We need to talk (more) about AI

About a month ago, I came to a sudden, disorienting realization: the widespread societal transformations I had occasionally imagined happening as a result of AI might in fact be quickly approaching, and we are (or at least I am) far from ready. In this post, I’ll walk through how I got here and what I’ve been up to since I had my little epiphany.

Remember the good old days, back in 2022? ChatGPT had just been released and I, along with the rest of the internet-connected world, was having a great time playing around with our new… friend? Toy? Whatever this thing was, it was certainly entertaining. And probably useful. But mostly fun.

Me introducing my family to the latest internet-craze

Eventually, we learned that we could do more with this tool than cook up imagined dialogues between our favorite non-contemporaneous personalities or generate lyrics for an “uncomfortable lullaby” (a personal early request). It could draft our emails, help us plan trips, and teach us new information. Occasionally that information would be false, but it usually seemed to check out.

Of course, we all used our newfound, variably competent personal assistants in different ways at work. As a software engineer, I could now feed error messages that I encountered directly into ChatGPT, and I could even copy in chunks of code as extra context. After reading my input, it would suggest a solution. Often that solution did not work. But the tool was useful enough that I kept coming back.

From there, things mostly plateaued for a while. Every few months, a new version number or model name would appear and some new quirk would start showing up in the AI’s output. Still, there were plenty of times when I preferred to get answers from a real human. AI was another tool in the toolbelt. Not always reliable, but good enough to be helpful.

Towards the end of 2025, my interactions with AI took a dramatic turn. Around November, my team and I began integrating something called “coding agents” into our workflow. That essentially meant having tools like ChatGPT and Claude that live inside the code editor instead of in the browser. They can read all of the relevant files directly and are able to edit those files on their own.

To illustrate the point, I’ll move out of the world of software engineering. Imagine you are writing a 400 page novel and, for whatever reason, you decide you want to edit a certain side character out of the story. You can’t just search for every time that character’s name appears and delete it. You’d end up with disjointed sentences. You would need to go through each instance individually and figure out how to best smooth over the text.

In this imagined scenario, you wouldn’t be able to copy your entire novel over into ChatGPT. The prompt would be too large. But what if there was a ChatGPT feature built into your text editor? You could ask it to make the change and it would search through and smooth over the text on its own. It wouldn’t be perfect, but it would probably make a solid first pass. That’s what coding agents are doing for software development. (I don’t use Microsoft Word much these days, but from what I understand something like this already exists.)

This was the crucial turning point for me. AI had quietly transformed from a glorified search engine to a strange twist on a junior software engineer that somehow also seemed to have decades of experience. In other words, it was actually doing my job, or at least trying to.

Slowly, I settled into this new reality and started to build a routine around it. I would feed the AI descriptions of features we wanted to build or bugs that needed fixing in plain English and it would explore the existing code and take a first pass. Inevitably I would need to correct, improve, and sometimes completely redo its work. Like a good engineer, I played around with my prompting strategies to try to make things flow more smoothly. But I sensed that something big had just happened – something that would take more than prompt-tweaking to address.

After a few months of winging it, I finally decided to sit down with Claude to try to understand what new skills I needed to stay ahead of the curve. It pointed me in two general directions. First, it suggested that I brush up on my management skills, since there was something analogous between managing AI tools and managing less experienced people. It also recommended that I level up on my bigger picture software engineering skills, like knowing how to organize a large codebase to keep it easy to navigate in the long term, because the AI tools had some clear gaps in thinking at that level. Both of these made sense to me, so I gathered some resources and started digging in.

Before long, though, another thought started bubbling in the back of my head: How long would these new skills remain useful?

If the technology were to stay where it was, I would be fine. My day to day workflow had already been turned upside down, but I would figure things out. The trouble was that all signs seemed to be pointing towards AI continuing to improve. I had heard people argue that AI development will hit a wall some time in the near future. Either the chips won’t be able to process more data, or the software will hit its theoretical limits. One way or another, the models’ capabilities would max out, leaving plenty of work for humans to do alongside them. But it seemed like, as a society, we were pouring all of our energy and resources into finding ways for those things to not happen. Sure, maybe they would, but that wasn’t part of the grand plan. If the stability of my career, and my financial future, rested on the assumption that “AI will stagnate soon”, I could be in for serious trouble.

And, of course, these concerns weren’t just about me. This new generation of tools had only entered my world a few months earlier. I hadn’t seen it coming, but suddenly I was calling the long-, and even medium-term stability of my career into question. It seemed at least plausible that similar shifts would come for other professions, and that they would arrive faster and hit harder than most people are prepared for.

I went back to Claude and shared my thoughts. After some back and forth, it pointed me to the names of a number of people and organizations that I might want to look further into. One of those names was Erik Brynjolfsson, an economics professor at Stanford of whom I had never heard, and the director of the Stanford Digital Economy Lab. Claude told me that he “takes the displacement question seriously” and left it at that.

A quick YouTube search led me to a CNN interview between Brynjolfsson and Richard Quest about a project called The Digitalist Papers. This sounded like it was right up my alley. Brynjolfsson and his co-editors had assembled thirty of “the top technologists, policy-makers, economists” – people like former Google CEO Eric Schmidt and Nobel Prize winner Joseph Stiglitz. This esteemed group was given the task of writing essays about “transformative AI,” its anticipated economic impact, and what we might be able to do to smooth out the transition into this brave new world.

I’m not going to use this space to summarize the content of those papers. If you’re interested, feel free to check them out here: https://www.digitalistpapers.com/. I haven’t finished reading them myself. For the moment, my focus is more on their “In the Media” page, and the surprisingly small number of links featured there – eight, to be exact. To be fair, I have never been treated to a seven minute interview on CNN, nor has any project of mine ever been featured in the New York Times. But I’ve also never been the CEO of Google, and I don’t have a Nobel Prize hiding in my back pocket.

From what I could tell, the general conversation around AI was not where it needed to be. Experts and academics were talking about changes at the scale of the industrial revolution and how we should prepare for them. Aside from a few brief pieces in major news outlets, hardly anyone seemed to notice or care. The reaction felt on par with those towards other “what if” disaster scenarios – things like nuclear war or another pandemic. Disturbing images, perhaps, but not something that many people feel compelled to actively prepare for. The difference, of course, is that as a society we at least pretend that we are trying to prevent those other scenarios. AI development, on the other hand, is something we actively encourage.

To be clear, I am not here to advocate one way or the other on any particular AI regulation or policy initiative. And I recognize the possibility that the technology will reach a plateau, or at least slow down enough to give us time to gradually reorient ourselves. My point is only that AI has the potential to trigger enormous changes quickly, and not nearly enough people seem to understand how sudden and dramatic those changes could be. We should be thinking and talking more seriously about what that might look like and how we should prepare for it.

In lieu of answers, I'll close with a question, echoing Will MacAskill in episode 213 of the 80,000 Hours Podcast, where he builds out a picture of AI developments causing a “century in a decade.” Think of all the changes you expect to occur in the world by the year 2126 – technological advancement, economic upheaval, scientific discovery, geopolitical conflicts. Now imagine all of that happening in the next 10 years.

Are you ready?