The business world is focused on the potential of Artificial Intelligence as a result of Open AI's offerings such as Chat GPT. CNS is examining the duality of AI - both how it might contribute to proliferation and how it may be used to counter proliferation. As part of this examination, Ian Stewart created an experimental GPT focused on the Export Administration Regulations. This post introduces the experiment.

GPTs are a tailored version of Open AI's Chat GPT that have been given specific instructions to undertake certain tasks. GPTs can also be given access to large quantities of domain-specific data on which to base their responses. The promise of GPTs is thus that they should be able to answer domain-specific questions in a thorough way. 

Analysis of trade data can provide insight both into general trends in trade and specific transactions of concern. Trade data can be sourced at both the transactional and aggregate level. Analysis of aggregate trade data often requires custom tools such as flow maps. Under this project, tools to collect, visualise and analyse UN Comtrade data are being developed. An example of this is the trade map, a flask-based flow map for UN Comtrade data. 

The tool leverages an HS correlation table based on the WCO STCE enforcement guide to identify HS code categories of higher interest.  

This project focuses on the curation of image sets related to weapons and dual-use goods and to the training of image recognition tools to recognise these items. 

This project leverages CVAT as an image and tagging repository. The tagging set corresponds with terms in the nonproliferation knowledge graph discussed elsewhere on this site.  Models are generally trained on local on high performance consumer GPUs as the dataset is modest in size. 

In the future, it is envisioned that the trained models will be available as API services. 

Scientists love to publish their work. As a result, important insights can be gleaned from scientific publications about who is working on what, for what purpose, an with what result. Scientific publication analysis has thus long been an important tool of scientific intelligence. 

The purpose of this project is to develop tools and workflows that enable bulk classification and analysis of scientific publications to identify new insight. The project uses APIs to bulk collect scientific publication information. It also involves the training of machine learning text classifiers and the use of NLP tools to extract data from relevant publications. 

International security issues are frequently covered by news outlets. Some news articles contain new information about particularly programs, entities and facilities. As such, media monitoring has been a key tool to monitor international security and proliferation issues. 

The challenge today is how to find the needles in the news media haystack. This involves collection of a large number of news articles across all relevant languages, the categorisation of news articles, the extraction of relevant information, and the adding of this new information to what is known. 

The Nonproliferation Archive (nonproarchive.com) is not a Fusion Labs project, as such, but it was built by Dr Ian Stewart alongside the other projects described on this site. The Nonproliferation Archive epitomises the types of project Dr Stewart is undertaking. The Nonpro Archive combines the old (historical materials) with the new (machine learning and AI).

Specifically, the Nonpro Archive leverages best in class open source software (Islandora) to house historic materials. It leverages machine learning approaches including Large Language Models (LLMs) and machine learning based transcription services (Whisper) to create document summaries and audio and video transcripts. 

The premise of Fusion Labs is that the nonproliferation domain, like many other domains, can be usefully conceptualsed and examined as a data domain. By compiling all available relevant data, new connections can be identified and new insights can be gleaned. 

The nonproliferation domain is full of highly interrelated data such as networks of individuals and organisations, scientific research networks, and trade networks. Graph databases offer potentially unique capabilities for storing and leveraging highly interrelated data. This project focuses on populating graph databases with nonproliferation data and building tools and approaches for its exploitation. The project is also exploring the costs and advantages of graph database approaches compared to the knowledge graph based approach which is being explored under a separate project. 

This project focuses on application of Large Language Models (LLMs) to nonproliferation problems. This includes identification of use cases, application of LLMs to these use cases, and exploring opportunities to train or retrain LLMs to make them better suited to these use cases. 

Ian Stewart discussed LLM use cases during a CNS seminar in January 2024.