Attack Of The (Language) Model Sniffers

They came for the tokens. They stayed for the credit.

“Image generated using OpenAI’s DALL·E.”

I swear, fellow progressives, that if I get one more email from someone supposedly in “technical management” waxing lyrical about the performance of some latest frontier model or other, I am going to go absolutely, gloriously postal.

And it will be in person, in an actual office, as this one will genuinely be worth the commute¹. Personal defenestration, the proper old fashioned way, as it was meant to be, complete with a cardboard box of belongings sailing out two seconds afterwards. Especially if it has a pot plant in it.

I’m sure you’ve come across this particular situation with a specific management droid who has recently developed delusions of grandeur.

They’ve taken out a privately funded API account — company expensed, of course — and now flush away the project budget with complete gay abandon, endlessly rewording Wikipedia articles, drawing little interconnected gearwheels from prose, while the rest of us fill out an Excel spreadsheet in triplicate and have to settle for one of the extremely limited slots on the corporate subscription.

It wouldn’t be so bad if they were making the talking cat videos, or those rapping sea lions that seem to be everywhere right now², as at least that would be a genuine contribution to society.

What we are dealing with here, dear reader, is a new genus entirely.

Not the engineer, who uses the “AI” thing quietly, swears at it occasionally, and then gets on with the actual work. No, this is the model sniffer — “Dirigens vapidus”, if we’re being properly Linnaean about it — the manager who has discovered that a large language model will talk back, and has mistaken that conversation for expertise, and needs to tell everyone all about it and how they’ve judged its competence.

I have catalogued the broader malaise before, in passing. The terminal inability to ship anything at all is one symptom, the generalist’s anxious pivot to whatever happens to be trending on a Tuesday is another. But the model sniffer has earned its own field guide entry, because it has evolved a courtship display all of its own, and performs it relentlessly, in meetings, on Teams, and in emails that open with “Quick one…” and never, ever are.

Here, then, is their modus operandi.

The model sniffer judges an astoundingly complex system — months of training, a trillion tokens, an entire data centre’s worth of GPUs quietly melting away somewhere in Virginia, by asking it to draw a bicycle, or a set of interlocking gears, direct from a Wikipedia article.

It then forwards the result to the development team with a one line verdict delivered as if from Sinai.

“Impressive! But it eats so many tokens!”

The frontier of human computation, evaluated on whether it can render an SVG cog without the teeth overlapping. Never mind if it can help with anything even tangentially related to the company business, just as long as it can amuse the management droids with something they deem to be both technical and relevant.

This is, naturally, the very same instinct that once judged an entire database migration by how colourful the resulting pie chart turned out to be. Or, commonly, the quality of the PowerPoint transitions over the actual content itself.

We’ve all been here before, you and I. The grift never really changes. It’s just the party trick that does, the one that makes the developers shake their heads in disbelief before rapidly moving toward the delete button and trying to eradicate the experience from their memory before it takes root and displaces something of actual importance.

Yesterday it was the blockchain, then the metaverse, then an NFT of a conspicuously sad monkey. Today it’s a hexagon with a little ball bouncing about inside it, a toothed gearwheel rotating, or something they don’t quite understand but that looks really technical and has a necessarily long explanation with lots of technical looking words, model names, and (most importantly) recent version numbers from the latest AI press release they saw on X.

Tomorrow, Cthulhu only knows, but you can be quite sure there will be a Teams, Outlook, and (more depressingly) LinkedIn post about it by lunchtime.

The shame of it all when it gets outside the company, that it might be read by someone actually technical, even if on LinkedIn that’s a very remote possibility, it still makes every job seeking developer cringe inside.

You can identify a model sniffer in the wild by a handful of reliable tells, should you find yourself cornered by one near the company kitchen, or in an “all hands”, all cameras on, Teams meeting.

There is the Token Accountant, who has no earthly idea what a token actually is, but knows with absolute certainty that there are far too many of them, and says so, loudly, at every available opportunity.

There is the Forwarder, whose entire technical contribution to the quarter is an email opening “Have you seen this?”, a link to the aforementioned rapping sea lions, and a single thinking face emoji deployed in lieu of any analysis whatsoever.

There is the Benchmarker, who has run precisely one prompt in their life — the gears, always the gears — and now refers gravely to “the model’s limitations” with the bearing of someone who has read the paper. They have not read the paper, of course. They don’t read papers, they just forward them with the aforementioned “Have you seen this?” and “Important stuff in here!!!” commentary.

And there is the Visionary, who, having watched the thing autocomplete a single function, has concluded that the entire engineering department is henceforth “a legacy cost centre.”

Bless them all, the poor luvs, for they know not what they do.

It is, frankly, exhausting. These are, by and large, the very same people who a year ago couldn’t state the nature of the software emergency if their bonus depended on it, and who now position themselves as some in house oracle on the entire frontier model space.

Amusingly, of all the management grades, it’s usually the CTOs themselves that start the ball rolling. That’s the one, if you hadn’t already experienced it, where we have to push back up the hill and out of the way, over and over again³.

So here we all are. Again. The one technology that might, on some distant and admittedly hopeful afternoon, actually automate the management droid out of existence, is instead being wielded by the management droid to draw cogs and feel briefly clever about it.

There is a certain Schrödinger grade irony in that, and I would laugh heartily if I weren’t quite so tired, depressed, and bitterly hostile to the whole thing altogether.

Judge the model by its output on real work, or extend us all the mercy of not judging it at all. But please, I beg you, for the love of Great Cthulhu below, stop drawing those bloody gears and sending on the email commentary that includes your witless pontifications.

Be wary, fellow software types. The model sniffer is not going extinct, it is breeding fast, it is being promoted, and it is now most likely writing the company’s “AI strategy”.

They have the tokens, we have the talent, and they will, as always happens, take the credit too.

They must be stopped.

[1]: Even with the astronomical price of petrol right now. [2]: A bit like the Badger Badger Badger one in the old times. [3]: Apparently the “T” in “CTO” stands for “Technical”. Hard to believe, really.