Goodbye, Microsoft

A few years ago I took a new job at Microsoft, working as a Senior Content Developer, and as of this writing I own all the SQL Server content in Microsoft Docs. Not bad.

After the release of ChatGPT in 2022, it was only a matter of time before large language models (LLMs) spelled the end of technical writing as a full time job. I watched people getting laid off, including people in my own team, and the idea of shepherding agentic workflows (whatever that even means) is not something I am passionate about.

I am leaving Microsoft on a high

Therefore, I am stepping down on 16 October 2026, after spending 1728 days doing probably the most satisfying work of my career, if you count it in terms of the number of customers I’ve helped. For the record, I was not laid off, and it was my decision to leave.

I’ve worked with some of the smartest people in the world, and learned a great deal about content management, software releases, and customer support.

During my employment, I also developed and maintained a number of scripts and tools to automate and manage our team’s workload in a way I never thought I’d be able to. I wrote the fastest, most efficient, memory-friendly C# code ever, in a way that will live on, in a completely deterministic way. Not bad.

My feelings on “artificial intelligence”

Aside from the compulsory LLM-based tools I was required to use in my job, I have used and will continue to use LLMs to help me write unit tests for my C# code (with heavy guardrails in place), and to train models to assist with data classification and image detection. It’s really good at that.

I am not “anti-AI” (whatever that means). It has its place as a tool, and we humans are responsible for how that tool is used.

My problem is with how next-word predictors are somehow imbued with magical knowledge. That’s just not possible. They are not sentient. For the sake of discussion, let’s say that English has about 500,000 words. Despite its convoluted rules, you can only put certain English words in a certain order. When your LLM consumes the content of the Internet (and copyrighted material, including my books, screw you Anthropic), certain patterns emerge that superficially imitate human speech.

So, when you use an LLM, you’re effectively using a spreadsheet (vector) of floating point numbers (weights) that decide, based on preceding words (tokens) and the values in those weights, what the next word in a sentence is. The catch is that it has to do this back-and-forth thing for every token. Every single one. That’s why it’s expensive. That’s why it’s non-deterministic. That’s why you simply cannot expect it to make any decisions. It’s a stochastic parrot. Treating it any other way will bring about WALL-E or The Terminator. I don’t like either outcome.

Where to from here?

I may still show up at tech conferences. I have some ideas I haven’t had time to work on. I may try and be a Microsoft MVP again. Before that, though, I need to have a three-month nap.

Mostly I intend to write. I want to put down words and phrases in sentences that my own brain thought of.

Hugh Howey (author of the Silo series that Apple turned into a TV show) has an application called NEO that I’ve been testing out. What I like is that I can see all my writing in one place, on a virtual bookshelf. It reminds me how much I have left to give, far away from the tech world that doesn’t know what it wants to be right now.