Running AI in a government cloud: FedRAMP, GovCloud regions, air-gapped models, audit logs and agents, as of September 2026. Same plumbing as anywhere else, with the boundaries drawn in ink.
· ml, government, security, compliance, explainer
Most questions don't need your most expensive model. Routing, cascades, distillation and caching: the four ways to send each request to the cheapest model that can handle it, and how to know it's still good enough.
· ml, inference, efficiency, cost, explainer
How AI models shrank from racks of GPUs toward laptops and phones, one fewer bit at a time. A kept-current timeline of quantization, from 2015's phone models to today's 1-bit LLMs.
· ml, quantization, efficiency, explainer
Jev is a decision model: you ask a question and list the allowed answers, and it returns calibrated odds instead of text. What that buys you, where it breaks, and why the pattern will outlast any one product.
· ml, decision-models, efficiency, explainer
API or your own GPUs? The cost crossover, speed, privacy, quality and who gets paged at 3 a.m., with September 2026 prices and a break-even calculator.
· ml, inference, cost, explainer
Prompt injection is SQL injection's younger sibling: instructions hidden in data. Why language models can't tell the two apart, why agents raise the stakes, and the defences that actually help.
· ml, security, agents, explainer
Months of compute and a fortune in chips end up as one set of files. Why AI labs treat model weights as crown jewels, how they get stolen, the controls that stop it, and the case for giving them away.
· ml, security, explainer
Training a frontier model is a distributed-systems problem with a very expensive failure mode. How the work is split across thousands of GPUs, how they talk, what breaks, and how labs measure the waste.
· ml, training, distributed-systems, explainer
From keypress to streaming words: tokens, attention, next-token prediction, temperature, and the two phases of serving (prefill and decode) that decide how fast an answer arrives and what it costs.
· ml, llm, inference, explainer
GPUs for people who run servers: why AI runs on thousands of simple cores, why memory bandwidth matters more than raw speed, and why batching is one of the biggest tricks in serving.
· ml, hardware, inference, explainer
Retrieval-augmented generation looks easy in a demo and breaks in production. Chunking, embeddings, keyword vs. vector vs. hybrid search, reranking, and why a wrong answer is often a search bug.
· ml, rag, search, explainer
A freshly trained language model doesn't answer questions; it continues them. Post-training explained: example answers, human preferences, RLHF and DPO, rewards a program can check, written principles, and the ways it all goes wrong.
· ml, llm, post-training, explainer
Evals, explained for people who ship software: test sets and the leaks that ruin them, why grading language is hard, LLM judges and their biases, and turning all of it into a gate your deploys have to pass.
· ml, evals, mlops, explainer
No math required: loss, gradient descent, learning rates and overfitting, with a line you can train by hand and a curve you can watch memorize the practice test.
· ml, training, fundamentals, explainer
Machine learning infrastructure is what's built around these AI models to turn a chatbot into a real business: hooking up tools, and housing it on architecture to serve millions.
· ml, infrastructure, mlops, explainer
If you're alone at a keyboard with no one to share your concepts with, you've got an infrastructure problem. Here's how to fix it with five people, a Discord server, and the discipline to keep showing up.
· webdev, discord, productivity, writing
