ML Before AI

You’re not a urologist just because you took a squirt without hitting your slippers in the morning.

Every dipstick uses AI now. Somewhere in the last three years, “AI expert” became a credential you could download with a subscription.

Before you fetch your pitchforks from the barn, I documented my setup and its rationale, and I built my own tooling. LLMs can be useful.

I built ML when AI mostly meant artificial general intelligence. As an engineer I trained R and Spark MLlib models on real manufacturing data from semiconductor fabs and automotive plants and deployed them for predictive maintenance and visual anomaly detection. I built a DNN on iOS that predicted handedness from sensor data with an experimental versions of both TensorFlow Lite and Core ML. And I deployed models to a hundred million customers with Kubeflow, back when Kubeflow sucked even more than it does today. As a PM, I owned an air-gapped Kubeflow stack with an SDK I designed to hide Kubernetes from data scientists, and I built weather forecasts for energy markets at a frontier lab, using hundreds of petabytes of data that required me to develop a data source valuation framework. Fifteen years of ML experience across engineering and product has become worth diddly-squat.

Nowadays, execs, PMs, and every gobshite in the universe “know AI” because they use ChatGPT, and hiring managers nod along, because apparently ChatGPT on a CV means the same as fifteen years of machine learning experience, including the unglamorous on-call years. And the Claude crowd, good lord, the Claude crowd, are somehow more insufferable, because they have decided ChatGPT is for the general public and Claude is for technical people. Piss off!

And there’s the titles… “AI PM” is absolute bollocks. Asking ChatGPT for market research you never validated, or vibe-coding a shitty demo, wow, that’s really something. You must have studied years to get good at prompting. Oh wait, you just tried it and it sort of worked. No, that does not magically make you a [technical product manager]/pm-interview-questions/). Just don’t tell the engineers they now have to own your mess.

If an (growling) “AI PM” writes evals today, most of them are pure putrid piss. They can’t find the cases that actually discriminate. They find an operational failure, codify it, and never re-test it across models or across contexts, so nobody ever learns whether it was a real failure or just a lucky prompt.

“AI-first” is also bollocks. Bollocks with knobs on, to be precise. It’s just a feckin’ tool. You’re not Rust-first or Git-first, are you? Sure, a lot of these folks are rusty gits, but that’s beside the point. “Data-first” was ridiculous too, but at least the data was worth something, sometimes even a moat. Someone else’s LLM is not a moat, because “AI-first” translates into “ Let’s run everything through an LLM!”—whether or not it makes sense. Congratulations, you’re training your future replacement, and while you’re at it, your work quality goes down the shitter faster than when it was just your human brain ruining your thinking.

Typing a question into Claude Code doesn’t make you a h4x0r because it runs in a terminal. I don’t mind that people build shitty prototypes. I mind that they think they’re doing anything AI-related. And most can’t even use Claude Code properly: no skills, no MCPs, nothing wired up. On the first day at my latest PM gig I received zero onboarding docs, and when I asked, people handed me Claude-generated 1-pagers that were actually between two and five pages. So instead I wired OpenCode to Notion, Jira, GitHub, Google Docs, and Slack with MCPs and built my own knowledge base. I curate what stays in context. But most don’t. Perhaps their company pays for Glean, and they read/write hallucinated shite without knowing it’s wrong. We are not the same.

Call it elitist, but fifteen years of my career are now worth as much as some n00b MBA who opened ChatGPT to do his friggin’ homework and write a cover letter that impressed some mountebank executive, the kind of executive who thinks the “Beautify this slide” button in Google Slides is innovation. I had one do exactly that: he complained about the look of a slide deck, then hit Beautify, and awed at the result: rounded and glossy with big colourful letters like the neon sign above a brothel next to the motorway to attract punters. It does bugger all for the content or the presentation. It’s the same school of thought as the bozos like Bezos who think botox in your lips and silicone in your tits make you beautiful. Google’s feature would be useful if it sucked the fat out of bloated slides, but it does not do that, because people prefer to add rather than subtract. Maybe one day Google will add an “Ozempic this slide” feature.

And then there’s the execs who proudly proclaim they bought a Mac Studio to run offline inference, as if owning the hardware makes you a hip AI engineer. How do you do, fellow AI engineers? Fine, thanks. Such a $10k machine, on which the best models that fit are 4-bit quantized open weights that trail the frontier by miles, is Silicon Valley’s idea of a midlife crisis. If you actually wanted offline inference, you’d buy your own damn GPU rack and deal with the whole CUDA mess that comes with it. Hey, at least they now have an expensive computer to chat with their accountant on iMessage.

“But people are more productive now!” Yeah, about that… An independent randomized controlled trial from 2025 found experienced developers took 19% longer with LLMs while believing they were 20% faster. The same team’s follow-up found no significant speed-up, mirrored by Google’s own, which found none either once it controlled for individual developer characteristics. The industry, meanwhile, advertises massive gains that even frontier labs cannot achieve. The apparent productivity boost comes from one parlour trick: dumping the hard work of verification on someone else. And that someone else uses an LLM too, which means random hallucinations checking up on others, not unlike a bunch of hippies on an LSD trip critiquing each other’s work. As long as everyone remains as high as a kite, no one will know how much dreck they produced.

Fine. You can call yourself a urologist if you want, but somebody still has to clean the bowl.