Sometimes I like to put my thoughts into words. Here's a collection of my thoughts in some of the topics I care about.
In autoregressive language models, not all tokens are created equal. The first tokens in a context window shape everything that follows — from attention patterns to generation trajectory. This is why short sessions beat long ones, and why compression exists.
Peter Gostev's BullshitBench exposes a disturbing truth about LLMs: most models will confidently answer nonsensical questions instead of pushing back. Even worse, reasoning models that "think harder" tend to generate more elaborate bullshit, not less. Here's what the data reveals.
Emerging research shows that AI tools are not reducing workloads — they are intensifying them. A deep dive into how artificial intelligence amplifies workaholism, erodes boundaries, and reshapes the meaning of "enough."
A University of Michigan foundation model called Prima can read a full brain MRI and diagnose across 52 neurologic conditions in seconds — reaching up to 97.5% accuracy. Here is why this matters for patients, radiologists, and the future of medical imaging.
Context windows define the boundaries of what an LLM can see, think about, and respond to. In this article we break down tokens, tokenization, and the invisible walls that shape every AI conversation.
An in-depth analysis of Claude Opus 4.6, covering its new features like adaptive thinking, 1M-token context, and agent teams, along with detailed benchmark comparisons against OpenAI GPT-5.2, Google Gemini 3 Pro, Meta Llama 4, and DeepSeek R1.
From 5,000 to 145,000 GitHub stars in two weeks, through a trademark dispute, a crypto scam, and critical security vulnerabilities—the wild story of Clawdbot (now OpenClaw), the self-hosted AI assistant that redefined what personal AI agents can do, and why security experts are warning everyone not to run it.
In a world obsessed with sleek GUIs and Electron apps, developers are quietly returning to text-based interfaces. From lazygit to btop, TUIs are proving that sometimes the best interface is the one that gets out of your way.
How Model Context Protocol servers can accidentally intoxicate AI agents with too much information, and why code execution might be the cure for this self-inflicted cognitive bottleneck.
A comprehensive technical comparison of single-agent and multi-agent AI systems, exploring architecture patterns, performance metrics, cost analysis, and real-world case studies to help you choose the right approach for your AI implementation.
Discover how low-code and no-code ML tools are democratizing machine learning. From Orange3 to PyCaret, explore the best platforms that let you drag, drop, and click your way from raw data to AI predictions—no PhD required.
Everyone hates Comic Sans—until you learn why it might be the most accessible font choice for dyslexic programmers. From undiagnosed developers to Comic Code and Comic Mono, we explore the intersection of typography and neurodiversity in software development.
Vector databases are specialized systems designed to store and search high‑dimensional numerical vectors called embeddings. In this article we'll analyze their structure, philosophy and challenges with this new protocols
Model Context Protocol (MCP) is utilised to provide tools to a LLM model in order to interact with an application. In this article we'll analyze it's structure, philosophy and challenges with this new protocols
Deep Learning has been the choosen tool to analyze complex data in many fields. Heart anomalies is no exception. There is a plethora of articles working on deep learning models capable of identifying anomalies in heart-rate data. This article present a different approach: instead of working with the raw data, we will process the signal to extract a simplified one that will contain enough information to train a simple model to identify anomalies. This approach will allow the simple signal to be recreated using other sensors and monitoring tools that won provide ECG data.