Notes
What gets read changes when generation is trivial
Before, it was enough to consider any piece of writing across roughly two dimensions: what it said, and who wrote it.
With AI, we now also have to ask where a piece of writing originated from. Was it borne from lived experience? Was it written by a human, an agent, or some collaboration? Where did the root thought come from? One reason this is difficult is that it challenges a fundamental tenet of many who saw themselves as truth-seekers. Does a piece of writing that is perceptibly true have less value if we don't know its provenance? Where does anonymity — and all the literature we manufactured about it — fall now that generating words is trivial?
This is one of my academic interests. I've spent many hours pondering and writing about it. If you'd like to fund research, share ideas about paths to explore, or otherwise engage, write me.
Three things help with AI-orchestration burnout
AI agents create a new kind of burnout. The work no longer drains you through typing — it drains you through judgment. More attention. More context-switching. More verification. More decisions per hour.
In my read, three things help: (1) focus toward a specified goal; (2) maintaining personal motivation that lives outside the project's outcome; (3) accepting that you will never be 'ahead' of the technology. It seems more sustainable to leverage these tools when one has a clear sense of the end they want to reach. Because the tools expedite both failure and success rates, you have to externalize continuity outside the outcomes themselves. When production is commodified — as it is now — you carve out targets for success outside whatever the latest harness can deliver.
Sport is where new technology shows up first
In the past two weeks, a human broke the marathon record in London, and a humanoid broke the half-marathon record in Beijing. It is increasingly the case that if you want to keep abreast of the edge of technological development, sport is the place to look.
The financial markets have already taken note. Sebastian Sawe's sub-two-hour record is the beginning of a much larger revolution. Technological breakthrough as footwear and augmentation — the super-shoes story — is one half. The other is the cooling tech inside the Beijing humanoid: liquid-cooling adapted from a smartphone product line, two micro-pumps reaching six liters per minute, capacity to handle sustained lower-joint motor load.
Sport is the demo arena. Watch sport for two cycles before the same technologies show up in the rest of the economy.
The first opportunity I held — ICIR, 2022
The first opportunity I held right out of high school in 2022 was as an investigative journalist back home in Abuja. On my first day, the editor assumed I was a recent graduate or a college student in journalism. I was eighteen, fresh from a two-year stint at the African Leadership Academy in Johannesburg, and I had pitched myself with two earnestly-written essays from my sports blog, BallerzBantz.
I intuited the editor had given me an unspoken pass to present myself as I wanted, so I let the ruse simmer while I learned the work. In four months I covered the World Cup, was deployed for two field-reporting stories, and got a taste for what frontier, meaning-making, edge-of-the-field work looked like. One exposé unraveled a ₦2bn sink in Nigeria's Sports Ministry — abandoned fire-trucks that had sat idle in the national stadium since the 2003 COJA games. (I traced the snapshot back through Google Earth.) Another investigated voter-registration suppression: roughly 1–2 million eligible teenagers disenfranchised every cycle.
These were preposterous, and in plain sight. That has become the multi-year obsession: investigating, empathizing with, and curating solutions to every domain I get into contact with. Sometimes that lands me in journalism. Sometimes football. Sometimes student governance. The current research and product sprint is just another flavor of the same itch.
When the moat is open-sourced before it's monetized
Andrej Karpathy shared an idea-file framework for coordinating AI agents to instantiate and self-sustain knowledge production in a file system — Wikipedia, but the verifiers, reviewers, and back-linking are done by agents.
In the days after, I tracked roughly ten platforms or founders building (or having raised significant capital to build) the same thing. Most had open-sourced their builds up to the moment of the tweet. Some published padded posts about how they sustained similar systems internally. Many of those posts went reasonably viral, which suggests the bottleneck is real.
A few of these were scoped as B2B products with humans-in-the-loop. But: significant consumer engagement. Many operations (mine included) had already built internal versions. And the surfaces are all now open-source — you can drop Karpathy's idea.md and have an agent build the whole thing. So what remains for the companies that scoped this as their primary product? I don't know that they have a defensible moat.
When the limit moves from execution to abstraction
If you ask a capable AI agent to compute a number, it will do so. If you ask it to find every instance of that computation in a codebase and replace, it will do so. If you ask it to build a function that does the computation, it will do so. If you ask it to refactor the codebase such that we never need the function again, it will do so.
The limiting factor in any technical endeavor becomes the operator's: (1) abstraction, (2) ability to hold complexity, (3) focus and vision, (4) acceptance that there are things they don't know that they don't know — and that the agent might be able to know, research, or implement some of these toward the operator's vision.
Mechanical vs. prompt-engineered harnesses
Across recent sprints, I've had to make decisions about using mechanical (script-based, JSON-enforced) harnesses or prompt-engineered ones.
The mechanical ones are useful — easy to append a UI layer so I can run the operation manually if the agent fails. But it takes considerable overhead to scope every edge case. The prompt-engineered harnesses prototype faster and afford the agent more creative freedom, useful for research where there is no single best path. The drawback: harder to debug, and harder to know when you need to debug.
One tool that helps: ask the agent, post-run, to reveal any tensions it experienced while attempting the task — then use the output to refine the harness. There's a third consideration too: models keep getting better, so a heavily-mechanical harness can constrain the perception of eventual improvement.