Bosses *do not know* that code is a liability, not an asset. That's why they won't shut the fuck up about the chatbots that shit out 10,000 times more code than any human programmer. They think they've found a machine that produces *assets* at 10,000 times the rate of a human programmer. They haven't. They've found a machine that produces *liability* at 10,000 times the rate of any human programmer.

27/

Maintainability isn't just a matter of hard-won experience teaching you where the pitfalls are. It also requires the cultivation of "Fingerspitzengefühl" - the "fingertip feeling" that lets you make reasonable guesses about where never before seen pitfalls might emerge. It's a form of process knowledge. It is ineluctable. It is not latent in even the largest corpus of code that you could use as training data:

pluralistic.net/2025/09/08/pro…

28/

*Boy* do tech bosses not get this. Take Microsoft. Their big bet right now is on "agentic AI." They think that if they install spyware on your computer that captures every keystroke, every communication, every screen you see and sends it to Microsoft's cloud and give a menagerie of chatbots access to it, that you'll be able to tell your computer, "Book me a train to Cardiff and find that hotel Cory mentioned last year and book me a room there" and it will do it.

29/

This is an incredibly unworkable idea. No chatbot is remotely capable of doing all these things, something that Microsoft freely stipulates. Rather than doing this with one chatbot, Microsoft proposes to break this down among dozens of chatbots, each of which Microsoft hopes to bring up to 95% reliability.

That's an utterly implausible chatbot standard in and of itself, but consider this: probabilities are *multiplicative*.

30/

A system containing two processes that operate at 95% reliability has a net reliability of 90.25% (0.95 * 0.95). Break a task down among a couple dozen 95% accurate bots and the chance that this task will be accomplished correctly rounds to *zero*.

Worse, Microsoft is on record as saying that they will grant the Trump administration secret access to all the data in its cloud:

forbes.com/sites/emmawoollacot…

31/

So - as Signal's Meredith Whittaker and Udbhav Tiwari put it in their incredible 39C3 talk last week in Hamburg - Microsoft is about to abolish the very *idea* of privacy for *any* data on personal and corporate computers, in order to ship AI agents that cannot *ever* work:

youtube.com/watch?v=0ANECpNdt-…

32/

39C3 - AI Agent, AI Spy

Agentic AI is the catch-all term for AI-enabled systems that propose to complete more or less complex tasks on their own, without stopping to ask permission ...

^{media.ccc.de (YouTube)}

Meanwhile, a Microsoft exec got into trouble last December when he posted to Linkedin announcing his intention to have AI rewrite *all* of Microsoft's code. Refactoring Microsoft's codebase makes lots of sense. Microsoft - like British Airways and other legacy firms - has lots of very old code that represents unsustainable tech debt. But using AI to rewrite that code is a way to *start* with tech debt that will only accumulate as time goes by:

windowslatest.com/2025/12/24/m…

33/

Microsoft denies rewriting Windows 11 using AI after an employee's "one engineer, one month, one million code" post on LinkedIn causes outrage

Microsoft told Windows Latest that the company does not plan to rewrite Windows 11 using AI in Rust after an employee's post causes outrage.

^{Mayank Parmar (Windows Latest)}

Now, some of you reading this have heard software engineers extolling the incredible value of using a chatbot to write code for them. Some of you *are* software engineers who have found chatbots incredibly useful in writing code for you. This is a common AI paradox: why do some people who use AI find it really helpful, while others loathe it? Is it that the people who don't like AI are "bad at AI?" Is it that the AI fans are lazy and don't care about the quality of their work?

34/

There's doubtless some of both going on, but even if you teach everyone to be an AI expert, and cull everyone who doesn't take pride in their work out of the sample, the paradox will still remain. The true solution to the AI paradox lies in automation theory, and the concept of "centaurs" and "reverse centaurs":

pluralistic.net/2025/09/11/vul…

35/

In automation theory, a "centaur" is a person, assisted by a machine. A "reverse centaur" is a person conscripted to *assist a machine*. If you're a software engineer who uses AI to write routine code that you have the time and experience to validate, deploying your Fingerspitzengefühl and process knowledge to ensure that it's fit for purpose, it's easy to see why you might find using AI (when you choose to, in ways you choose to, at a pace you choose to go at) to be useful.

36/

But if you're a software engineer who's been ordered to produce code at 10x, or 100x, or 10,000x your previous rate, and the only way to do that is via AI, and there is no human way that you could possibly review that code and ensure that it will not break on first contact with the world, you'll hate it (you'll hate it even more if you've been turned into the AI's accountability sink, personally on the hook for the AI's mistakes):

pluralistic.net/2025/05/27/ran…

37/

There's another way in which software engineers find AI-generated code to be incredibly helpful: when that code is *isolated*. If you're doing a single project - say, converting one batch of files to another format, just once - you don't have to worry about downstream, upstream or adjacent processes. There aren't any. You're writing code to do something once, without interacting with any other systems.

38/

A *lot* of coding is this kind of utility project. It's tedious, thankless, and ripe for automation. Lots of personal projects fall into this bucket, and of course, by definition, a personal project is a centaur project. No one forces you to use AI in a personal project - it's always your choice how and when you make personal use of any tool.

39/

But the fact that software engineers can sometimes make their work better with AI doesn't invalidate the fact that code is a liability, not an asset, and that AI code represents liability production at scale.

In the story of technological unemployment, there's the idea that new technology creates new jobs even as it makes old ones obsolete: for every blacksmith put out of work by the automobile, there's a job waiting as a mechanic.

40/

In the years since the AI bubble began inflating, we've heard lots of versions of this: AI would create jobs for "prompt engineers" - or even create jobs that we can't imagine, because they won't exist until AI has changed the world beyond recognition.

I wouldn't bank on getting work in a fanciful trade that literally can't be imagined because our consciousnesses haven't so altered by AI that they've acquired the capacity to conceptualize of these new modes of work.

41/

But if you *are* looking for a job that AI will definitely create, by the millions, I have a suggestion: digital asbestos removal.

For if AI code - written at 10,000 times the speed of any human coder, designed to work well, but not to fail gracefully - is the digital asbestos we're filling our walls with, then our descendants will spend generations digging that asbestos out of the walls.

42/