forward network: transforms each position
  → logits for possible next tokens
  → decoding policy selects one token
  → append token and repeat until stop condition

For one attention head, the common conceptual form is Attention(Q,K,V) = softmax(QKᵀ / √d) V. Learners do not need to derive it, but must understand that attention is learned weighted information flow—not a factuality or authorization mechanism.

Block 5

5. LLM lifecycle: from pre-training to serving

data collection and filtering
  → tokenizer and architecture selection
  → large-scale pre-training
  → base checkpoint
  → supervised / instruction fine-tuning
  → preference alignment: RLHF, DPO, or related methods
  → safety, capability, and task evaluation
  → optional domain adaptation: full fine-tune, LoRA/PEFT, adapters
  → quantization or distillation
  → registry and release approval
  → serving, routing, monitoring, updates, and retirement

Block 6

6. Embeddings and retrieval-augmented generation

RAG combines a generative model’s parametric knowledge with retrieved external information. Ordinary RAG does not retrain the LLM. It changes inference context.

OFFLINE / INGESTION
source → parse → normalize → split into chunks → preserve metadata and ACL
       → embedding model → vector / lexical index

ONLINE / QUERY
user and tenant identity → query → optional rewrite/decomposition
  → dense, sparse, or hybrid retrieval with authorization filters
  → optional reranking → selected chunks with provenance
  → prompt augmentation → LLM generation
  → citation and support validation → response

Block 7

7. Agents, tools, memory, and MCP

An agent is a system pattern, not a special magical model. It connects model output to state, tools, and repeated execution.

goal and initiating identity
  → observe context and state
  → plan or choose next step
  → propose tool and structured arguments
  → application validates policy and authorization
  → optional human approval bound to exact action
  → execute with scoped workload identity
  → observe result and write controlled memory
  → repeat within time, cost, step, and authority limits

MCP architecture

An MCP host is the AI application. It creates one or more clients, each connected to a server. MCP separates a JSON-RPC data layer from transport. Servers can expose three core primitives:

Clients may also support sampling and other capabilities. “Model-controlled” does not mean “model-authorized.” Applications must authenticate endpoints, validate schemas, enforce user and tenant authorization, constrain tools, display material actions, bind approvals, log definitions and calls, and control egress.

Memory types

Block 8

8. Evaluation, model serving, and MLOps

Evaluate the complete system

Serving path

client → authentication and quota → model gateway / router
  → prompt and policy processing → selected model runtime
  → accelerator and cache → streamed or complete output
  → validation / tool orchestration → telemetry and billing

MLOps adds versioning and governance for data, models, code, configuration, evaluation, registry stages, deployment, monitoring, rollback, and retirement. An incident record must identify the exact model, tokenizer, adapter, prompt template, retrieval index, tool definitions, policy, runtime, and route—not only a vendor model name.

Block 9

9. Practical: trace three architectures

Trace the same user question through three progressively capable systems:

1. a hosted chat LLM;
2. a tenant-aware RAG assistant;
3. a RAG agent that can create a ticket through MCP.

Terminology clinic

Correct this statement using the glossary:

“The AI learned our PDF during inference, stored it in its neural-network
database, reasoned deterministically, and securely called the MCP API.
The 95% accuracy score proves the agent is safe.”

10. Knowledge check