A Masterclass on AI: How do LLMs Work and how does is Impact Investing?

Classical ML required manual feature engineering

Traditional models like linear regression depended on humans selecting relevant input features (bedrooms, bathrooms) and failed to process unstructured data such as images, text, or video.

Three distinct learning paradigms

Supervised learning uses labeled examples like house prices, unsupervised learning discovers hidden patterns in unlabeled data like customer segments, and reinforcement learning optimizes through reward/penalty systems used in AlphaGo and self-driving vehicles.

Scale constraints limited early models

Classical machine learning ran on single CPUs due to matrix multiplication requirements, restricting model complexity until cloud computing enabled distributed processing across multiple GPUs.

Neural networks use hierarchical feature extraction

Deep learning models automatically identify patterns through layered processing—detecting edges in layer one, shapes in layer two, and objects in layer three—eliminating the need for manual feature selection.

Parameters vs. weights explained

When Meta releases 'Llama 3B,' the number refers to 3 billion parameters (variables like X, Y, Z); training generates weight coefficients (A, B, C) that constitute the model's proprietary intellectual property and secret sauce.

LLMs are massive linear equations

Despite generating human-like text, models like GPT-4 with approximately 530 billion parameters essentially function as enormous linear equations solving for optimal coefficients through mathematical optimization.

From explicit programming to context orchestration

Traditional software required programmers to write database queries and deterministic rules; modern AI agents independently determine which tools to call by interpreting large context windows without explicit human instruction.

Reasoning models bridge the capability gap

The critical innovation between basic LLMs and agents is reasoning capability—models now figure out intermediate steps and orchestrate tool calls rather than simply predicting the next token in a sequence.

Code follows fixed rules making it vulnerable

Unlike poetry or prose, programming relies on rigid syntax and repeatable patterns, making software engineering particularly susceptible to automation as AI masters deterministic rule-based tasks.