I explore, I work, and I have hobbies.
Jul 13, 2025

LLMs Ace SQL… Just Ignore the Footnotes

Few days ago, I came across BIRD benchmark that evaluates LLM performance on natural language to SQL. The model namely XiYan-SQL, at the 5th position is open source with all of paper, weights and codes available. The 32B parameter model XiYan-SQL performs better than GPT-4o

Jul 12, 2025

The AI War: The Ultimate SVG Generation Battle

AI World welcomes another

brand new open source Model: Kimi K2, and in non-thinking mode, it not only rivals (Claude Opus 4) but surpasses (GPT-4.1, Claude Sonnet 4) the performance of leading models in SWE-bench.

I'm putting four different…

Jul 7, 2025

Calculus for Words: Optimizing Text with the Power of Gradient Descent

In the world of machine learning, gradient descent is the engine of progress. It's the mathematical process that allows models to "learn" by incrementally minimizing their errors. But what if we could apply this powerful concept not just to numbers,…