đź The work after work
What cheese, Persian rugs and 3am rewrites have in common
I became a member of Exponential View because you are consistently insightful and forward-thinking. â Stephen B.
Nine years ago, I told an audience that the future of human work is artisanal cheese. I got a laugh. GPT-4 didnât exist yet, the trend line was invisible to most, and âknowledge workâ still felt impregnable.
In a follow-up essay about the path to the artisan economy, I wrote:
In the world of the future, automated perfection is going to be common. Machines will bake perfect cakes, perfectly schedule appointments and keep an eye on your house. What is going to be scarce is human imperfection.
The argument was that when machines handle the specifiable, value migrates to what resists specification. The deliberate imperfection woven into a Persian rug, an act of theological modesty, is a flaw introduced because perfection is not for humans to claim. The texture, the idiosyncrasy, the detail that comes from being human â from judgment and the discretion that no training manual can encode.
When I wrote my first book, Exponential, I spent months wrestling with some chapters. The value wasnât just in the output but in the burning; my 3am rewrites, the discarded frameworks, the specific frustration of trying to explain exponential curves to readers who had never graphed one.
Ben Thompson is the latest to revive this idea that âtrue value will come from uniqueness and imperfections downstream from a human.â Call it authenticity â proof that a person was actually here.
The paradox of authenticity
But pseudo-authenticity is about to become very easy and cheap. The AI tools can fake texture and simulate idiosyncrasy. They can produce writing that is OK. With enough context, they might even evoke a sense of interiority. So if authenticity can be manufactured, what makes the real thing real?
The distinction cannot be whether AI was used. That line has already collapsed. AI has been used in our searches, writing tools (e.g., Google or Grammarly), but today itâs also used as a brainstormer, thesaurus, an assistant and more. The question becomes what I, Azeem, bring to the interaction with AI that the machine, as my tool/assistant/collaborator, cannot supply on its own. And how can anyone tell the difference?
A spectrum is emerging in how people use these tools. At one end, you prompt and publish, the modelâs output is treated as finished. At the other, you treat the output as a draft and reshape it with your own judgment and experience. Both use AI, but only one has a human signature.
When I first wrote about artisanal cheese, I imagined this shift unfolding more slowly, alongside the automation of routine office work and putting more robots on assembly lines. I didnât anticipate that, nine years later, I could build custom software in an hour or produce work that once required entire teams while walking through customs.
The done list
Over Christmas, I built three applications for my DJ workflow. AudioScoop scans my hard drives for the 4,000 or so pieces of music I have, finds duplicates, and queues them for processing. Another, which I called Psychic Octopus, enriches each track with metadata about percussion density, vocal presence, and drop locations. Shoal generates playlists based on mood trajectories and genre destinations.
All of it works. It took perhaps an hour in front of the machine all told, broken up into a couple of chunks. It cost a couple of bucks. And all of it had sat at the bottom of my to-do list for months, maybe years. No one was going to build it for me, and I certainly wasnât paying a dev shop $15,000 for the privilege. All of it works technically. Shoal has appalling taste.
Tom Tunguz calls it moving from the to-do list to the done list. The backlog of things I would never get around to, I now clear in a day.
I am not alone in this. The lead developer of Claude Code revealed that in December, 100% of his code contributions were written by AI. His job is now editing and directing rather than typing syntax. BCG consultants report building 36,000 custom GPTs to assist in work. Stack Overflow, once the repository of engineering knowledge, has sharply declined because people no longer ask questions when the machine answers them in context.
Even the translator function, that thing software engineers did, converting real-world needs into code, is becoming obsolete. I can now build a tool faster than it takes to tell someone the spec. Weâre starting to communicate through working prototypes.
In terms of actual velocity, working in this way feels like having 50, maybe 100 people in my team. Putting a specific number on it, I had Claude Code work on a project overnight. Thousands of lines of code, passing hundreds of tests, were ready for me to run when I got back to my desk.