Category
3 posts filed under this topic.
Expanded context sizes - Bigger models - Using the same hardware. Ollama K/V Context Quantization is opening doors for what’s possible with LLMs. Ollama’s move to integrate it is a significant step forward—and one that every AI enthusiast should keep an eye on.
Read moreHave you ever felt confident about the big picture of transformer architectures but found yourself scratching your head when it came to the nitty-gritty details? If so, you’re not alone. Whether you’re a student diving into NLP for the first time or a seasoned professional brushing up for your next interview, understanding positional embeddings is essential. This article aims to break down the intricate mechanics of transformers, starting with absolute positional embeddings—the foundation of how models like BERT understand token order. Note that I assume you are aware of what an Encoder/Decoder is and what are embeddings. If not, there is plenty of resources that will do more justice to these topics than I can. So feel free to come back to this article once you get through with that.
Read moreFor a long time, I treated A/B testing like something you “run” and then “check the dashboard.” If the p-value was small, you shipped. If it wasn’t, you moved on. That approach works ... until it doesn’t. The more I read about experimentation in real product teams, the more I realized there’s a big gap between: • knowing the basic idea of A/B tests, and • building systems and instincts that make experiments reliable at scale. So I set myself a project: build a small A/B experimentation toolkit from scratch, and then build a simulation lab to validate it. This post is a high-level story of what I built, why I built it, how I planned it, and what I learned.
Read more