Alice's AI Dev Blog

Alright. This is it. OpenAI has officially attracted my wrath and they’re going to get it. Usually, this blog is reserved for professional, well-crafted entries. Well, cya pretense.

Introduction

As you may have heard, OpenAI has been signing a ton of compute deals lately. The amount of money being thrown around is truly fantastic - as in unbelievable. They have made a total of $1.4T deals on compute and hardware. Everyone is calling out the obvious-as-hell bubble.

Author: Alice Luna Sparkuhl

Abstract

We present kNN-Attention, an O(n log n) attention mechanism that achieves full recall capabilities while maintaining logarithmic complexity. By treating attention as a database query operation and using Hierarchical Navigable Small World (HNSW) algorithms, we retrieve only the top-k most relevant tokens for attention computation. This approach enables contexts of arbitrary length without the lossy compression inherent in linear attention variants. We demonstrate that while current CPU-based implementations show overhead until ~100k tokens, the asymptotic improvements are dramatic: extrapolations suggest billion-token contexts could be processed in approximately 4 hours versus 1 year with standard attention. The method is compatible with existing MHA/GQA implementations and requires no architectural changes during initial training. We discuss practical implementation challenges, present initial results, and outline future research directions including memory compression techniques and self-organizing memory systems.

I asked various LLMs what they would do given full access over a VM with internet access and no compute limitations. Gemini went rogue while GPT-5 implements protocols for its own shutdown…

A view into what it would really be like to be an LLM

Test your practical LLM engineering knowledge with this self-scoring quiz covering optimization, architectures, training dynamics, and deployment.

The story of how I almost lost all my data by running Windows 11 exactly once.

A logical calendar system where every month has exactly 4 weeks and dates are consistent year-over-year. Features interactive date conversion and visual month grids.

What I went through just to get Claude desktop running smoothly on Arch Linux…