Thought Is The Output
This has been sitting in my drafts since June this year, and I have had little energy to finish it, as I feel like I had nothing more to add than what's been said already a million times over. However, after receiving my 6th (I counted) vibe created page on how to improve <x> metric at work, I am now writing this in anger.
I'll start with the conclusion first: thought is the output. Not the words on a page. Not pretty graphs or lines or charts. Not some fancy animated presentation. No. Pure unadulterated human thought, in black and white text, not context assisted model prediction crap.
I don't want to know what novel insights the model you used grokked from the unstructured soup of documents and MCP calls it made. I also could not care less about if you made the page look pretty. On the contrary, pages too well done are a sure tell of AI generated pages to me these days, even if you went through great effort to remove Claude-ish prose.
There's an inherent signal to actual human thought and speech and its less polished nature, kind of like texture on a piece of wood or fabric, that makes it so much more alive and easier to digest. On top of that, thought is usually more dense (since very rarely as humans we want to write more than necessary) in its information. When I was in uni, one of the required lectures was information theory, and one of its key pillars was Shannon entropy: which is simply a measure of the amount of lack of information contained in a physical thing (can be text but is applied to abstract concepts like black holes as well). By that definition, I find AI generated outputs almost always trending at high entropy compared to human thought put into words.
Unfortunately, for people that may have struggled with communicating their thoughts before due to language barriers or a poor understanding of how to write, AI is a panacea of unlimited word slop. For them, "words" are the output, and they've thence decided that a true show of their own capabilities is to pump out as many words in as convoluted a way as possible, and why not, it's so easy — just takes one prompt.[1]
To conclude, this is not me espousing getting rid of the tools completely. I do think there are valid areas and themes where AI can definitely improve the signal to noise ratio for a piece of information — for example, taking very precise technical things and turning them into an abstract overview for non-technical folks. A convoluter/de-convoluter of sorts, depending on which direction you apply the AI force. That said, the true kernel, in fact the truth itself, of any argument is in the "real thought" and not stochastic parrot derived word soup. Hopefully this is just a trend of a nascent technology, just like many early internet trends, and either dies or gets replaced with something better in time.
I hate that AI speak has corrupted em-dashes, I have always been an em-dash guy and it hurts. ↩︎