Intro

0001-01-01

This is a tvtropes-like set of short pages meant to explain how agentic AI goes wrong, and why it will always go wrong.

It’s written from the perspective of someone who has used this stuff since before the transformer, who ran llama.tar when it leaked, and who has used every model in every harness in every problem. I use millions of tokens every day. Truly. I have, and continue to, do it all.

I’m not a pessimist, LLMs are incredibly powerful tools that have changed how software is written (for the better, mostly). But those with less experience often over-estimate it, and people with organizational decision-making power often fail to grasp what exactly this tool is, or what its limits are. Discussions about this often lead to them presuming that you just don’t "get it", and recommending elementary things like "have you tried plan-then-build?"

This series is intended to demonstrate a set of principles that aren’t negotiable. This is how it is. Our opinions need to be shaped by the reality around us, or we proceed into delusion at our peril.

All the pages in this series link to each other in some way, but here’s an overview.

• [[Models Have a Limit]]

• [[Architecture Beats Models]]

• [[Documentation is Always Wrong]]

• [[Userland is Not Enough]]

• [[Software Managers have Dunning-Kruegar]]

• [[Alignment is hard actually]]

• [[Deinvestment Causes Layoffs not AI]]

• [[Maximalism of the Gaps]]

All site content protected by CC-BY-4.0 license