In the current era, AI chatbots like ChatGPT and Claude are capable of various tasks such as drafting emails and organizing travel. These chatbots utilize large vision-language models (VLMs), which are AI systems trained on extensive datasets comprising books, websites, code, and images. The AI algorithms are further refined using vast amounts of human feedback to ensure they follow instructions and avoid undesirable outputs, enabling them to generate text or images in response to user inputs. While there are limitations, chatbots are beneficial for numerous tasks, including areas traditionally requiring specialized skills, such as coding.
Joshua Lynch, a U.S. Air Force cadet, worked on a project under the U.S. Department of the Air Force–MIT AI Accelerator’s Phantom Program. With mentorship from Laura Niss, a technical staff member at MIT Lincoln Laboratory, Lynch explored whether he could develop a fully functional program as a coding novice. He employed a method called “vibe-coding,” using generative AI chatbots to write and refine code based solely on prompts. His objective was to enable those with military knowledge, irrespective of technical skills, to create useful software, bypassing traditional military software development constraints.
Niss explained, “The Phantom student wanted to see if he could create a useful application through self-identified vibe-coding, without any previous experience.” She aimed to observe how Lynch’s perception of AI evolved with usage and explore how nontechnical military personnel could leverage AI. Lynch sought to develop an application for his tactical team to minimize collateral damage while enhancing mission survivability, featuring AI-assisted target recognition, modular intelligence, and more.
Throughout the project, Lynch completed AI development courses and explored both military and civilian AI applications. He utilized Anthropic’s Claude, OpenAI’s ChatGPT, and Google’s Gemini for code generation, mainly relying on their chat functions rather than a development environment. The final application, created with Google AI Studio App, interfaced with the Gemini API and integrated AI in the development process.
Over three months, Lynch crafted the Remote Operating Modular Augmentation Device (ROMAD-AI). He learned techniques to improve code output, such as breaking problems into smaller parts and maintaining focus. Recognizing AI limitations and adjusting accordingly took much of the project time. As the project progressed, Lynch shifted focus from battlefield assistance to basic document processing, such as analyzing tactical maps and generating mission documents, due to AI and time constraints.
The resulting prototype, though not fully secure for its intended use, demonstrated the potential for such applications. Niss commented, “I was quite impressed with this final product, and it showed me how powerful these systems can be at prototyping designs from nonexperts.” She noted that AI could help nontechnical users communicate problems and solutions to technical experts.
During the project, Lynch’s understanding of AI capabilities evolved, leading him to scale down his initial expectations. Observations of AI systems over time revealed that Claude was more stable than ChatGPT in aspects like likeability and perceived intelligence. Lynch found AI to be a valuable tutor but noted inaccuracies in areas he knew well.
The project highlighted AI’s potential to enable nontechnical service members to create software for unique problems, serving better as a prototyping tool than a full production solution for sensitive applications. Security risks arise from improper code vetting, as seen when Lynch’s application sent input documents to a Gemini AI model instead of local parsing. Despite AI’s ability to generate functional code, code review remains a critical bottleneck.
Niss concluded, “For me, this project reinforced the expanse between experts in different fields. No matter how good AI gets, I think we’ll always need to collaborate to get to the best solutions for the most important problems.” The research was sponsored by the Department of the Air Force Artificial Intelligence Accelerator under Cooperative Agreement Number FA8750-19-2-1000.
Original Source: news.mit.edu
