Nimchimpski Server

Welcome.

This is my self-hosted Ubuntu Server — equal parts creative lab and production-grade workshop - here I prototype machine-learning tools that prioritize innovation and robustness.

Early interest in ethical questions took me first to a degree in Philosophy and Psychology at King’s College London; later Computer-Science research left me focused on algorithmic bias, alignment, and the subtle ways human assumptions seep into machine logic.

As a sideline I'm developing a theory that any cognitive architecture in AI that integrates instrumental intelligence with symbolic intuition would exceed any previous ideological system in scope and subtlety.

Ive also been known to produce experimental Electronic music (as Noderandori and Circog), publish travel articles, work as TV editor and producer and supply editorial and advertising photography, (oneclearshot, MarineVisual) But my real passion lies in exploring the intersection of technology and human experience, learning form that, and building things that do some good.

ML Sythetic Aperture Radar analysis for disaster response
Still working on getting a demo version of this here...a bit more server RAM/CPU/GPU will help :]
Reinforcement Learning
Othello Game Engine. An implementation of the Q-learning algorithm, which is a type of reinforcement learning. The game engine allows you to play Othello against an AI that learns from its experiences. The AI uses a Q-table to store the values of state-action pairs, updating them based on the rewards it receives after each move. The more it plays, the better it gets at making optimal moves.
At the heart is the wonderful Bellman equation:
Q(s,a)←Q(s,a)+α(r+γmaxa′Q(s′,a′)−Q(s,a))
This version is pure self learnt via about 10000 games of self-play (on a laptop), with no pre-programmed heuristics. Play the game
Cooperative LLM network with Retrieval Augmented Memory
This project explores the concept of a cooperative network of Large Language Models (LLMs) enhanced with Retrieval Augmented Memory (RAM). Each avatar in the network has its own memory, which can store personalized information and experiences. When a user interacts with an avatar, the system determines whether the response can be generated from the avatar's memory or if it requires assistance from another avatar in the network. This creates a dynamic and collaborative environment where avatars can share knowledge and provide more accurate and context-aware responses.

Key features include:
- Memory Retrieval: Each avatar can store and retrieve information from its own memory, allowing for personalized interactions.
- Cooperative Knowledge Sharing: Avatars can request help from others in the network, enabling a distributed and collaborative approach to problem-solving.
- Customizable Personalities: Users can customize their avatars to reflect their own personality, preferences, and knowledge domains.
- Dynamic Decision-Making: A decision function determines whether to retrieve information locally or request assistance from another avatar.
Video demonstration of the project.
All project code is at my github page

Some of my photo and video work (under the monikers oneclearshot and MarineVisual)...
photography and electronic music

video showreel
more vimeo
Gsnowboarding snowboard instructor website

If you like music and sounds at the experimental end of the spectrum, you can check out (or buy one of!) my projects on Bandcamp
some code video demonstrations here

ARCHIVE

The Minimax Algorithm

This simple but powerful algorithm is a foundational technique in decision-making for games like chess, tic-tac-toe, and Go. It’s brilliant because it enables an AI to anticipate an opponent's best moves and respond optimally, assuming both players play perfectly. The smple simple powerful recursive approach formalizes strategic thinking — maximizing your Utility while minimizing your opponent's.
However:
Its computationally expensive: Its brute-force nature makes it impractical for games with massive move trees unless heavily optimized (e.g. with alpha-beta pruning).
Assumes perfect play: Real opponents might not always act optimally, making the algorithm sometimes overcautious.
Doesn't learn: Minimax doesn't adapt or improve from experience — it's static and rule-based.

Thanks for visiting!

contact alexwebb at pm.me