Post #4389600

2026-08-05 02:50 UTC

@tbortels@infosec.exchange @mhoye@cosocial.ca For that specific task, it feels to me like a purpose built tool could be made that would ultimately be faster to use than an LLM. I'm imagining something where you paste in a JSON blob, then select a part you want, and it generates a `jq` command to extract that from the structure you pasted. I think I'll add that to my project list, but anyone else can feel free to make that. I'd be interested to know if that ever exists, whether it was in response to this post or not.

Replies (1)

@tbortels@infosec.exchange 2026-08-05 04:53

@mhoye@cosocial.ca @Ryan1729@mastodon.social There are utilities that do that. Many of them. The problem, at least for me, is that a lot of my real-world sources tend to be very not-regular. One log stream will have many different record types depending on the event. I have yet to find a more flexible utility than python and incremental iteration - and again, the advantage here is the AI is that it accelerates that process, pretty reliably. The filter/summarize utilities don't *interpret* and infer patterns - and it turns out that's useful. It's almost as if a purpose-built tool for analyzing streams of tokenized text and making predictions based on what it sees is useful for analyzing streams of tokenized text and making predictions based on what it sees. I repeat: the problems around AI, the real ugly ones, are all *what humans do*. Humans train on sources without permission. Humans trade stocks to create a bubble. Humans don't do their due diligence on validating results. Humans lay off humans because they incorrectly think the AI is something it's not. Humans ascribe actual thought to a pattern matcher with a feedback loop. If some asshole suns rampant in the streets with a chainsaw and hockey mask - it's not the chainsaw's fault.

Open ##4389599