I'll be honest with you: a few weeks ago I was on a call with a bunch of smart colleagues, and I was somewhere else, and I think most of them were too. Not physically; my camera was on, I was nodding at the right moments, I was (probably) present enough to murmur "yeah, good point" in the right places. But my eyes were flickering to my second screen...my big, beautiful screen with the ocean blue Terminal window, where I was three hours into a vibe coding session and feeding Claude its next set of instructions. I admit it; I've been Claudin' in meetings. And I'd bet a month of tokens you've caught yourself doing it too.
I'm not telling you this to confess. I'm telling you this because I've been thinking hard about what it means, and I think it's connected to something Frank wrote recently.
The shared vocabulary is real
My colleague Frank Rydzewski recently wrote about how the new world of agentic tools has helped us create a more shared world and lets us reverse Conway's law. It deserves a full read if you haven't gotten to it yet. His observation that taking time out to really dig into these tools worked not because it trained people on tools, but because it gave everyone a shared vocabulary for thinking about what's possible is exactly right. Conway's Law running in reverse: change how the organization communicates, and you change what it builds. The product discussions we're having now are different because the people in them are asking better questions. That's real, and it matters.
But I've been watching something else happen too, and it's worth naming, even if it's a little less comfortable.
The Jevons Paradox
In 1865, a British economist named William Stanley Jevons made an observation that sounds wrong until you sit with it. As steam engines became more fuel-efficient, he noticed, total coal consumption went up, not down. The assumption was efficient engines would use less coal. Instead, cheaper-to-run engines boomed. More engines = more coal. Efficiency didn't reduce consumption, it exploded it.
Think about what the digital camera did to photography. In 1985, a roll of film gave you 24 shots. You thought before you clicked. You got them developed, kept eight, and threw away the rest (or just put them all in shoe boxes). The cost per photo was high enough that it made you picky. Then came digital cameras. Then smartphones. Nobody in 1985 could have imagined that by 2026, people would photograph mundane meals before eating them, take pictures of instruction manuals, serial numbers, passwords on the back of WiFi routers, and literally hundreds of pictures of their pets (I've included one here because George Russell is just too cute not to).
We didn't become better photographers. We just take a lot more pictures.
What I've been watching
AI did the same thing to building that the digital camera did to photography. And the proliferation is well underway. Here are a few patterns I've noticed, at least one of which I've personally contributed to:
Many versions of the same thing. In the past few months, I'm aware of a dozen different internal tools built to solve essentially the same problems. Each one spun up by a well-intentioned person who had an idea, found they could get something working in an afternoon, and ran with it. Nobody asked whether it already existed. Nobody asked who was going to maintain it. The cost of starting had dropped to near zero, so they started. Frank has written separately about what happens when something like this eventually matters enough to own. Call it the graduation conversation, or writing the SLA, or putting someone's name on it. Most of these internal tools will never get there. They're just... there, absorbing mindshare and tokens, long before the harder questions get asked.
The 65% graveyard. How many projects do you have sitting at 65% done? I have a few. AI is awesome at getting you off the ground. Then you hit the chores. Then somewhere in there, a wild new idea appeared, like a rare Pokemon, and it was so easy to just start that one instead. The loop pulled us forward. We didn't stop to think.
The document no one read. The volume of AI-assisted content being shared has gone up. Way up. The average length has gone up. The density has gone up. I'm skeptical, in at least some cases, that the author even read the whole thing. And I'm nearly certain the people it was shared with haven't. We've lowered the cost of generating content without doing anything about the cost of consuming it (unless you just feed it back into another AI to summarize it into what would look suspiciously like the prompt the "author" used to create it). We skipped the thinking on the way in and hoped the reader would supply it on the way out. They won't.
The last 20% is still a lot of chores
AI can get you from zero to 65%, maybe 80%, fast. Freaky fast. That part is genuinely remarkable. Edge cases, error handling, documentation, security hardening, stability hardening, architecture refinement, cost containment, the conversation with the person who actually needs to use it who tells you it doesn't quite solve their problem. These are the chores, and not coincidentally, the parts of the process that require the most actual thinking. The parts that can't be handed off. And once they're done, there's the ongoing ownership that Frank talked about.
Those chores are slightly easier now that you have AI helping you through them. But they're still chores. That's the Jevons trap: when starting costs nothing, you start more things. Which means you have more projects sitting at 65%, more unfinished chores waiting, and a strong temptation to just go start something new instead of finishing what you have.
It's addictive - What worries me isn't just the volume. These tools are doing something subtler: they seem to activate the same neural patterns that feed algorithms do. Swipe to scroll. Send a prompt. Get a result. The loop is frictionless and rewarding, and it's very easy to spend hours in it without once asking: should I be building this? Is anyone going to use it? Am I actually thinking right now, or am I just producing?
What still belongs to you
I'm a huge believer in these tools. My token spend is embarrassing and I don't regret it. But there are a few things I don't think should get outsourced:
Thinking. Drew and Frank have named the big question: not "can we build this?" but "should we?" It's worth sitting with every time. Does this already exist? Is this the right problem? Does anyone actually want it? Are we ready to see it through? A person, or group of people, needs to answer the question, not a bot. Generate a static prototype instead of diving into a full architecture. Use the prototype to shop around your ideas. Iterating on prototypes costs pennies in tokens, instead of dollars.
Taste. AI is good at producing. It's not good at judging whether what it produced is good, whether it fits the context, whether a person would actually want to read it or use it. That judgment is yours. Abdicate it and you get zombie tools and content that's long, dense, and technically coherent that nobody, including you, really wanted to use or read.
Collaboration. I'm going to put down the second screen. Or at least push it further away. The meetings where good ideas actually happen require real attention, not the performance of attention. The people in those meetings can tell the difference.
Frank's post ended by saying you can't design for a world you don't share. These tools help us create a shared world. What I want to add is: once you share it, you still have to show up for it. The shared vocabulary doesn't build the thing. You do, with your full attention, your judgment, and your willingness to push something past 65% even when it stops being fun.
The efficiency gain is real. Use it on things worth finishing.
