Yurii Shevchuk's blog

The article addresses the reliability challenges of Large Language Models (LLMs), particularly in structured text extraction tasks, and presents two straightforward methods aimed at enhancing both the reliability and efficiency of these processes.

The article shows relations between the sigmoid, softmax, and tanh functions. Specifically, it shows how one of these functions can be used to represent another straightforwardly.

The article covers a rather simple proof of the peculiar property of the generalised binomial distribution

In the article, I'll show you that gradient boosting can be seen as a "blind" gradient descent. The gradient boosting pursues the same goal as the gradient descent, except it doesn't have a function for which we want to find a minimum, and yet the steps it takes can get you closer to it. That's what I mean by "blind". Gradient boosting optimizes the function which it doesn't see.

In this article, I will show that many of the properties of Knuth's numbers can be understood by imagining a rather impractical memory device that can be directly connected to Knuth's numbers. In addition, the recursive construction process of the numbers can tell us how to create such a device.

The article aims to correct the claim about the martingale betting strategy in the Numberphile video and show that the actual probability is equal to 50% rather than 1/e, as stated in the video.

Three solutions to one of the problems from the Putnam's exam.

Solution to the FiveThrityEight's puzzle which includes general intuition and prove for game with an arbitrary number of milk and black chocolates.

Range is quite an unusual data structure for SQL queries and chaining these ranges is not straightforward. This article shows how to implement a solution to a very specific problem involving ranges. In addition to the solution there is a proof which shows that the proposed algorithm works for every possible set of ranges.