Seems fun

Some companies have brought this on themselves by incentivising employees to burn tokens, making it a performance metric, hence “tokenmaxing” became a logical employee response. In something like coding where AI capabilities are already quite productive, some kernel of this maybe made sense as a way to get employees to adopt the new tech into their workflows. But obviously it’s easy to take things too far. An agentic workflow can burn quickly through tokens because the stuff triggering the AI to “think” isn’t limited to manual human input anymore.

Some bits of that comment Norman found didn’t ring true from a system administration or security perspective. You wouldn’t want to give AI free rein over absolutely all your sensitive files, and why would it be necessary?

There’s also the issue of AI context windows. When they’re initially trained at enormous data centres for billions of dollars, an AI can handle unfathomable quantities of training data. Only the big players can do that. But you can then take their model and train it up on more of your own data to make it more specialist in what you need. At this stage it can’t handle very much before you get to the point learning one more thing means forgetting something else. A manager on a complex software product told me they can’t even train an AI on their documentation to serve as an assistant for end users, because there’s just too much of it. They’re having to produce a massively cut down set of docs as the first step. Similarly building AI support agents for workers at a large company – even if it was desirable to make an all-in-one agent that could cover all roles and processes, there’s too much for it to learn. So it has to be more granular, not a free for all.