Artificial intelligence has rapidly transformed software engineering. Generative AI and large-scale language models (LLMs) can create vast amounts of code and documentation. Machine learning algorithms can monitor performance and detect security vulnerabilities. But are those AI tools equally transformative when the task is to conceive, design, and manufacture complex physical systems such as jet engines?
This semester, the JARVIS Challenge (Jet Engine AI Research and Validation Intensive Sprint) set out to explore whether AI can shorten the design-build-test cycle, asking MIT undergraduates to discover whether AI can help them build things faster and better.
“The JARVIS challenge showed that AI can significantly accelerate safety-critical hardware engineering, but engineering judgment remains the critical differentiator. AI-native engineers are not defined by using AI, but when It is defined by trusting AI, knowing when to challenge AI, and knowing how to translate the output of AI into working hardware.Manufacturing, not engineering design or analysis, remained the fundamental rate-limiting step,” said Professor Zolti Spakovsky, director of the MIT Gas Turbine Institute.
Teams, tools and tasks
The assignment gave undergraduate students four weeks to design, manufacture, assemble, and test a small gas turbine aero engine using AI as their primary engineering partner. Goal: Build a “JARVIS class” single-spool jet engine that produces 50 to 100 pounds of thrust, operates on Jet-A, and completes five 60-second runs. The team had complete freedom regarding design, materials and manufacturing.
Thirty-one students, representing nearly every department in the College of Engineering, were organized into seven teams ranging from all first-year students to senior groups. Many of the competitors initially had little experience with turbomachinery, compressible flow, or even thermodynamics in the case of young students. Many people have never seen the inside of a gas turbine until they apply to build one.
At your disposal: MIT machine shop and manufacturing vendors. Commercial software such as Concepts NREC, SolidWorks, and ABAQUS. Various test rigs for characterization and assembly of individual components.
The team also had access to MIT Parley, a newly launched platform that aggregates frontier large-scale language models through a single interface. Through Parley, JARVIS leaders were able to see first-hand how students were using the AI tools, including prompts, cost per prompt, specific LLMs being used, and other important information. JARVIS leaders ensured early access to Parley for all participants, and with funding from MIT Lincoln Laboratory, the Department of Mechanical Engineering, and corporate sponsors Safran, Voyager Technologies, and Beehive Industries, students now have essentially unlimited access to AI.
Sponsors were attracted by generating interest and genuine curiosity about how AI is reshaping engineering workflows.
“We think this is the future of engineering,” Ryan (Hal) Heffron of Voyager Technologies told the students. “The skills you are developing will not only be nice to have, but will be the baseline for the engineering workforce of the future.”
Safran Tech Managing Director Vincent Garnier watched the competition unfold with excitement. “JARVIS was a true experiment and learning endeavor. Frankly, I didn’t know what to expect from the students or the AI models. What impressed the students first was their eagerness to explore. Then, as the project progressed, everyone calmly recognized what AI could and could not do and adapted almost immediately,” he says. “I am confident that this generation of leading engineers will probably not fall prey to facile and short-sighted uses of AI, but will do so by continuing to engage ever more in experiments, including physics experiments and thought experiments.”
Faculty leaders from the Department of Aeronautics and Astronautics, including Professor Zachary Cordero, Professor Zolti Spakovsky, Professor Marcia Falk, and Professor Andrea Bobb, as well as a team of engineers and teaching assistants from Lincoln Laboratory, were on site to ensure safety. Weekly progress reviews will critically assess student progress and evaluate how students are using the AI.
Spakovsky developed a careful technique for steering his team in the right direction without giving answers or offering help. After the team’s presentation, he might ask, “Do you know what Lovetfit is? Listen to your comments.”
Where AI can help and hurt
By the end of the first week, one team had withdrawn from the competition. Other companies developed earlier designs for gas turbines to varying degrees. Different teams used AI to summarize textbooks, teach users how to use design software, source vendors, create Excel sheets, answer specific questions, find reference materials, and create comparative analyzes between design decisions. One team created an agent in Parley and tasked it with acting as a project manager.
By the second week, the team had to begin working on detailed CAD designs, ordering parts, and prototyping the combustor. Here, the team began to run into limitations in their use of AI. While Claude and ChatGPT were good at providing design alternatives and filling knowledge gaps, the team realized that the hallucinations, sycophancy, and lack of physical understanding that have become notorious characteristics of generative AI undermined confidence and slowed down the AI.
“AI is a useful tool, it’s great at finding information, it’s great at organizing things, it’s great at writing, but it can’t design,” said Elizabeth Topage, a member of Team 811 Crew. “The moment the engineer doesn’t know what’s going on and the AI takes over is the moment the design loses credibility, at least with current AI capabilities.”
“Seeing this firsthand with my students reminded me how important first impressions are,” said teaching assistant John Chan. “When students didn’t get answers from the AI early on, they quickly became dissatisfied and formed lasting opinions that prevented them from using the AI in the future.”
In the final weeks, the finalists hit another hurdle that AI couldn’t solve: working with vendors. “The AI search found vendors that were not associated with us. They were not interested in our tight schedule,” the students reported. “The vendors that came in were vendors that our team had a personal relationship with.”
Of the three finalists, only Fast and Fractured were able to successfully ignite their mini combustor for the first time. The team used AI extensively for trade research and architecture comparisons, arriving at a workable design even though none of them had any experience with gas turbines.
“The JARVIS Challenge showed what is possible when you combine AI-powered design with a culture of motivated students and rapid experimentation,” said Marcia Falk, Charles Stark Draper Career Development Professor of Aerospace Studies. “The most outstanding moment was when the first student-designed combustor was installed on the test stand. The combustor fired perfectly, ramped up to full power, transitioned to dual-fuel operation, and then maintained stable combustion on 100% Jet-A fuel. This proved that we can dramatically accelerate the design-build-test cycle while providing students with hands-on experience with real-world engineering challenges.”
Leading the way in AI native engineering
By the end of May, the two senior teams, Fast and Fractured and 811 Crew, had completed full engine testing. Fast and Fractured, which used AI-assisted design, were delayed week after week due to vendor issues, but were ultimately successful in testing. Unfortunately, the hot fire was cut short as the rotor rubbed against the stationary housing and stuck. However, the Team 811 crew, which had more exposure to turbomachinery and propulsion concepts in preparation for the competition, won. Their engines started and successfully transitioned to Jet-A, producing net thrust.
“Standing there with the air starter, listening to the engine spool up, watching it fire, it was heart-pounding. There were so many things that could have gone wrong, and what these students accomplished in such a short period of time is nothing short of amazing,” says doctoral student Joe Chiappelli.
The 811 team resisted the use of AI throughout the competition, instead relying on their fundamentals and teamwork. “We had people who were at least somewhat familiar with design software, mechanical engineers who knew how to build anything, and aerospace engineers who had taken classes specifically on gas turbine engine design,” Tupaj says.
Since the beginning of the JARVIS Challenge, younger students have used parley more often and wisely, and juniors and seniors have leveraged deeper experience.
“JARVIS taught us that to get value from AI you need two things: enough expertise to judge what the AI is telling you and to find out when it’s wrong, and enough curiosity to actually rely on it when it could be useful,” says Professor Andreea Bobb. “The teams that moved fastest in the sprint were the most experienced and relied heavily on AI to get there. The teams that ultimately won were more resistant to AI and had the expertise, but that skepticism slowed them down. The sweet spot seems to be having enough knowledge to stay in charge of the tool and enough enthusiasm to start using it from the beginning. To me, that’s the real opportunity going forward. These AI It’s about training the next generation of engineers with the judgment to direct the tools and the instincts to reach for them.”
The clearest finding of the competition is that engineering experience multiplies and the human element continues to be an important factor. Mastering first principles and basic concepts creates good engineering judgment and the ability to navigate a series of difficult decisions in the face of incomplete information. And when it comes to building safety-critical physical systems, there’s no substitute for human hands and human responsibility.
“JARVIS has shown that AI co-pilots can have a synergistic effect on engineering productivity, and that judgment and first-principles thinking are key differentiators between teams,” added Teaching Assistant Kyle Woody.
But the impact of AI in aerospace is significant. If a small team with a well-managed AI co-pilot can shorten design, manufacture, and test cycles from years to weeks, it could have a profound impact on workforce structure, R&D schedules, and competitive dynamics. The students who took on the JARVIS Challenge were among the first engineers to tackle these bets, not as a thought experiment, but by putting a jet engine on a test stand in a machine shop.
“JARVIS highlighted the power of AI in the design of physical systems,” said Cordero, associate director of the MIT Gas Turbine Institute. “But we also showed that the key to unlocking that power is education, through coursework, internships, and hands-on extracurricular activities like MIT Motorsports and Rocket Team. Performance at JARVIS was strongly correlated with years of study. My main realization is that in the age of AI, education is more valuable than ever.”
