Every developer with a ChatGPT Plus subscription seems to be calling themselves an AI expert right now, while the actual ML Engineer works through something much messier. Roughly $200 billion of investment is riding on the assumption that LLMs will produce profits the math does not really support. I have been around long enough to see the dot-com era in those numbers.
Stephanie Kirmer, a staff ML Engineer with nearly a decade in the field, was blunt about how the job is changing when we spoke recently. Better prompts are the small part of it. The bigger part is the sociological impact of the code we ship, and what to do with the “scut work” that used to eat whole afternoons.
The sociology of the ML engineer
Kirmer’s background in sociology shapes how she engineers. Instead of “move fast and break things,” she asks: “What are the social inequalities at play here?” and “How do different groups experience this tool differently?” That is not academic fluff. I run into the same thing in the WordPress ecosystem, where a badly implemented AI search feature quietly alienates anyone who does not fit the “standard” data model.
For how these data models land in our niche, I wrote up 3 machine learning lessons for WordPress development. It is a reality check if you think you can hook an LLM into a site and call it done.
Automating the scut work
Unit tests and boilerplate eat time. Kirmer says the daily work of an ML Engineer now leans on code assistants for that “scut work,” which leaves more room for the problems specific to your own system. Handled carelessly, it also produces a new flavor of technical debt in AI development.
If an LLM is generating logic for a custom WordPress plugin, working code is not the bar. It has to be secure and able to scale. Here is a helper I use to check a generated snippet against our internal standards before it reaches a staging environment.
<?php
/**
* Naive evaluation of LLM-generated code blocks for specific banned patterns.
* Prefix: bbioon_
*/
function bbioon_validate_llm_snippet( $generated_code ) {
$banned_patterns = [
'eval(',
'base64_decode(',
'$_POST['
];
foreach ( $banned_patterns as $pattern ) {
if ( strpos( $generated_code, $pattern ) !== false ) {
return false; // Found a security bottleneck
}
}
return true; // Potentially safe to refactor
}
The $200 billion bubble
Kirmer’s argument is that the bubble comes from investment being out of proportion to what the tech currently does, not from the tech being useless. So companies keep making outlandish promises to keep the VC wheel turning. I have watched this happen before. When it pops, the engineers still standing tend to be the ones who were solving actual problems.
Winning back a skeptical public means demystifying the thing. LLMs are advanced statistical models, and treating them as magic helps nobody. The full interview with Stephanie Kirmer is up at Towards Data Science if you want more on the social side.
If this ML engineer work is eating your dev hours, I can take it on. I have been working with WordPress and custom integrations since the 4.x days.
Where that leaves the job
The ML Engineer job is drifting away from pure data science and toward software engineering with a sociology and ethics component. The $200 billion of noise is a distraction. Build systems that are reproducible and that people can trust, whether that means Python or custom blocks against the WordPress Developer Resources. Solve the problem instead of shipping the hype.
