Model Mechanics

Today Explore Practice Library Exam Certificate

Noesis

Deep, finishable courses on prompting, psychology, money, health, and more — hands-on, in English and Khmer.

Learn

Today Explore Practice Library All courses

Certify

Take the exam Certificate

Account

My account Sign in

© 2026 Noesis. All rights reserved.Built for people who never stop learning.

Today Explore Practice Library Profile

Noesis — Deep, finishable courses.

How AI Models Work

It's not magic.
It's mechanics you can understand.

26 lessons on parameters, tokens, context windows, agents, and how Claude, GPT, and Gemini really differ.

💭 Honest take

26 evidence-backed lessons across 6 tracks. Six flagship engineering lessons now use a complete learning loop: production incident → system map → expert trace → attempt-first prediction → guided lab → retrieval practice and transfer close. The other 20 remain concise scenario lessons while this higher-rigor format is tested and expanded. The course avoids permanent provider scoreboards and teaches you to measure current models on task quality, Khmer performance, latency, cost, safety, and lifecycle. Lesson bodies are currently authored in English; Khmer localization still needs fluent review and parity evaluation.

Start learning Explore topics

26

lessons

26

real prompts

3

models covered

6

role perspectives

Current track 3. Tools, Agents, and Protocols0 / 4 · ~16 min left

Overall Progress: 0 / 26 lessons completed

✓Distinguish weights, training, inference, context, and product memory — then test which layer caused a behavior

✓Design tool and agent loops with validated arguments, least privilege, approval, idempotency, and verified completion

✓Evaluate multimodal and Khmer performance as separate tasks instead of trusting a generic capability label

✓Choose prompting, retrieval, fine-tuning, tools, or deterministic code from the measured failure each one can solve

✓Select and operate models with a reproducible eval, quality–latency–cost evidence, safety controls, routing, monitoring, and migration

Up next

The Agent Loop: Observe, Decide, Act, Check

An agent is a controlled software loop around a model—not a new species of model — Beginner

🎯 Learning Tracks

Curated paths through the course — pick one and finish it, instead of scrolling a thousand lessons.

0 / 6 complete

Start here01Inside a Language Model0/4 lessons·Start 02Controlling Model Behavior0/4 lessons·Start 03Tools, Agents, and Protocols0/4 lessons·Start 04Multimodal and Multilingual Systems0/4 lessons·Start 05Grounding, Adaptation, and Evaluation0/6 lessons·Start 06Choosing and Operating Models0/4 lessons·Start

Beginner

0 / 8

The Agent Loop: Observe, Decide, Act, Check

An agent is a controlled software loop around a model—not a new species of model

Write a Model Requirements Card

Choose from the task backward, not from the provider logo forward

Weights, Training, and Inference

When an AI behaves differently, find what changed before you blame the model

Multimodal Inputs: More Than Text

Separate the base model, modality encoder, API surface, and generation tool

Message Roles and Instruction Priority

Higher-priority text guides behavior; it does not create a security boundary

Tokens and Embeddings: Text Becomes Numbers

Measure English and Khmer instead of trusting a words-to-tokens slogan

Open-Weight and Local Deployment

Control increases, but so do licensing, serving, security, and total-cost responsibilities

Model Lifecycle: Version, Monitor, Migrate

A model change is a production change even when the endpoint stays the same

Intermediate

0 / 13

Hallucination, Uncertainty, and Verification

A plausible answer is not evidence, and a confidence sentence is not calibration

Function Calling: A Proposal, Not Execution

Schema-valid arguments can still be wrong, unauthorized, or dangerous

Run a Quality–Latency–Cost Bake-Off

Current numbers belong in a dated experiment, not permanent course prose

Documents and Screens: OCR, Layout, and Vision Failure Modes

The right crop or text layer can outperform a larger screenshot

In-Context Learning: Instructions, Examples, and Ambiguity

A prompt conditions behavior; a test set tells you whether it worked

Embeddings, Retrieval, and RAG

Retrieve evidence first, then evaluate retrieval and generation separately

Execution Owners and Trust Boundaries

Map credentials, data, approval, and side effects—not merely “client” versus “server”

Routing, Fallbacks, and Failure Budgets

One model does not need to handle every request—but every route needs a reason

Attention and Context: What the Model Can Use

Capacity is not recall, and a long window is not a database

Multilingual Models: Token Cost, Transfer, and Khmer

A Khmer interface is not a Khmer-quality evaluation

Sampling and Reproducibility

Understand the probability controls, then verify which controls your API still honors

Next-Token Generation: From Logits to a Response

A fluent answer is a sequence of local choices, wrapped by a serving system

1–12 of 13

Page 1 of 2

Advanced

0 / 4

Prompt, RAG, Fine-Tune, or Tool?

Choose the lever from the failure—not from which technology sounds serious

MCP and Delegation: Two Different Architecture Problems

Standardizing tool access does not make a system multi-agent—or safe

Computer Use: Perception–Action Loops

Every click depends on what the system saw, inferred, and was allowed to do

Reasoning Models: More Compute, Not Magic

Spend extra inference only where it improves a measured outcome

Expert

0 / 1

Capstone: Defend a Model Selection Decision

Produce the evidence another team could challenge, rerun, and operate