Long-form writing

Blog

Longer essays and posts from content/posts, separated from the working digital garden.

Loading blog filters...

Static HTML

All blog posts

Every long-form post as a regular HTML link.

2025-11-232025-11-22 Why I Focus on Java, Rust in the AI EraRight now AI writes more and more code. Because of this, using statically typed languages is even more sane. Languages like Java and Rust give clear rules and strong checks. I also prefer frameworks that work on a higher
2025-11-22Too Many Things, Not Enough TimeWhen I was kid, time moved slowly. I remember getting my C64, and the small manual that showed how to draw a moving ball in BASIC. I read that book many times. Until I was about 9/10, it was the only IT related book I ow
2025-07-19Everyone will say that 2025 is the year of AI agents. But I think it's far from readyEveryone will say that 2025 is the year of AI agents. However, even with 99% accuracy at each step, longer workflows will often fail. Imagine a 50 step flow, it would have only about 60% overall accuracy and in reality,
2025-02-03Zastosowanie Deep Learning w Identyfikacji Narzędzi na Podstawie Śladów NarzędziowychZastosowanie Deep Learning w Wspomaganiu Porównywania Śladów Narzędziowych Rozwój metod deep learning i uczenia metryk podobieństwa umożliwia budowę systemów wspierających ekspertów kryminalistyki w analizie śladów narzę
2025-02-02The Principle of Least Action - Unifying the Laws of PhysicsThe Principle of Least Action: Unifying the Laws of Physics Rather than describing nature with a multitude of separate laws, almost every phenomenon from the path of light to the motion of planets can be understood as na
2025-01-30Reinforcement Learning with GRPO Fine-Tuning a Small Language Model for Chain-of-Thought Math Reasoning. Similar to Deepseek R1 trainingI found this lovely gist the is similar to DeepSeek R1 training. This demonstrates how to train a small language model Qwen or Llama on the GSM8K math dataset using the new GRPO algorithm from the Hugging Face TRL librar
2025-01-21Deepseek R1 noteThe first generation reasoning models, DeepSeek R1 Zero and DeepSeek R1 , bring advanced reasoning capabilities to AI. DeepSeek R1 Zero: Reinforcement Learning Without Fine Tuning DeepSeek R1 Zero is trained using large
2025-01-13Rejection Sampling NoteRejection Sampling rejection sampling is a technique from probability and statistics used to generate samples from a target distribution $p x $ when direct sampling is difficult. Instead, you sample from a simpler "propo
2025-01-07CES 2025 notesCES 2025 seems to have a great start in LAS VEGAS. Very strong announcements from NVIDIA. Some of these really mind blowing how quickly technology is moving on NVIDIA GeForce RTX 50 Series GPUs NVIDIA introduced the GeFo
2025-01-07NVIDIA Cosmos - CES 2025 AnnouncementNVIDIA Cosmos: CES 2025 Announcement At CES 2025, NVIDIA introduced Cosmos , a computing platform that accelerates the development of autonomous vehicles and robots. Cosmos uses synthetic data to train models, reducing t
2024-12-30Inside InnoDB: A Straightforward Look at Undo Logging and HistoryImagine you have a little library where people come in and update books all day long. Some mark pages with notes, some correct typos, and some even rip out pages. Yet, no matter what, each visitor wants to see a neat, co
2024-12-24Note on paper Training Large Language Models to Reason in a Continuous Latent SpacePaper Title: Training Large Language Models to Reason in a Continuous Latent Space Link: 1 Motivation and Main Idea Large Language Models LLMs usually do what’s called “chain of thought” CoT reasoning, where they generat
2024-12-23A Course of Pure Mathematics – G. H. Hardy (1921) [pdf] - notepdf.pdf pdf.pdf The way calculus was taught changed after Hardy's A Course of Pure Mathematics came out. People in general like how direct this book it is. It is easy to understand things like what limsup and liminf mean
2024-12-21What Gödel and Turing Do—and Do Not—Tell Us About Truth, Minds, and MachinesFour Different Questions Gödel's and Turing's results are often invoked in discussions that move too quickly between very different questions. Keeping those questions separate is the beginning of a useful conclusion. For
2024-03-15Understanding the Attention Mechanism in Transformer ModelsTransformers rely on a concept known as attention that helps in language processing and understanding. I try to explore the complexity of attention within transformers, analyzing its role in capturing cues and achieving
2023-12-14The Art of Algorithm Tuning: Why Optimization Still MattersWhen I first got Programming Pearls by Jon Bentley, I was maybe 10/11 years old. I remember reading it many times. It showed how to improve algorithms, how to write code that is simple and fast, and it told many stories
2022-12-09How to Build Simple Image Classifier (torch, torchvision, CIFAR)Image Classifier with CIFAR 10 <img alt="A photo of a man on an elephant with an ML generated overlay showing objects in the frame" src=" %28Johnson et al.%2C 2016%29 %28cropped%29.png" width=450px An image classifier is
2022-12-07Tackling the Exploding Gradient Problem in Neural NetworksTackling the Exploding Gradient Problem in Neural Networks When you're training deep neural networks, you might run into what's called the "exploding gradient problem." This is a tricky issue, but don't worry, it's manag