THE WORK WE PAY FOR IN TOKENS, A SCRIPT DOES FOR NOTHING
Two million tokens to wake up in the morning, read the email, search the keywords and check the weather. Plus the subscription to the search service. The example is Tim O’Brien’s, I put his name in the first comment, and that’s the exact same job. It’s a script that burns zero tokens. The invoice doesn’t notice it, so it tells you how many tokens you burned, which models you called, how much the calls cost. It doesn’t tell you whether the system did anything worth paying for. And that’s the part nobody measures.
O’Brien compares it to databases, and the comparison holds. A database costs money, and from the cost you can’t tell what it is for, how it was designed, what happens if it stops. And with Artificial Intelligence it’s the same. Well, it’s actually worse. Because you put a network of agents inside, and it never gives the same answer twice. An agent writing code in place of hours of human work, a classifier sorting documents, and an orchestrator coordinating dozens of processes can all close the day on the exact same figure, and they are not the same thing.
Then there is the idea that a token is a token. It isn’t. A million tokens spent by an orchestrator are not worth a million tokens spent on a hard task. A call costing a few cents that produces bad work costs more than an expensive call that gets it right the first time. And running a billion tokens of simple stuff on the frontier model instead of a light one is tens of thousands of dollars thrown away. So picking the model isn’t enough. What counts is the reasoning level, how we handle the memory, and before anything else, whether that task needs a model at all. Because maybe it just needs deterministic software.
In the companies I go to, I’m starting to see the next step. I don’t like it. Because the token consumption of a single employee is used as a measure of how much they produce. As if whoever burns more is working more, and whoever burns less is hiding. Yes, okay, let’s count tokens, fine. They are, you know, the only visible thing. But the number that matters is another one. What it costs us in total to get the thing that system exists for. Nobody puts that on our bill, on the company’s bill. And we have to work it out ourselves. And anyone selling optimization by looking only at the bill is selling the easy part. What do you think?
#ArtificialDecisions #MCC #TokenCost #Automation #Productivity