How to work around all of that abundant intelligence

Let’s take a sec to take stock. I watched a video, or listened to it on my commute. And I could not help but pause and wonder how this is the pinnacle of progress.

Nate B Jones is (no idea how you can have a middlename initial with no period) is a what you must call a tech-savvy, eperienced and sophisticated user of AI. He comes from the industry side — product, Amazon, the practical machinery of getting things shipped — and now publishes a firehose commentary on AI for executives and knowledge workers who want an edge.

The video I have in mind is called “My AI Workflow Has Changed (Here is What I Learned)”. It is short, about six minutes, and it is part of a format where he explains how he has been using AI that week.

The workflow he describes, what he does “when I want serious work done.”, is this: He asks Codex¹ to look across his local file system and find files by description. Not exact filenames. Not neat paths. More like: find the thing I made around this time, about this topic. Codex finds the files, copies them into a clean working folder, and then he starts a fresh Codex chat pointed at that folder. The folder becomes what he calls a “clean context window.” If he has detailed instructions, he puts those in the folder too. Then the agent works from that staged environment. Nate says this lets him work with 30,000 to 50,000-word documents, spreadsheets, code, and multiple drafting tasks.

I do not want to strawman this. It is a good workflow. I can see why it works. I do versions of the same thing myself.

But that is exactly the part that made me pause.

If someone like Nate needs this workaround, what does that tell us? And what does it tell us that he thinks that this is the part he wants to share with his quite substantial audience of AI-forward people?

I do not mean “needs” as in: this is a nice productivity trick. I mean: if a technically proficient, AI-obsessed, professionally incentivized user, whose work is to test and explain these systems, finds it useful enough to teach a large audience that serious work means first assembling a clean little folder-world for Codex, then this is … a fact about the current shape of the technology.

The strongest users are still building prosthetics around the strongest models.

So the operator has to know what matters. Know which files belong together. Decide if and when the context is too noisy. Understand what should be excluded. Instruct as to which standard applies. Keep track of and intervene when a plausible answer is unacceptable because it is solving the wrong version of the problem.

This is why the “AI will soon do the work” story is true and false at the same time. True, because the models can, under certain conditions, do surprisingly large chunks of work. False, because the work only becomes safe and useful after someone has done the meta-work of shaping the environment in which the model operates. Which as of yet is bespoke, on the fly, and not done once and then never again.

Anyway, that is the current cutting edge, and I would not argue it is different, from my daily battles with the tech. If your mileage varies significantly, you should become an influencer/creator and send me your stuff, please.

Footnotes

1. Just to be clear: you can swap Codex for Claude, or whatever. It is the latest and greatest, available to the public at this time. ↩