Google's TimesFM vs the Stock Market: Every Way I Fooled Myself
I tried to beat the market with a time-series foundation model. It didn't work — and the reasons why turned out to be worth more than the strategy would have been.
33 minute read
Categories
4 pages
Machine Learning
Exploring AGI and the integration of System 1 and System 2: JEPA, Transformers, and Dynamic Memory
An in-depth exploration inspired by Yann LeCun and Bill Dally's GTC 2025 discussion, detailing my thinking process on integrating JEPA with transformer models and developing dynamic learning and memory management in AGI.
4 minute read
Ring Attention Explained: How Modern LLMs Remember Long Contexts Without Losing Their Minds
A deep dive into Ring Attention—how models like Gemini and Claude handle long contexts using clever memory tricks like sliding windows, compressed memory, and selective token referencing.
4 minute read
Context Keeper: How Storing Long but Indexing Short Affects Retrival Accuracy
How breaking text into chunks can improve similarity detection for question pairs, outperforming traditional full-text embedding methods
3 minute read