Building an LLM from scratch in 60 minutes
Armed with nothing but C# and the .NET libraries, we'll endeavor on a whirlwind tour to build a toy GPT-like LLM from scratch.
Expect a hardcore deep dive coding session that demystifies the essential building blocks of LLMs, including tokenization, embeddings, layers, self-attention, transformers, auto-differentiation, gradient descent, and more.
By the end of the session, we'll train our LLM on a corpus of text and attempt to have a meaningful - or gibberish - conversation with it.
About the speaker
Bart De Smet
Originally from Belgium, Bart relocated to work at the Microsoft headquarters in 2007 after graduating from Ghent University while being a C# MVP.
After having worked on WPF and .NET for a few years, Bart focused on reactive programming for quite some time, helping to build Reactive Extensions (Rx) and large scale event processing systems powering various cloud services.
After a brief stint in AutoML focusing on Python performance tuning, Bart is now focusing on building next generation distributed systems in the office of the CTO.