The Fourth Rule of The Foundry
My AI agent team's investigator was skimming, and I only caught it by testing it, so I added a tenth agent whose whole job is reading. Part two of two.
# blog
Blog Posts
Writing on networking, security, IT leadership, and the business of technology.
browse: /categories /series /tags /archive
Reading the Whole Cookbook to Write One Word: How Inference Works
Training is one enormous bill, paid once. Inference is the meter that never stops running, and an old network saying about bandwidth and latency explains why a chatbot's words arrive one at a time. Part four of six.
[ Ai and llms ] Part 2 of 6 3 min read
Prompting ClaudeIt Copies Everything
I'd like to claim I've never taught a model a bad habit by accident. A model learns from every part of an example, including the parts you didn't mean to teach, so here's how to pick them, on one chart. Part two of six.
[ Tech field day ] [ Security ] 5 min read
Fear and Logging in Las Vegas
I'm headed to Tech Field Day Extra at AWS re:Invent to hear what AWS has to say about security, how it compares with Azure, and where the rest of the industry is going.
I Understand Most of the Words I'm Saying: An Introduction to How LLMs Work
An introduction to a six-part series on how large language models work, from a network engineer who understands most of it and a little of the math.
[ Ai and llms ] Introduction 3 min read
Prompting ClaudeMost of What You Know About Prompting Is From 2023
Prompting advice ages faster than the models it was written for, and a lot of what's still online, including my own notes, now backfires. An introduction to a six-part series where the infographics do the heavy lifting.
[ Ai and llms ] Part 1 of 6 2 min read
Prompting ClaudeStop Yelling at the Robot
I used to write prompts the way I wrote change-window emails to the NOC: in capital letters, with the word CRITICAL in them. Eight prompt habits that helped on 2023-era models and backfire on the current ones, on one chart. Part one of six.
Two Strong Oxen or 1,024 Chickens: The Hardware and Network Behind LLM Training
A frontier training run is a networking problem dressed up in a math costume. The industry chose the chickens over the oxen, so how do you get 16,000 of them pulling in the same direction? Part three of six.
[ Ai and llms ] Part 2 of 6 9 min read
How LLMs Actually WorkFail Better, Fifteen Trillion Times: How an LLM Is Trained
Training turns roughly 405 billion random numbers into something useful through one boring move: guess the next token, get graded, nudge every number a hair, and repeat. Part two of six.
[ Ai and llms ] Part 1 of 6 11 min read
How LLMs Actually WorkJust a Lot of Multiplication: What's Actually Inside an LLM
A large language model is an enormous pile of numbers, and everything it does comes from multiplying other numbers by that pile. Part one of six.