Introduction

As developers, we face a relentless stream of emerging tech, with Artificial Intelligence sitting squarely at the center of the noise. It’s easy to read high-level summaries and feel like you understand a concept—until you try to build it yourself.

To move past the hype, I built three distinct side projects targeting three core pillars of modern AI: Machine Learning (ML), Large Language Models (LLMs), and Artificial Super Intelligence (ASI). Here is what each concept actually looks like under the hood, and what building them taught me about the limits of today's systems.

1. Machine Learning: It’s Just Matrix Multiplications

Let’s start at the foundation. What actually happens when a machine 'learns'?

A classic entry point is the MNIST handwritten digit classification challenge: given 60,000 28x28 grayscale images of handwritten numbers (0–9), write a program that learns to recognize them.

Strip away the marketing terminology, and the solution isn't magic—it's high-dimensional curve fitting via matrix multiplication. The process breaks down into two core phases:

  • Learning Phase: The model iterates through a training subset of images, comparing its predictions against ground-truth labels and adjusting numerical parameters (weights) via backpropagation to minimize classification error.
  • Inference Phase: The model locks those optimized weights and applies them to new, unseen images to evaluate real-world performance.

Seeing a Convolutional Neural Network (CNN) classify digits with over 98% accuracy demonstrates how linear algebra can mimic pattern recognition.

Code & Deep Dive: Check out my implementation on GitHub: github.com/p-dirac/python-keras-cnn

2. Large Language Models: Compression, Weights, and Software Engineering

Scaling up linear algebra brings us to Large Language Models (LLMs). While LLMs rely on the same fundamental matrix operations as simple classifiers, the scale and architecture (Transformers) change what they can accomplish.

A helpful mental model for an LLM is a massive digital library compressed into a non-deterministic matrix. When an LLM is trained on millions of documents:

  • Tokenization: Raw text is broken down and converted into numerical vectors (tokens).
  • Weight Synthesis: Instead of storing the text directly, the architecture encodes semantic connections and structural relationships across billions of parameters.
  • Generative Reconstruction: When you prompt the model, it predicts probability distributions over tokens, converting latent weights back into coherent text.

Beyond drafting articles, LLMs can model formal structured logic—like synthesizing financial filings, validating regulatory schemas, and authoring domain-specific code.

Code & Deep Dive: Explore my SEC compliance and document generation prototype: github.com/p-dirac/disclosure-pilot-ai

3. Artificial Super Intelligence (ASI): Can AI Out-Think Physicists?

While ML classifies and LLMs synthesize, Artificial Super Intelligence (ASI) is hypothesized to invent entirely new conceptual models. Most discussion around ASI remains abstract, treating it as simply a 'smarter LLM.' But can today's architectures genuinely perform creative discovery?

Consider the history of physics: Kepler’s planetary laws, Newton’s gravity, Maxwell’s electromagnetism, the Bohr atom, Einstein’s relativity, and Schrödinger’s quantum mechanics. Could an AI discover these breakthroughs without human guidance?

To put this to the test, I focused on a specific 19th-century anomaly: the unexplained precession shift in Mercury's orbit. Newtonian gravity failed to account for the observed measurements, leading historical physicists to propose modified force laws (such as Hall, Gerber, and Yukawa potentials) before Einstein transformed the framework with General Relativity.

When evaluating whether current AI models can discover novel physical laws to resolve such anomalies, the answer is clear: current pattern-matching engines cannot bridge conceptual gaps that require discovering fundamental paradigm shifts.

Code & Deep Dive: Read my analysis on alternative gravity models and ASI boundaries: github.com/p-dirac/physics-asi-concepts

Conclusion

Building across these three tiers highlights both the capabilities and limits of modern AI:

  • Machine Learning excels at numerical pattern classification.
  • Large Language Models excel at semantic synthesis and structured automation.
  • Artificial Super Intelligence remains an unsolved challenge requiring breakthroughs beyond probabilistic next-token prediction.

If you want to dive deeper into the code, mathematical derivations, or project documentation, feel free to explore the GitHub repositories linked above.

"What are your thoughts on current AI limitations in scientific discovery? Let’s discuss in the comments below!"