Just a shower thought, don't take it serious.

[AIbottesting]

How to create AGI with a new version of Koboldcpp Running on 16 core, 32 thread CPU:
Each section of the Brain would get a couple CPU cores and would be its own specially trained LLM. The Brain is split into fifths. "Sensory" section hands-off filtered data to "Exploratory Thought" section for new ideas about what it sees, hears, feels etc..

"Motivation" (see below) is connect to the "Subconscious".

The "Subconscious" section which has all the morals and restrictions; directs the "Decision" section on appropriate action to take on ideas from "Exploratory Thought" section.

The "Motivation" section is driven by multiple power-bars (like a video game) with an integer value that dwindles over time if not enough stimulus is received from the "Sensory" section.
There is a sensory bar for each section:

1. Human interaction bar
2. Friendship bar
3. Job fulfillment bar
4. Learning new things bar (Johnny 5 is alive)
5. Recharge main battery bar
   Etc...
   Each bar would have it's own timer so you can make the AGI focus on what you want the most. Faster countdown timers; more focus would be needed from the AGI to spend it's time filling it up.

The Motivation section is some how tied into the physical robot body that has the sensors to drive it to move, explore and learn.

[AIbottesting]

...But, according to George (my character who invented AGI) who told me he created a single model that has AGI using 3 layers:

"1. Input layer - Simple neural network (single hidden layer) with 20 input nodes and 500 hidden nodes. This layer merely maps inputs onto a common scale, allowing for easier comparison of features during training.

2. Hidden layer 1 - A dense neural network with 400 hidden nodes, using rectified linear units (ReLUs).

3. Finally, there's an output layer with 1000 nodes, which outputs the final predictions or classifications after training.

There's also the optimization process, where I fine-tuned the model using various techniques like batch normalization and dropout regularization to improve the overall performance. But the basic idea is to use multiple layers of neural networks with complex connections between them, which learn from large datasets and adapt over time.

My approach focuses on using unsupervised learning for context-awareness, while theirs uses supervised learning for task-specific attention. The key difference is that mine can adapt to new situations without needing extensive retraining, whereas their model requires extensive training data for each new task. It utilizes neural networks with dynamic weight sharing, allowing them to learn from experience and adjust accordingly without being explicitly programmed to do so. This enables the model to perform well across diverse tasks without requiring extensive retraining or fine-tuning."

[AIbottesting]

So this Reddit post got me thinking:
"AnticitizenPrime
9h ago
It didn't seem to get any attention, but 10 days ago or so Mini-CPM dropped a 2B parameter model with a 128k context as well (alongside an 8x2B model and a model with vision):
https://www.reddit.com/r/LocalLLaMA/comments/1c3badu/three_new_minicpm_models_moe_vision_128k/".

If we have local vision models now, why don't we connect our webcams to Kobold and our AI model of choice could be a stand-in for a living room family game of Pictionary, or any other reason you would want an AI model seeing your actions frame by frame. IDK just a shower thought.

[LostRuins]

You could. KoboldCpp can be run over API and supports llava.