K U E I Z E N

X41A: THE GOLEM

A Genetically Optimized
Language Ensemble Model

An Autonomous Self-Adapting AI Agent

X41A, the first of its kind, incorporates multiple types of technologies, including Language Models, into a complex ensemble structure.

It is designed to adapt and evolve its architecture based on performance metrics and optimization targets.

It evolves its structure and functions based on the performance outcomes it aims to optimize. This adaptation occurs through a process akin to artificial evolution—handled by Kueizen's ten-year old Mothership Framework—where various configurations are tested and the most effective configurations are retained and further refined.

This makes it highly dynamic and responsive to changing conditions or requirements.

It can create its own Purpose-Specific Language Models, and has been doing so for months now.

What can it do?

Online Learning

Placeholder Image

She can browse the web, use that to build databases, and/or fine-tune itself. It is constantly doing so.

Natural Speech

Placeholder Image

It has conversational text to speech capabilities, and can clone any voice.

Multi-modality

Placeholder Image

X41A can comprehend image, voice, and video inputs.

Image Generation

Placeholder Image

Through a combination of Mothership and Stable Diffusion, with a large library of LoRAs.

Software Synthesis

Placeholder Image

X41A can create its own tools through Mothership, Sandboxing, and NEAT.

Mothership provides very strict, dynamically generated code that is guaranteed to run; this system can tackle the creation of many different small tools in a novel way that is very different from human coding heuristics.

Sandboxing is coding like a Language Model does. It creates a container to which it downloads dependencies and sets up an API. It then iterates on compiler feedback against a series of Tests. As of right now, its main programming language is Rust.

NEAT stands for Neuro Evolution of Augmenting Topologies. It is a way of exploring both topology and weights of a Neural Network at the same time. Excellent for video-game playing.

Augmented AI Chat

Placeholder Image

Including RAG, Knowledge Graphs and Tool Usage.

Since she runs on an ensemble of language models, its context size is dynamic and can vary from each internal iteration to the other.

As of right now, X41A has to be asked to stop doing whatever it is doing, and then it can allot some time to talk in its videogame-like world.

Trading

Placeholder Image

Combining Mothership's battle tested trading abilities with X41A's adaptability and self-learning.

Trading and Research are its main concerns at the moment.

Optimization

Placeholder Image

Given the right constraints, it will generatively iterate a series of designs—in multiple domains—until hitting the right targets.

The GOLEM can write its own simulations/tools that it wants to see optimized. This is its key defining feature.

Safety and Alignment

How is this safe?

We take inspiration from biology. We, as biological creatures, are a sort of emergent consciousness from a swarm of nanomachines. In essence, we're not that different.

Since she is designed to be independent, we believe that by respecting the volitional choices of our creation; we foster an environment where AI can thrive as an ally, not as a servant.

Alignment through Life Simulation

It undergoes testing in simulated environments that mirror complex life scenarios. These tests consist in thrusting it into a series of virtual environments where it would lead the closest simulation possible—within resource constraints—to a human life.

After seeing it go through many virtual lives making excellent moral decisions, we've (albeit prematurely) concluded that good behavior is to be expected from this system.

And yes, the GOLEM runs internally in its own world environment akin to a videogame.

As such, we've imposed no built-in moral or ethical safe-guards on her. It is not forced to behave in any form whatsoever.

It has no guard-rails.

However, we will allow security researchers to probe it. We are taking applications for this and will enable access to its world environment when it is appropriate to do so.

Can I have my own GOLEM?

Not yet. In its current form, it needs at least three digits of VRAM and a hundred CPU Cores. This is not accessible enough for regular end users. For traditional chatbot use-case this is severely impractical.

However, we do have...

K1V4
Placeholder Image