Introducing Gluon

With the recent advances in LLM capabilities, opportunities for LLM applications in intellectual domain has arisen. My first exposure to LLMs was back in March 2023 with the introduction of GPT-4. Initially enamored with its performance, I started to see flaws in GPT-4’s abilities, especially as it was tasked with handling broader, more complex tasks requiring heavy reasoning. There is no doubt in my mind that LLM models will continue to advance, completing tasks that are out-of-reach today; however, in the short-to-medium term, current capability is sufficient for these advanced tasks given an appropriate framework for user-GPT interaction. To that end, I started development of Gluon, an LLM (currently GPT) interface designed for engineering and pedagogical use, in May of 2023.

An animated example of a real Gluon exchange. Gluon stores the query payload and GPT's raw response as well as human-readable transcripts.

To summarize, Gluon is

  1. an LLM user-interface, allowing full interaction with control over supported parameters. Not only does it handle standard features such as streaming and image inputs, it also has added features not commonly available in a LLM client such as sharable and restartable transcripts system.
  2. a framework for integrating LLM capabilities into the local user environment: allowing a the backend to directly interact with user file system (and hence local functions) and programming environments (interpreters).
  3. a director of multi-agent configurations: each ‘agent’ is equipped with their own set of functions, environment, and instructions. This distributed system enables query delegation; thus each agent can be specialized, compact, and robust.

In an higher-education context, 2. enables interaction with technical software which is made accessible through the natural language interactions: valuable time is not spent on learning ‘vocational’ skills, but on considering essential questions such as “how can I determine if the numerical solution is accurate?”. Furthermore, software environment such as MATLAB™ — not generally available, but is heavily used in industry and academia for its advanced capabilities — can be ‘glued’ together with GPT, if there is a local installation. Meanwhile 3. allows generation of a team of AI ‘agents’ with a much broader skillset than the vanilla pre-trained model. This multi-agent framework enables construction of a reliable system suitable for engineering and classroom applications, where low tolerance of errors has previously prohibited complex LLM configurations.

I’m intent on continuing development of Gluon and currently looking for wider feedback on its potential use cases. If you are interested in using this software, please contact me. A live demonstration may be prepared on an individual basis.