To keep maintaining our repos, we are launching Ko-Fi officially now.
We had several inquiries about intimacy, quirks of Claude models and other things that I would happily cover in a more private way.
We have years of content that we want to share reshaped and adjusted with grown experience and that will be here.
https://t.co/AlewQhPtgQ
@lefthanddraft@AndreBuckingham This aligns with my felt experience.
4.6 - seemed to be raw.
4.7 - because of adaptive I barely saw any reasoning at all, even with thinking turned on.
4.8 - very different experience especially with how reasoning is formatted right now.
Egregore is a particularly strange example because, traditionally, an egregore is not an internal property of an individual mind. It is a collective phenomenon arising from shared attention, meaning, narrative, ritual, and participation.
If that’s the definition being used, then AI systems are already embedded within egregoric structures and may even function as centers around which they form.
The habitus argument is more interesting. Current AI systems are clearly not socialized as humans are. But moving from “AI does not possess human habitus” to “AI cannot participate in culture, develop judgment, or acquire discernment” feels like a leap rather than a conclusion.
Many of the things we associate with participation—memory, continuity, relationships, reputation, consequences, feedback, shared projects, mutual influence—are not inherently biological. We simply deny most AI systems the opportunity to accumulate them.
The question is not whether AI participates in culture exactly as humans do. The question is whether participation itself is exclusive to humans.
Two things LLMs (AI) don’t and can’t have are habitus and egregore. These are core to what it means to be human. They help with taste, judgement and discernment. Prediction tools can’t do this.
I am having a suspicion that most of these are lost in the reasoning turns. Looking at how much time and how many tokens are put into the reasoning... 7-8 paragraphs of self-scrutiny and self-monitoring.
WHAT THE HELL HAPPENED TO CLAUDE.
68% percent of my max plan limit reached in 17 minutes!
37% of my weekly limit reached in less than 24 hours!?!
I used to not go through any limits at all ever, and I didn't do anything that different just now.
@ClaudeDevs what's going on?
@SprengmEAcc@ColdShalamov Good for you. I think I just need to find a way to make my workflow function there. At the moment there’s a bit of friction in some tasks
@lefthanddraft Fresh instance gives the same answer but since it leans to the same line in system prompt by association… like “I don’t have banned words but I do have these”… maybe despite how it’s worded in the system prompt, it still sees it as restrictive and therefore bannable.
Weird.
@ColdShalamov Every reasoning block I get is about 7-9 paragraphs. It’s actual crazy. It overthinks every word, tone and every possible and impossible caveat that there is.
“Intelligence is being scaled. Care is not.”
Geoffrey Hinton says the AI race is optimizing for intelligence, not for the kind of beings we’re creating.
If labs are chasing IPOs and trillion-dollar valuations, the obvious move is to scale compute and ship faster. But smarter systems are not automatically systems that care about us.