For serious Nuero
Why have one or the other when you can have both
"Personality, Reasoning, And Memory are all part of one thing/concept/idea"
It is a compound concept that doesn't have a clean English name, so that's what I came up with.
From my studies, the three have massive overlap, and the more you define them in detail, the more they converge.
Hence they are a "compound concept"
Personality and Reasoning overlap heavily, but I separate Personality because it's so broad. There are hundreds of ways to categorize and define it.
Reasoning — over centuries and especially the last 26 years, thousands of standardized reasoning methods have developed, both implicit and explicit. Think of it like hundreds of checklists for how to do stuff, ranging from highly general to highly specific.
Memory — I think of this in terms of storage, sorting, and summarization, more computer-focused here given the AI/LLM context. Personal Knowledge Management techniques apply well here, though PKM leans more toward reasoning than pure storage. It pairs well with active memory management.
I believe the combination of Personality, Memory, and Reasoning written in first-person perspective is the best kind of data for LLM training.
Neuro is strongest in Personality, notably weaker in Reasoning and Memory.
Within Reasoning I place heavy emphasis on being structured, organized, logical, and factually correct. Training data needs complete, streamlined Reasoning CoT that fits those criteria. For emotional content, I'd defer to mental health professionals for what standardized practices apply, given the sensitivity.
So how is this relevant to Neuro-sama? Training data, obviously.
My idea is to structure her training data as first-person CoT Reasoning that includes personality and active memory management. Active memory requires tool calls baked directly into the reasoning data, otherwise there's no real memory.
This works significantly better training from scratch vs fine-tuning an existing model
#Neuro Personality-Reasoning & Memory
1 messages · Page 1 of 1 (latest)
Note: this is mostly for Vedal. I have further details on Reasoning specifics available. Idk what the existing memory system looks like but will likely need modification, but the personality is all you Vedal — you created her.
Fell free to reach out with questions.
Y'all should probably not leave this up too long btw or just hide it/make it private idk
Isn't this just an overexplanation of agentic AI, Chain of thought and RAG, all well known LLM concepts?
No not really there is nothing on the market and I mean nothing that does all three in the way I'm talking about
I'm talking about data level integration of
First person perspective Chain of Thought into the bones of the data
Which hasn't been done before
The type of reasoning I'm defining does not exist in the AI field as of yet there is work similar but not as broadly encompassing and or organized/structured as I'm talking about
I'm also pretty sure no AI has tools integrated as deeply as I'm talking about
Especially pertaining to having usage/active memory management stuff
It simply hasn't been done
whenever I see "simply hasn't been done" I am immediately skeptical
That is totally fair so am I
There is supporting evidence for each piece of it I just haven't seen exactly what I'm talking about done
anyways, you need to simplify it and have an actual project/plan otherwise it'd probably be left on the wayside
I did
I've spent a really long time 5 months of focused dev and 1 year of informal research
I'm not saying its revolutionary or the greatest thing ever
It still needs validation I know that
could you please make it so it's not one big wall of text then, it's really hard to follow
probably put it up on a github repo then and include a shorter, more concise explanation with a further reading link that links to the github repo
Can do