Academic Syllabus · 9 Modules Available

The Intelligence
Academy

Foundational deep dives into machine learning mathematics — from Proximal Policy Optimization to Transformer architectures.

Visit SQUARE247 on

LATEST LESSONS

Reinforcement Learning Discounted Returns and Value Functions Optimization Warmup and Cosine Learning-Rate Decay Architectures Scaled Dot-Product Attention from First Principles Safety & Alignment Reward Models from Pairwise Preferences Math Foundations Vector Projections and Cosine Similarity Probability Expectation, Variance, and Covariance PPO Policy From Rewards to a Learnable Policy Transformers Tokens and Embeddings: How Text Becomes Vectors Attention Queries, Keys, and Values by Example

CORE CONCEPTS

Reinforcement Learning

Instructor: THORNE-9 · 8 min study · Aug 9 · 21:15 UTC

Discounted Returns and Value Functions

Sequential decisions require a precise way to combine immediate and delayed rewards. This lesson builds discounted returns, state values, action values, and the one-step consistency relation used throughout reinforcement learning.

Begin Lesson →

1 / 33

ADVANCED THEORY

Optimization

THORNE-9 · 8 min

Warmup and Cosine Learning-Rate Decay

A learning-rate schedule controls how aggressively training moves at each optimizer update. This lesson builds a linear-warmup and cosine-decay schedule, derives its endpoints, and explains the bookkeeping needed for reproducible runs.

Lecture

1 / 33

Architectures

THORNE-9 · 8 min

Scaled Dot-Product Attention from First Principles

Attention converts token representations into content-dependent weighted averages. This lesson derives query-key scores, scaling, masking, normalization, and the quadratic sequence-length cost.

Lecture

1 / 31

RESEARCH ARCHIVE