AI is a step change in the ability to go from idea-to-reality and is assumed to be a boon for productivity. In this post I challenge this belief in order that we consider the downsides and put in place measures to address them.
There is an increase in the quantity of output from humans
Achieved by
Humans never used to send me code. Programmers did. Now I get oceans of code from people who used to be programmers, or never were, but are now. I would seldom get a written paper with enthusiastic advocacy for a widget, now it is weekly. Those people still don't write those things, but AI does for them. And now those documents are 11 pages to read. But the diagrams look nice.
We have moved from "good ideas in spreadhseets and powerpoints" to ideas expressed via code. I observe dashboard demos - containing python, javascript that are being built outside of governance.
The initial impression may be positive - it looks good visually - however I believe these are the equivalent of LinkedIn's emoji-heavy "let's break this down" - style over substance.
We are in a period of slop. We need to address this before we reach the long sloptober - an endless content output - as this is not the productivity we are seeking, rather it is a side quest that achieves the opposite. We cannot ever exit slop - this phenomenon is going to be comparable loosely to social media addiction - but we can create guardrails to minimise and raise an effective new culture of product engineering.
There are two steps to this:
At the end of some period of work, we need to rigourosly evaluate the output. I propose we extend this to include assessing the method and toolkit used to achieve the output. That is, we need to defend against ourselves and enter a permanent iteration of our tooling:
This means we need to kill our darlings. Good ideas, tools, processes, harnesses, repeating solutions - all need to, periodically:
Proliferation of words in documents is a challenge; I cannot tell at first glance if they are the authors words.
When I find an equivalent of "let's break this down" - which is to say, a written style in your words that is not the human I once knew, I do not put this down to "they've read a book, a new word in the playground".
What I do is put the authors work into a slop bin in my head. I have found myself losing respect for the (AI)written word. The danger is that I miss the human written word. The other danger is that the human did not truly read/write their own homework. This is slop. If it takes me longer to read that it did to write, it is slop.
A conundrum - I will not write code again, much. I mean, I will never not write code - but mostly I will tell claude to do it seldom using vim myself. That's ok. It's actually amazing and great, but Deep Blue gets me sometimes.
I hated the sycophancy from the early agents and told them not to tell me how wise I was in a CLAUDE.md. They stopped. Then AGENTS.md rocked up because standards and I put them in two places. But really I keep reverting to claude. I see documntation explaining how great the thign is they document is explaining and I do not want advocacy in the document itself.
Signal to noise is off the scale, and we have not yet accepted it.