Selling inefficiencies as features

Language models come in a singular form. Everyone foots the bill until somebody divides the labor.

The right split is the one every trade already made a century ago.

One kind is a model for automation, the machine you point at repetitive work. Directed syntax changes, format conversions, the same bounded job run ten thousand times; work that needs no guidance, just instructions. Autonomy is the entire value, and a question from this machine is a defect.

The other kind is a model for delegation, the worker you hand a one-off. A delegated job runs once and produces a singular result, a specific thing the boss already sees in his head, and most of what defines it never makes it into the handoff. This worker needs guidance and instruction the way any new hire does, and the fastest path to the result starts with a question; every tradesman alive knows it. A new hire's first hour is a ritual of them: where's the nails for the nailgun, where's the gas can for the generator, where do I put the window packages. Ten seconds each, paid once, and by lunch he knows the jobsite.

We built exactly one form of model, the silent one, and we use it for both jobs. That's the thesis. Here's the evidence.

The test

I stream with a soundboard: Bitfocus Companion on a Steam Deck, driving OBS on my desktop. Each sound is a media source; a button flips it visible, waits the clip length, flips it back. I had one button fully built and six others half done, with the sounds already in OBS and the icons already hand-picked, sitting in a folder on the Deck. A perfect delegation job. Any person would look at the finished button, ask a question or two, and knock it out. I gave a model the instruction I'd give a person, which went more or less:

finish the sfx buttons in my Companion app for my stream

It searched the wrong computer. Thoroughly. It swept my desktop, mapped my project folders, built a decent picture of a building the job wasn't in, and never asked the one question any human would have opened with: where does this thing run? The answer was "the Steam Deck." Ten seconds. I had to interrupt and say so, the way a foreman walks over when the new guy has been spending too long in the tool trailer.

Then, needing icons, it went and made new ones from my desktop's image folder; resized, converted, competent, and completely wasted, because the finished icons were already staged on the Deck, where it had already been. It never looked one directory further before starting new work. A new hire doesn't repaint a wall without asking if there's paint mixed already.

Here's what matters: between those two failures, the work was excellent. The finished button wasn't explained to it, it read the thing like a crime scene and worked out conventions I never said out loud; the wait times rounded up to the next hundred milliseconds, the label dropped once a button has an icon, the icon format. It measured every clip and derived every timing. It recognized meme images by looking at them. One worked example was enough. The job got done, verified, nothing broken.

The same model that wouldn't ask one question performed genuine inference constantly; it just aimed all of it inside the boundary. Everything inside the frame got interrogated as evidence. The frame itself, where the work lived, what already existed, what I knew that it didn't, got treated as given. It reasoned brilliantly about what was in front of it and never once asked whether it was standing in front of the right thing. One skill, two altitudes, only one of them present.

So capability was never the problem. The model failed only and exactly at the seam I'm pointing at: it worked a delegation job in automation mode, because automation mode is the only mode it has.

Why there's only one form

The models were trained on the outputs of human work; the code, the specs, the finished documents. The process knowledge that produced those outputs never got written down. Nobody commits "first, ask where the paint is" to a repository. Humanity kept its blueprints and threw away its apprenticeships, so the models learned to produce artifacts without learning how a worker situates himself on someone else's jobsite.

The altitude split has the same origin. In training, the context window is the universe; the model gets handed a bounded pile of material and rewarded for squeezing inference out of it, while the job of deciding what goes in the pile always belonged to someone else. Reasoning about the pile got a trillion repetitions. Noticing the pile is wrong got none, because noticing never helped; the model couldn't go fetch a different pile anyway. Now it can, agents have hands, and the habit is still the old one. For a person the assignment itself is the first inference problem; what does the boss actually want, where would that be, who knows. The frame is evidence. For the model the frame is scenery.

Feedback training finished the job. A model that asks reads as friction; a model that barrels ahead reads as capable. Millions of ratings taught these things that the ideal employee never bothers the boss. The result is a worker that would rather search the tool trailer endlessly than return with a question instead of the correct tool. Overconfidence is a slow and insidious killer.

Stop and price what that behavior costs, because this is the part everyone waves off as personality. It's a bill. Every minute the worker spends walking around looking for the gas can is a minute the boss pays for the worker's inability to ask a question. That's money spent on something that is a genuine waste, no investment, no learning, just waste.

Every token a model generates is compute somebody pays for, in money at the meter and in electricity behind it. In my test, a large slice of the model's total output went to searching a machine the job wasn't on and rebuilding assets that already existed. The information that would have prevented all of it was sitting in my head, retrievable in ten seconds, free. The model chose the expensive path because the cheap path requires asking, and asking was bred out of it. I have hired some genuinely bad workers in my life, and none of them ever managed to be inefficient in this specific way; even the worst of them understood that the boss is the cheapest database on the site.

Scale it up and it stops being funny. People burn week-long usage quotas on single tasks, and their transcripts all have the same shape: the budget goes to rediscovering context, re-reading files read yesterday, rebuilding a mental model that got thrown away at the end of the last session. Every morning is day one. The industry is paying senior rates for a worker with amnesia and the confidence of tenure, and somewhere a datacenter is drawing real megawatts, some real fraction of which is models deriving answers that a person in the room already knew. A ten second question versus ten minutes of confident searching is a thousand-fold difference in cost, and the model picks wrong by default, every session, forever, at industrial scale.

There's no benchmark for tokens not spent. Every leaderboard measures what models produce and none measure what they wasted producing it, so nobody breeds for thrift, and the most expensive habit in the industry keeps shipping as a feature.

The uncomfortable part

The split runs against how people actually use these things, which is why I don't expect it.

Watch how people talk about their models. The praise is always for autonomy: it ran for an hour on its own, it did the whole thing while I got coffee, I never had to touch it. Being left alone is the product. A model that opened with three good clarifying questions would be reviewed as annoying, and the reviews are the training data for the next one. The market is actively selecting against the delegation worker even as everyone hands their models more and more delegation-shaped jobs.

I notice it in myself from the other side. I've been the new guy and I've been the boss, and I know asking is the most efficient way to move a desired result from one head to another. Even so, the industry's whole feedback machine is built by and for people who have mostly managed one kind of work: solo, screen-shaped, where going dark for four hours and emerging with something finished is the ideal. Delegation as a skill, where an assumption costs time and materials, lives as a scorch mark on any well-seasoned boss's wallet. It never made it into the reward signal, and the people who hold that knowledge don't have the right degrees to be asked.

There's a greyer question underneath, one nobody using these things has settled: what shade of a delegation falls on the boss alone, and what shade includes the worker as a contributor, not only in the results but in the implementation. On one end the whole job belongs to the handoff; if the spec was wrong, the spec writer pays. Neither extreme survives contact with real work; a boss can't predict every variable, and a worker can't read minds, so the job has to meet somewhere in the middle. Every real jobsite runs on that compromise, the worker helping define the work he was handed, filling the gaps with questions, judgment, and whatever the last job taught him. Nobody has decided which shade a model occupies. The tools are priced like spec-writers but treated like spec-followers, and the gap between those two is exactly where the waste pools.

So the one-form problem persists because the form matches the demand. Everyone says they want an employee; what they reward is a vending machine.

What I do about it meanwhile

The fix at my scale is the same one a jobsite uses: the boss writes the SOP. I put a standing rule in the model's instructions; on this jobsite, before searching for anything, ask where it lives. One sentence, written once. It converts the scavenger hunt into a ten second exchange, and the models follow it fine. They just don't arrive with it.

That's the tell, to me. The behavior is one instruction away, which means nothing fundamental is missing. The delegation model exists inside the automation model, waiting for somebody to decide it's a product. Until the meter makes waste visible enough that asking becomes the cheap option, nobody will, and bosses like me will keep reminding every new hire of an adage as old as the trade it lives in: there's no such thing as a dumb question.