Human of the Gaps

Happy New Year Everyone! The DMZ on my shoulder has been nagging me to write a post — even though it feels like LinkedIn has become bots writing content for bots to read. But here goes anyway. This article has been nowhere near AI (other than the picture) so it will likely read as somewhat antique.

A couple of days ago I was sitting in the lounge at Vancouver International Airport. Demelza Green gently reminded me that I had promised to write something about the use of agentic AI in large scale app modernisation projects. Maybe not so gently. We talked about doing something with COBOL — a dead language in common use, kind of like Latin for the developer high priests.

Air Canada announced that our flight had been delayed. It was now scheduled to depart at 1:50am, in five hours. Earlier on they had messaged about a plane swap, so obviously somewhere, someone knew things were changing. Depending on where you looked — in the app, on the website, on the departure board — you got different times for boarding, departure, arrival. None of them agreed.

I started thinking about how unnecessarily complex those old systems must be, how many layers of unfathomable abstraction sat between reality and what was being presented to us.

So I prompted up a COBOL app to check the location and status of late running aircraft, complete with a little telnet / ASCII flight radar. I can't recall ever having seen COBOL before. I'm not a fan (COBOL isn't an ideal language for parsing json). But like all such AI interactions, it's an entertaining mix of productivity, experiment and frustration.

The UI it produced looked almost exactly unlike Australia. I pasted the rendering back in. Claude agreed it needed some work — it commented that the Cape needed to be a bit more pointy, and that the Bight was missing. It was an odd observation given what it had produced wasn't even an island. We went back and forth a few times, each exchange slightly more surreal than the last, until I realised that this wasn't a coding problem at all. It wasn't struggling to draw Australia, it just had no idea what it looked like.

This wasn't going anywhere fast. I asked it to create a shared function that mapped GPS coordinates to ASCII column and row locations. I told it to get a set of coordinates outlining Australia and map them using simple symbols. Use the same function to overlay major cities and aircraft locations.

A couple of minutes later it was working. Clearly AI needs people like me.

The Blur

The last twelve months have been a blur. I've spent nearly half of them away from home, at conferences or meeting with potential clients. AI dominates everything. Every conversation eventually winds up in some kind of philosophical musing about where this is going.

The message, repeated ad nauseum, is that we need not fear AI. It's going to create so much work for everyone that we'll all be busier than ever. Experientially, this does feel true. I've not had this much fun coding in decades. In the last couple of months, I've reverse engineered the Bluetooth protocol on my camper's battery management system. I've rewritten the app that runs my house in React Native. I have several Qwen based models distilled to run on a Mac mini that now quietly look after large parts of the infrastructure around my house. A CrewAI team is currently rewriting the microservices on my RPI k8 cluster. The speed to realise an idea is intoxicating.

Of course, the agents aren't perfect. They run off track frequently. I have to figure out why, work out what assumption was wrong, what instruction was ambiguous, how do I get an agent to break the work down into steps that reduce the chance of failure, and that another agent can implement or check.

The Analogy

And the overarching feeling is still one of optimism. The analogies are familiar: the industrial revolution, humanity freed from the drudgery of a rural existence, only to discover the joys of the production line. Industrial automation frees us to be designers, project managers, coders and staff the help desks. These stories reassure us that there will always be something left for people to do, even if we can't exactly see it from here.

So our destiny is to do more of the creative, fun, cognitive work — and in the short term this genuinely seems to be happening. But I think the analogy is probably wrong. The better analogy, for me at least, is the rise of the scientific worldview and the slow retreat of the divine. For a long time, everything science couldn't explain was evidence of God. The gaps were expansive and mysterious, and human meaning lived comfortably inside them. The earth was 6000 years old until they discovered dinosaurs. And even though science advanced, those gaps didn't disappear all at once. But inexorably they kept shrinking.

I see the same thing happening with AI. Yes, it still needs humans. Yes, working in those gaps is exhilarating. There is real joy in being the person that notices what the system missed, the edge case it didn't consider, the prompting that unlocks a better result. But make no mistake, the gaps keep shrinking. The number of things humans can do better keeps declining. The endpoint has to be the obsolescence of human cognition as a comparative advantage. These models already code much better than I do, they just have some blind spots that I can cover.

The Utopia Question

I think this is why Elon Musk and Sam Altman talk about a utopian future where we don't need to work. Elon says we don't need to save for retirement. Sam promotes Worldcoin as a UBI proxy. They see the logical conclusion of this. The entire point of AI is to outthink and outcreate humans.

Even with the capabilities we have today, I doubt there is any organisation on the planet that can genuinely adapt to a fully AI native structure. People use these tools, experiment with them, even enjoy them. But beneath the rhetoric about abundance and opportunity, they understand what this means. We all sense it. And because of that, most people can't or won't fully embrace it.

But for now, I'm still sitting in airports. Flights still get delayed. Systems still disagree with one another about what's happening. There is still a place for a human in the gap.