Jae H. Park·5d agoWhy LLMs Need KV Cache and Chunked PrefillThe Matrix Math Behind these Optimizations
Jae H. Park·Aug 30Heading Straight for the Peak: Why the Gradient Vector Points to Steepest AscentImagine you are climbing a mountain and want to find the steepest (quickest) path to the peak. Which direction should you step? This is the…A response icon3A response icon3
Jae H. Park·Mar 17Importance Sampling in Disguise: Inside the Objectives of VAEs and Diffusion ModelsVariational Autoencoders (VAEs) and Denoising Diffusion Probabilistic Models (DDPMs) are often presented as different generative models…
Jae H. Park·Jul 16, 2025From Noise to Clarity: The Math Behind Denoising Diffusion ModelsWhen I first encountered the Denoising Diffusion Probabilistic Model (DDPM), I struggled to understand how simply adding and then removing…A response icon2A response icon2
Jae H. Park·Mar 27, 2025Finally! A Clear Derivation of the VAE KL LossWhile analyzing the loss function of the variational autoencoder (VAE) model, I noticed that the Kullback-Leibler (KL) loss term has a nice…
Jae H. Park·May 21, 2024Forget Keywords, Think Meaning: The New Era of SearchingIn 2020, during a seminar discussion about using AI to improve corporate search, I expressed skepticism about embedding-based search with…
Jae H. Park·Aug 19, 2023How can you measure “Uncertainty” ?How can you measure “uncertainty”? This was a question I have posed to my kids when they were learning probability. Understanding…A response icon1A response icon1