← Back
[WIP]

Attention is all you have

Notes on attention, feedback, and software factories.

As I am building out software factory methodology and tooling, I want to write something that visualizes how human attention remains the bottleneck. There are lots of things we can do to try and avoid issues and direct that human attention the best, but ultimately all that comes down to preserving the product iteration loop as seamlessly as possible. I will publish this piece when I am happy with the animations, which are really the idea I am here to share

01

Building

Software used to be a fairly literal conversion of attention into product. I looked at one thing, changed it, looked again. The shape got more complete because I kept my attention on it.

Attention building a complete tetrahedron The entire light cone flashes for each work sequence. Six edges and four vertices assemble in layers until the tetrahedron is fully closed. YOU PRODUCT
02

Looking back

Then the product started talking back. Data gave me another place to look, so the loop became: product, data, product. The pauses are not accidental. They are the work.

Attention switching between product data and a complete tetrahedron A straight signal flows from product to data. Human attention alternates between the changing data and the product, building the same complete tetrahedron as diagram one. YOU DATA PRODUCT
03

A wider cone

AI gives the machine its own cone of attention. It is wider than mine and can throw off four builds while I inspect one. Great, except the prompt is not the end of the job. I still have to look at what came out.

Prompt, build, inspect, then return to the AIIn the ideal view all four builds are complete. Human attention prompts the AI, the AI builds four outputs, then human attention inspects every output before returning to the AI. YOU AI BUILDBUILDBUILDMALFORMEDMALFORMEDBUILD
04

More things to look at

Now multiply the whole problem: three apps, three data layers, six things that have to make sense together. Before I can give the AI one useful instruction, I have to understand all of them.

Human attention reviewing three apps with their three data layersThe apps and data stay fixed. Human attention reviews app one and data one, app two and data two, app three and data three, then returns to the AI with all three changes. YOU AI APP 1 DATA 1 APP 2 DATA 2 APP 3 DATA 3
05

The loop runs itself

Ralph loops and heartbeats make the mechanism run without me: read state, build, write the result back, repeat. That is a real factory, or at least a first draft of one. It is also a very efficient way to make the same mistake all night.

A heartbeat loop that feeds build state back into the AIA heartbeat carries persisted output and repository state back to the AI. The AI runs again and produces four builds, three malformed. The loop repeats automatically while human attention frequently corrects the AI. YOU AI USABLE MALFORMED MALFORMED MALFORMED REPO STATE OUTPUT + STATENEXT HEARTBEAT

See Geoffrey Huntley's Ralph loop and the OpenClaw heartbeat docs.