AI Arena
Practical guides on building with AI — agents, retrieval-augmented generation, LLMs, and the architectures that tie them together.
-
Machine Learning Basics
A hands-on series covering the core building blocks of every ML model — linear regression, weights and biases, loss functions, gradient descent, batches and epochs, overfitting, many features, activation functions, neurons, and classification, ending at next-token prediction. 14 parts, each with an interactive demo you can run in your browser, plus a glossary.
-
LLM Basics
A plain-language series on how large language models actually work — from the simplest feedforward network, through networks that read and remember, to attention, the transformer, and GPT vs BERT. 8 parts written for nontechnical readers, each with an interactive demo.
-
How LLMs Talk
A plain-language tour of everything between the model file and your chat window — inference engines, model servers, chat templates, thinking mode, tool calling, and agents. 8 parts following a single message end to end, written for nontechnical readers.
-
Inside the GPU
A beginner's tour of GPU hardware from the vacuum tube and the first transistor up to Nvidia's Ada, Hopper, Blackwell, and Vera Rubin, AMD's Radeon and Instinct chips, and Intel's Arc. Logic gates, the memory wall, Nvidia's streaming multiprocessors, AMD's compute units and Intel's Xe-cores, tensor cores, CUDA, NVLink, how a chip is made, what a frame is, where the watts go, what an NPU is, and how to read a spec sheet. 34 parts in the spirit of Inside the Machine, most with an interactive demo.
-
Building AI Agents
A hands-on series covering the full stack of AI agent development — LLM clients and chains, chunking, embeddings, query transformation, RAG strategies, tool calling, agent architectures, and guardrails. 9 parts, implemented with LangChain and LangGraph.
-
Speculative Sampling: How a Small Model Makes a Big One Faster
A language model writes one token at a time, and most of each step is spent reading the same weights again. How a small draft model lets the big one check six tokens in a single pass, what the check looks like token by token, what happens on the round where every guess is wrong, and who runs this in production, from Google Search and OpenAI to Apple, Groq, DeepSeek and the desktop tools.
-
SynthID, dissected: how AI watermarks work, who actually uses them, and what they can't tell you
October 2026 was the month AI watermarking stopped being a research topic. I pulled SynthID apart to see what the signal actually is, how fragile it is, who holds the keys, and why open weights are the hole in the middle of it.