STO-Activities: (no title)
Activity title:
Military Applications for Large Language Models
Activity Reference:
IST-207
Panel:
IST
Security Classification:
NATO UNCLASSIFIED
Status:
Active
Activity type:
RTG
Start date:
2024-01-09T00:00:00Z
Actual End date:
2027-01-09T00:00:00Z
Keywords:
AI, ChatGPT, Generative PreTrained Transformer, GPT4, Large MultimodalMultiAgent Language Models
Background:
The development of AI-based systems is progressing at an ever-increasing pace. The success of ChatGPT and GPT-4 shows in an impressive way what large (multimodal/multi-agent) language models are capable of, both for searching information and for authoring. They can answer questions and make discussions, design websites based on sketches and write/refactor program code, analyze images and videos, solve mathematical tasks, and write summaries of the contents of books. They also exhibit some degree of creativity, as they write poems and short stories, create pictures based on textual description etc. The societal impact of these new AI-enabled capabilities cannot be fully assessed, yet. With respect to the defense sector, they impose advantages and risks. Therefore, it is necessary to continuously monitor the rapidly advancing developments and examine use cases both from NATO's perspective and from the perspective of a potential adversary. The development of very large models is currently being driven by a few companies in the civilian sector, which are investing billions of dollars. Most research organizations cannot afford such an investment. Therefore, it is also necessary to investigate which models are available and how they can be used safely by NATO members for military purposes.
Objectives:
The group seeks to get a better understanding of the technical approaches, capabilities, and limitations of large (multimodal/multi-agent) language models over the course of the next three years. The technical advancement progresses at high speed (outside the defense sector) and this group is meant to follow the most recent developments and share insights among the members.
The group will identify military use cases of large language models both from NATO’s and the enemy’s perspective. For each use case, the maturity of the available language models is going to be assessed.
The group will also investigate the availability and accessibility of large language models both in the civilian and military world and how they can be technically addressed via APIs. To achieve the best results, proper prompt writing is essential. The group will analyze the impact of prompt writing and investigate best practices.
Topics:
• Language Model Technology: The state-of-the-art approaches to build and use large language models will be analyzed to the extent possible. (Note: The underlying architecture of these models is usually shared by the companies but not the training data)
• Identification of Military Use Cases: The group will continuously perform a survey of civilian use cases and analyze whether and how they can be transferred to the military domain. Moreover, the group will collect military-specific use cases.
• Dealing with Hallucination: One of the key problems of large language models is that they tend to “hallucinate”, i.e., the result is not always fact-based. The group will look at how to cope with this phenomenon from an application/end user perspective.
• Bias: The output of such large language models may be biased (also with respect to political attitude). The group will analyze how this affects potential military applications.
• Assessment of AI-based models: The group will develop methods to assess the capabilities of AI models with respect to selected use cases. This allows to evaluate the capabilities of models over time and to compare different models.
Contact:
Open2Partners:
Title:
Created at 10/07/2023 16:00 by System Account
Last modified at 16/05/2024 15:00 by System Account
Go back to list
Home(NATO STO)