Tech & AI Journey

📊 From Output to Insight: The Value of Preservation

Having discovered that with some careful prompt engineering even chat interfaces could yield staggering insight and value (sometimes!), the digital prepper in me kicked into overdrive. I began thinking about how these "outputs" (completions) could be preserved and organised so they don't vanish into the digital ether (output storage remains, sadly, a much neglected topic!).

Building my own modest CRUD interface to store these outputs ("Output Hub"!) then led to the second realisation¹: while numbers and statistical models are what's happening under the hood, text is the layer at which humans provide information to LLMs and get information back from them. Curating it is key to leveraging it to maximal effect.

⛏️ Contextual Data Mining: Building a Personal "Context Repo"

An unintended benefit of carefully curating outputs was discovering that those same outputs could be mined for contextual data and then reintroduced in subsequent interactions. My boring but demonstrative example is a five-line markdown file laying out the key details of my OS and hardware, which can single-handedly do everything from making highly useful recommendations for local models to run to helping me pick out a hardware upgrade without tedious component checking.

I developed an interest in deliberately and selectively curating a repository of personal information (like a few others toying around with the idea, I naturally gravitated to the language of "context repo").

🔍 The Deep Challenge of Contextual AI

Imagining how to develop a "context repo" is a work in progress and by no means easy. Inevitably, it leads to rather stark realisations about life: without contextual awareness and memory, human life would be meaningless!

Our very cognition is a byproduct of the passage of innumerable memories, thoughts, and feelings. Without the context of our lives determining the value of information to us, the sea of information surrounding us is irrelevant (because it is bereft of meaning).

Our life's "context," moreover, is in a constant state of flux. In the drab terms of database systems, it might be described as an immensely nodal network of data with a constant level of file operations that defy the type of technical architectures we have at our disposal today. So, I think it's no exaggeration to say that the technical challenge of reflecting this in the type of AI systems needed to develop transformative and personal experiences is significant, worthwhile, and hugely important to how humans will best harness the technology.

🧪 Experiments in Contextual Data Generation: Multi-Agent Systems

On a more practical level:

One of my experiments-in-progress involves a multi-agent system for proactively developing contextual data. The mechanics are simple: one agent interviews you, another tries to distill it to the contextual clues, and finally, it's factorized in a context database. I don't view this as the ultimate solution or direction that personalized AI systems will take, but it is proving an interesting experiment in how one might kick-start a personal RAG system of sorts.

✨ By Abstracting The Mundane Details Of Technology, AI Allows Creativity To Flourish

My favorite part of exploring AI at the moment, however, is actually the perhaps unloved art of writing system prompts: the configurations that take the final product of the LLM produced by the vendor and imbue it with subtle or definitive direction, personality, and nuance.

System prompt writing exemplifies how generative AI tools have provided creatives with access to technology in a way that is transformative and wonderful. It allows us to think beyond the mechanics of orchestrating API calls, storing context, and choosing database types. It allows our minds to think freely about how humans can leverage intelligent computers to improve not only their lives, but the world around them and their relationships with others. I've written a collection of over 700 system prompts to date, of which more than 600 are open-sourced (here, on Hugging Face, and other nooks and crannies of the internet).

🔮 My Long-Term Vision

I'm motivated by the idea and conviction that beyond the hype surrounding the latest and greatest models with reasoning powers that may lack current applicability is a less glamorous but longer-term transformation.

This transformation will flourish not over individual release cycles or months but over the course of decades through the collective efforts of technologists, ethicists, creatives, thinkers, and everybody else who believes that AI is a change worth nurturing. If I can play even a microscopic role in that direction, it will be a positive contribution.