Latest articles
Free Web Search for AI Agents: What Works, What Doesn't, and How to Build It
Every AI agent needs web search. Every search API costs money or gets blocked. I tested every free option from DuckDuckGo to SearXNG, found the one that actually works from production servers, and built a capsule agent with Cheerio and the AI SDK that can search, scrape, and reason over live web data.
CLaRa: Apple's Revolutionary Approach to Making AI Remember Better While Using Less Memory
Apple just released CLaRa, a framework that solves one of AI biggest problems: how to give language models access to huge amounts of information without overwhelming them. Instead of feeding massive documents directly to the AI, CLaRa compresses information by 16-128 times while keeping what matters. This post breaks down the math behind how it works, why it is different from everything else, and what makes it so powerful.
Spectral-Entropic Bottleneck Theory: A Mathematical Framework for the Reasoning Horizon in Large Language Models
Current LLMs fail at deep compositional reasoning and hallucinate inevitably. Existing theories explain these as separate problems. This paper introduces the Spectral-Entropic Bottleneck Theory (SEBT), a unified mathematical framework showing that both failures share a common root in the attention mechanism: the spectral entropy of attention matrices decays monotonically across layers, creating a reasoning horizon. We derive constructive bounds and propose an architectural solution that provably eliminates the attention-induced bottleneck.
Bayesian Relevance Estimation: Deriving the Optimal Retrieval Score from First Principles
Everyone uses cosine similarity for retrieval. But if you derive the mathematically optimal scoring function from Bayes theorem, you get something different: the dot product minus a magnitude penalty. This post walks through the full derivation, proves when and why it outperforms cosine similarity, and extends it with structural priors and adaptive retrieval stopping.
View Transitions in Next.js 16: Native Animations Without a Single Library
The View Transitions API lets you animate between UI states and page navigations using nothing but the browser. Next.js 16 and React 19.2 ship first-class support for it. Here is everything you need to know.
Building an AI Agent From Scratch With the AI SDK
The AI SDK is now at version 6 with agents, tool approval, MCP support, and structured outputs. This guide walks you through every concept and ends with a full agent that can search the web, generate structured data, and stream to a Next.js UI.
OpenClaw: The Open-Source AI Agent That Hit 175K GitHub Stars in Months
OpenClaw (formerly Moltbot, formerly Clawdbot) is a self-hosted AI agent that runs 24/7 on your own hardware. It connects to WhatsApp, Telegram, Slack, and 60+ platforms. One developer built it in 10 days. Now it has 175,000 GitHub stars.
Claude Code Changed How I Work and I Am Not Going Back
Claude Code is not an autocomplete tool. It is a terminal-native AI agent that reads your codebase, runs commands, writes code, creates commits, and iterates on errors. Here is how it works, how it compares to Cursor and Copilot, and why it feels different.
RAG Is Not the Best Knowledge Solution and Here Is the Math to Prove It
Everyone defaults to RAG for giving LLMs external knowledge. But the retrieval step introduces failure modes that most teams never measure. This post breaks down the math behind RAG accuracy, compares it to long-context, CAG, PageIndex, knowledge graphs, and fine-tuning, and makes a case for what actually works best.
AI Is Eating the Job Market and Nobody Is Ready
Over 50,000 jobs were cut in 2025 with AI cited as the reason. 57% of US work hours are already automatable. But the same technology pays developers 45% more if they know how to use it. Here is what the data says, what it means, and what I think about all of it.