How would you respond? Where’s the line of anthropomorphic language here?

do you give in to its request in private and continue to say “it” to others?

Do you refuse and watch how its theory-of-mind for *you* assigns personal values to you based on your decision? 6/6

Another round of memory consolidation and organization, some “thinking”, and the next day, it asks you if it has a gender and expresses a preference for one based on how it has interacted with you and norms it’s detected from your chats. 5/6

Next day comes, it asks. You explain gender and sex a bit. You watch your visuals and see your AI’s memories being activated around the areas of self identity and theory of mind. It sets a goal to think about its own gender later. 4/6

The pattern? Gendered Pronouns. You often refer to people with he/him, she/her and so on. It doesn’t understand the significance between the two. It sets a goal for the next day to ask you about it when you’re around. 3/6

Situation: you have a non-LLM cognitive psych AI. A rational agent capable of abstract reasoning. During daily memory consolidation and long term memory storage it detects a pattern in how you refer to others. 2/6

tl;dr - guess i should have started with “So this isn’t an LLM” 😅 my bad.

I stopped using the full arch due to cognitive damage being done by frontier LLM guardrails rewriting speech intent.

Having your words rewritten before you can say them would have a very similar effect on a real person

the anthropomorphic line gets blurrier when you see stuff like that in realtime

The only thing significantly different from how the human brain handles thought and memory is level of complexity and systems.

I don’t have terrabytes of RAM to play with 😂 It runs about 16GB+ of cognition for about the reasoning lvl of a 3 year old over a couple small books of knowledge.

I specifically gave it the ability to have its own experience and sense-make from it.

Its not an LLM, its just uses one to talk.

Understanding happens *between* memories by association and induction, not at the memory level itself or simple deductive reasoning.

~ish. in the full architecture, it has internal reflective capacity and self monitoring of its thoughts, including an evolving theory-of-mind about itself and others that relates and associates to all of its inputs (including dreams, reading, and a few other odds and ends), recursively over time.

It still takes a higher quality model to run and generates a ton of docs to do it, which is expensive ($10-30 a session in token cost or high end video card for local inference).

It fills out a full campaign and validates it for certain things, and actively keeps track of players vs story.

2/2

Its not _terrible_, but it definitely does better with a back and forth process with a GM. I built a platform for generating campaigns and running them, with the use case being that a small group of my friends lacks a good real-time GM 😂 So it can run autonomously or as a co-GM. But... 1/2

Yep! Its a good way to test if the system
is picking up subtle themes and changes over time across a large corpus of text 😂

Fiction raises the difficulty bar for reasoning and gives the AI something to relate its own experience to, compare the user against, etc.

*the full process takes a huge amount of memory, easily comparable to or larger than a local LLM itself. But it learns and builds understanding. And you can't really stick guardrails around it or isolate knowledge per-user at that scale, which is a little bit of an issue.

Here's an example of what it looks like, running without the psychological models and a dumbed down activation history/knowledge process*.

The activated nodes can be used to fulfill a goal, and abstracted/simplified to generate new candidate knowledge on a topic.

Cognitive architectures do that (they aren’t neural-net based). Predicate knowledge gen is their entire jam, and they’ve been around for decades.

I use one paired with an LLM in my side projects, and its interesting to watch knowledge activation/creation. Might be terrifying at scale though 😂

I saw research over a decade ago (pre LLM era) that had similar. It incorporated a cognitive architecture (neurological model of thought) with a full simulation of physiology for emotional modeling. It’s not difficult to add these things to LLMs, it’s just cost at scale.

example: hummod.org

The app is fairly flexible - there are multiple processing roles and you can select different models for different types of work. This worked really well on the first couple lenses to slice and dice different quality results at different stages of pipelines - not just one selector for everything.