Skip to content
Adrian Mircea Nenu computo, ergo sum
← Research & Projects

// Project

mvGPT - Minimum Viable GPT

A language model from scratch, to understand one from the ground up: a GPT defined in PyTorch and taken through the whole training pipeline - pretraining, fine-tuning and reinforcement learning - with the evaluation and metrics that show what it learned, then deployed so anyone can talk to it in the browser. Built episode by episode on YouTube, with the models on Hugging Face.
PyTorch LLMs RL Video series

Latest videos

YouTube ↗

Latest writing

Substack ↗

Training a 14M GPT to Generate Children Stories in 20 Minutes: Pretraining, SFT, and Direct Preference Optimization

End to end implementation, parameter and architecture analysis, and light evaluation of the checkpoints

Read on Substack ↗

Could Actor Systems Be All We Need for Agentic AI Workflows

Are We Reinventing the Wheel of Actor Systems as We Develop Agentic Orchestration for AI?

Read on Substack ↗

Is Computer Science a Worthwhile Pursuit Today?

As of 2026, It Will Most Certainly Endure the Test of Time, Albeit With a Raised Bar of Expectations

Read on Substack ↗