ASKiIMAN™ REALM 14 / COMPUTING
--:-- · -- --- ----

The Intelligence Engine

AI, Machine Learning, Neural Networks & Agents

Central Doctrine

Traditional programs are explicitly instructed what to do. Artificial intelligence introduces systems that can learn patterns from data, make predictions, recognize complex structures, and generate novel content. This realm maps the transition from rigid, rule-based algorithms to probabilistic neural networks, Large Language Models, and autonomous agents capable of intelligent behavior.

DATA → REPRESENTATION → LEARNING → MODEL → INFERENCE → PREDICTION / GENERATION → ACTION → FEEDBACK

The Internal Chambers

Ontology of Machine Intelligence

CHAMBER 01

What is AI?

🧠
→
PREDICTGENERATEDECIDE
  • Machine intelligence, reasoning, and learning
  • Perception, prediction, and decision making
  • Language understanding and automation
CHAMBER 03

Trad. Code vs Machine Learning

RULES+DATA
→
CODE
→OUT
DATA+OUT
→
TRAIN
→MODEL
  • Traditional: Human writes rules to process data.
  • Machine Learning: Machine discovers rules from data and examples.
  • Deep Learning: ML using multi-layer neural networks.
CHAMBER 04

Machine Learning

⚙️
  • Learning algorithms extracting patterns
  • Features (Inputs) and Labels (Expected Outputs)
  • Training the model to generalize to unseen data
  • Prediction and Inference
CHAMBER 06

Supervised Learning

[1.2, 3.4]→"CAT"
[5.1, 0.9]→"DOG"
  • Learning from Labeled data
  • Mapping Features (X) to Targets (y)
  • Minimizing Loss to improve predictions
  • Classification and Regression
CHAMBER 07

Classification

  • Predicting discrete class labels (Categories)
  • Binary (Spam / Not Spam) vs. Multiclass
  • Learning a Decision Boundary to separate classes
  • Probability prediction and thresholds
CHAMBER 08

Regression

  • Predicting continuous numerical values (e.g., Price)
  • Linear and Polynomial regression
  • Minimizing Residuals (Mean Squared Error)
CHAMBER 09/10

Unsupervised & Clustering

  • Learning from Unlabeled data
  • Discovering hidden patterns and density
  • K-means, Hierarchical clustering, DBSCAN
  • Customer segmentation and anomaly detection
CHAMBER 12

Reinforcement Learning

AGENT
→ ACT →
ENV
STATE
←+1 REWARD←
  • Learning via trial and error to maximize Reward
  • Agent, Environment, State, Action, Policy
  • Exploration vs. Exploitation trade-off
CHAMBER 17

Train / Val / Test

TRAIN
VAL
TEST
  • Train: Learn parameters.
  • Validation: Tune hyperparameters.
  • Test: Final unbiased evaluation.
  • Preventing data leakage and testing generalization.
CHAMBER 18/19

Overfitting & Underfitting

  • Overfitting: Memorizing training data; fails to generalize (Test error rises).
  • Underfitting: Model lacks capacity to learn the pattern.
  • The Bias-Variance Trade-off.
CHAMBER 21/22

Loss & Optimization

  • Loss Function: Measures prediction error (MSE, Cross-entropy).
  • Gradient Descent: Iteratively updating parameters to minimize Loss.
  • Learning rate, Epochs, and Batches.
CHAMBER 25

Confusion Matrix

TP
FP
FN
TN
  • True Positives, False Positives (Type I), False Negatives (Type II)
  • Precision: Accuracy of positive predictions.
  • Recall: Ability to find all positive instances.
  • F1-Score and ROC-AUC.
CHAMBER 28/29

Trees & Ensembles

  • Decision Trees: Splitting data via feature thresholds.
  • Ensembles: Combining multiple weak models (Random Forest, Gradient Boosting) to create one strong predictor.
CHAMBER 33/36

Neural Networks

  • Input, Hidden, and Output layers
  • Deep Learning = Multiple hidden layers
  • Learning hierarchical representations of data
CHAMBER 34/35

The Artificial Neuron

Σ
  • Inputs × Weights + Bias → Weighted Sum
  • Activation Function: Adds non-linearity (ReLU, Sigmoid, Softmax) allowing complex pattern learning.
CHAMBER 37/38

Backpropagation

  • Forward Pass: Compute prediction and calculate Loss.
  • Backward Pass: Apply the Chain Rule of calculus to propagate error gradients backwards.
  • Update weights to minimize future error.
CHAMBER 40

CNNs (Vision)

→
  • Convolutional Neural Networks
  • Kernels/Filters slide over images to extract spatial features (edges, textures).
  • Core tech for Computer Vision (Image recognition).
CHAMBER 41

RNNs (Sequences)

  • Recurrent Neural Networks
  • Maintains a "hidden state" memory to process sequential data (Time, Text, Audio).
  • LSTMs and GRUs solve vanishing gradients.
CHAMBER 42/43

Transformers & Attention

FFN
ATTN
  • The architecture behind modern GenAI.
  • Self-Attention: Weighs the importance of every word in a sequence relative to every other word simultaneously.
  • Replaces sequential RNNs with parallel processing.
CHAMBER 44-46

Large Language Models

The cat sat on the mat.
  • Models trained on massive text to predict the most probable Next Token.
  • Tokenization converts text chunks to IDs.
  • Autoregressive probabilistic generation.
CHAMBER 47/48

Generative AI

"A dog"→
🧠
→
  • Foundation models capable of multimodal generation (Text, Image, Code).
  • Diffusion models (Images) vs. Autoregressive models (Text).
CHAMBER 49

Prompting

SYS: Act as an expert.
USR: Explain quantum.
AST: Quantum is...
  • Guiding foundation models without retraining.
  • Zero-shot vs. Few-shot prompting.
  • Chain-of-thought (forcing step-by-step reasoning).
CHAMBER 57/58

Transfer & Fine-Tuning

  • Pretraining: Expensive, unsupervised learning of general representations.
  • Fine-Tuning: Cheap, supervised updating of parameters on a specific target task.
  • Avoids training massive models from scratch.
CHAMBER 61

AI Explainability (XAI)

🔍
  • Deep neural networks act as "Black Boxes."
  • Interpretability vs. Performance trade-off.
  • Techniques (SHAP, LIME) to identify which features drove a specific prediction.
CHAMBER 77/78

AI Agents

ACT
  • Systems that combine Models, Memory, and Tool use.
  • Agent Loop: Observe → Plan → Act → Feedback.
  • Autonomous execution of multi-step tasks via API interaction.
CHAMBER 79-81

RAG & Embeddings

RETRIEVE→
LLM
  • Embeddings: High-dimensional vectors capturing semantic similarity.
  • Vector DBs: Search for nearest neighbors (relevant docs).
  • RAG: Injecting retrieved facts into LLM context to prevent hallucination.
CHAMBER 94

Model Lifecycle (MLOps)

DATA
→
TRAIN
→
DEPLOY
↑↓
RETRAIN
←
MONITOR / DRIFT
  • Models degrade as real-world data distributions shift (Data Drift).
  • Continuous monitoring, evaluation, and retraining pipeline.
FINAL CHAMBER

The Bridge to the Future

REALM 14 taught:
DATA → LEARNING → INTELLIGENCE → AGENT
↓
But intelligence does not exist in isolation.
How does AI alter HUMAN CIVILIZATION?
  • We have built machines that can learn, predict, and generate.
  • We now step back to look at the entire computing universe.
Proceed to REALM 15 → SYNTHETIC CIVILIZATION