Personal Research NotebookVol. I · 2026
Bereket Atakilt · ሕሩይ

Bereket Atakilt

Heruy

Ethiopian Software and AI/ML Engineer

My name is Bereket Atakilt but you can call me Heruy. In Ge'ez Heruy or Hiruy means 'the Chosen one' or 'the Exalted one'. I am an Ethiopian software and AI/ML engineer working, building and learning around machine learning, language models and AI agents, computational systems, simulations and reinforcement learning with one unifiying theme - the strange ways simple rules can produce complex behavior.

I intended this site to be a sort of journal or logbook as I explore these ideas in English and code.

Ahem...

... Nature, complexity, life, intelligence and consciousness. Their unifying theme being simple constituents and simple rules resulting in complexity. Themes, unifying ones, are becoming a unifying theme in this writing, Dear Reader and recurrence too has been a recurring one.

Read essay

Selected experiments, implementations, and computational studies.

፩

Language models

LilGPT

A small GPT implementation built to understand tokenization, embeddings, attention, transformer blocks, training, and autoregressive generation.

Python · Transformers · Deep Learning

፪

Deep learning tools

Lil Torch (not a rapper)

A minimal PyTorch-inspired deep learning framework exploring tensors, automatic differentiation, computational graphs, modules, and optimization.

Python · Autodiff · Systems

፫

Reinforcement learning · Evolution

Natural Selection 2D

An experimental computational ecosystem where agents evolve and learn to survive in a 2D world using neural-network policies shaped by evolutionary fitness.

Python · Simulation · Evolutionary computation

View all projects
፩

Ahem...

... Nature, complexity, life, intelligence and consciousness. Their unifying theme being simple constituents and simple rules resulting in complexity. Themes, unifying ones, are becoming a unifying theme in this writing, Dear Reader...

View all writing

Questions before answers.

Artificial intelligenceLanguage modelsReinforcement learningComplex systemsScientific computingComputation