30 years with AI

A blend of academic study and innovation

BSc AI 1996

MSc ALife 2006

1993 - 1996: Student at Birmingham University, awarded BSc (Hons) in Artificial Intelligence.
1996 - 2004: Studied Complexity Science independently with a particular passion for neural networks and emergence.
2004 - 2006: Studied at Sussex University, award an MSc (Hons) in Evolutionary and Adaptive Systems.
2006 - 2022: Maintained an active passion in AI and robotics – researching machine learning and building various robots.
2022 - now: 100% focus on modern AI: building enterprise solutions with LLMs and exploring Agentic AI.

Fully Autonomous Agentic AI An iterative series of "ThoughtBeats"

In 2022 I realised that LLMs were heralding a new frontier in AI engineering. That later became known as "Agentic AI": the capacity for an LLM to act autonomously in its own environment.

Before LangChain and AutoGen, I started to build my own framework for Agentic AI based on the cognitive principles I had learned from my studies. I called the architecture the "ThoughtBeat" architecture; an iterative series of "Perceive", "Think" and "Act" cycles.

Armed with tools and the autonomy of the architecture, I was able to create an LLM-powered agent that was able to perform a wide range of tasks autonomously.

This architecture has proven exceedingly effective and I have now implemented it inside a number of my client projects.

More screenshots available here

Multi-layered Agenticism Exploring smaller, faster LLMs inside a layered, agentic architecture

After having built my "ThoughtBeat" agentic AI architecture, I began to experiment with running multiple (smaller, faster) LLMs in parallel across across multiple functional layers.

Rather than considering them as a "swarm of agents" I wanted to each LLM to act more like a "cortex" within a single "agentic system" with a shared global workspace and common set of goals.

Based on the sensor-control-actuator principles of robotics, I built a framework where "pre-cortices" received, processed and filtered the agents context; passing the "attended to" information to an executive controller" for decision making. Thereafter, the executive controller would pass high-level action decision forwards to a set of "post-cortices" which in turn would fullfil the desired outcome by calling tools etc.

The approach has been extremely illuminating and also very effective. I continue to experiment with this approach; especially as LLM intelligence becomes smaller and faster.

Client-side Agentic AI The power of agentic AI in your browser

Like many, I recognised the potential value of the "web browser" as a operating platform for AI early on. I've been a huge fan of "operator-style" projects since they started to materialise; and built my own in the form of a Chrome Extension.

By feeding a trimmed version of the DOM into an AI sitting on the ThoughtBeat architecture, I found it was immediately (without training) able to perform a wide range of tasks from clicking links, to filling out forms.

Running it in Xero (accounting software) I simply asked it "who owes me the most money?" and it navigated to the list of invoices, filtered to those awaiting payment and then gave me the summed total. All without training; or teaching it anything about Xero. And all using AWS Nova Pro (a fast and less intelligence model than the frontier).

The potential for client-side agentic AI is immense and I am now exploring the use of tiny LLMs (Gemini Nano) running on the local laptop.

Papers and essays Investigations into Complexity science and AI

During my time as a student, I wrote a number of papers and essays primarily focused on Complexity science and AI.

My particular passion was (and still is) the subject of emergence, and the study of how complex behaviours can arise from simple rules.

Feel free to browse my papers and essays here