INCOSE Meeting Explores Living in a Generative World

By Shane Klestinski, Associate Editor

 

The Orlando chapter of the International Council on Systems Engineering (INCOSE) discussed generative artificial intelligence (AI) applications during its monthly meeting in Orlando, Florida, Sept. 12.

 

Judd Cheatwood, president of INCOSE’s Orlando chapter, described INCOSE’s mission as “making a better world through a systems approach.” This involves helping to develop products better to help solve problems and without negative environmental or social impacts. The content covered at INCOSE’s monthly meetings is often technical or semi-technical, and it involves the defined practices and standards of systems engineering.

 

Barclay Brown, Ph.D., associate director of AI research at Collins Aerospace, presented his content, “Generative AI and Large Language Models for Systems Engineering: An Interactive Workshop,” virtually to an in-person audience of INCOSE members in Orlando, as well as other virtual attendees.

 

Brown designed the session to be an interactive workshop on using ChatGPT and other large language models (LLM) and generative AI in systems engineering. The topics covered included prompt engineering, application architecture, and retrieval-augmented generation, among other related aspects.

 

The presentation also gave attendees the opportunity to engage with Brown’s personal AI assistant. This assistant was an LLM named “Hobson,” which shared the namesake and response style of the snarky butler character from the 1981 movie, “Arthur,” that earned John Gielgud an Oscar for best supporting actor. Throughout the evening’s presentation, attendees could get more experience interacting with an LLM by emailing Hobson questions and receiving responses.

 

While Brown’s presentation was intended for an audience of systems engineers, he noted that the information was also useful “to any kind of knowledge worker,” to include general applications to any kind of engineering work.

 

He began by discussing the modern state of generative AI, and he displayed examples of his own experimentation with LLMs that highlighted the future potential of AI, as well as the current restrictions involved in using the technology.

 

Brown noted that it required considerable effort to get generative AI to correctly spell “living in a generative world” in the image above, but it still misspelled “box office,” which indicated an AI limitation at the time he generated the image. He also played the first half a song, “Living in a Generative World,” which he generated with ChatGPT that called on its understanding of popular trends in current music. The song featured the possibilities of using customized songs for presentations at conferences or other gatherings. Click here to hear the song. (The AI-generated lead vocal sounds similar to Taylor Swift.)

 

“The purpose of these experiments is to demonstrate and teach people how to do things,” Brown said. “This program generated this entire song by me giving it about 20 words of description of what I wanted the song to be about, and the style I chose, ‘electronic pop.’ It did everything.”

 

Many tech professionals who have worked with generative AI are familiar with “hallucinations,” which can manifest as minor factual inaccuracies, significantly misleading information (to include citations of nonexistent sources), unrealistic images, or unreasonable conclusions. Brown emphasized that while AI is a powerful and entertaining tool, it still needs human supervision.

 

In an early draft of the presentation’s opening pop song, the AI originally wrote the line, “These AIs are my cooks,” as “This AI is my landscooks.” At first, Brown was impressed by the program’s vocabulary that used a word so obscure that he had to look it up – until he discovered that “landscooks” is not a word. The final version of the line that made sense and maintained the song’s rhyme and lyrical cadence, required an edit from Brown, which highlighted the collaborative nature of working with AI.

 

“I think we’re going to use ‘generate’ as a verb more and more in coming years,” Brown said. “It’s not going to be, ‘I wrote this paper,’ it’s going to be ‘I generated this paper’ – and what does that mean? It means ‘it’s my work and I had a partner in this AI, and we did it together.’”

 

In a generative world, according to Brown, people “take the place of a conductor, instead of a composer.” More specifically, people are the conductors of an orchestra of “trained animals” that can all do special things and have special abilities, and humans coordinate their work. Brown described the overall process of generating content with AI as “directing to create.”

 

“The kind of skills you need for a world like this are different from the past,” Brown said. “If you want to generate in this generative world, you have to imagine stuff specifically – more specifically than you’re used to. We have to imagine clearly… and then describe it clearly. I find that people who have great skill with language and great vocabularies, who know a lot of adjectives, do a lot better in generating stuff.”

 

He explained that if users are generating written content using AI, they must instruct the AI on how to write it. Users who don’t know the difference between writing styles that are “breezy,” “academic” or “authoritative” – among several other adjectives that describe writing – will have difficulty obtaining the products they want. However, if they can accurately describe their ideas, they are more likely to get the desired outcomes.

 

One specific takeaway of Brown’s presentation included the advice: “Your prompts are too short.” According to Brown, prompts often don’t describe enough when users are trying to get a sufficiently usable product from generative AI.

 

Brown said that assigning certain personas can be helpful when prompting an LLM to describe the operation of, for example, the Saturn V rocket to different audiences for different purposes – in so far as what an LLM has been “pre-trained” to know. A prompter would get a much different result if they instructed an AI to generate a presentation of the rocket’s operation as explained by a senior NASA engineer, compared to a kindergarten teacher.

 

“This is the domain of prompt engineering,” Brown said. “People thought prompt engineering would be a new profession a few years ago, [but] I don’t think it’s going to be a new profession. I think it’s going to be a skill… and we’ll see more tools built around large language models that can do things for [users].”

 

This “new skill” he outlined – of users creating things in their minds and describing them well enough for an AI to generate them – can also be seen in generating software. Brown said that if he can describe a program, the AI can generate it, but he will notice that the AI may do so poorly. It might produce “weirdly structured code, which might run, but it’s not that good,” because the code can be difficult to maintain, understand, and it may introduce bugs if he doesn’t guide it enough to sufficiently describe the software architecture.

 

Professionals in the modeling, simulation and training (MS&T) community have also noticed the increased popularity of LLMs and AI-generated content in recent years, and they have started exploring the possibilities of such applications to the MS&T field. Since MS&T professionals who develop training – specifically for defense purposes – have experimented with such applications, it’s reasonable to assume that hostile nations have done so as well. When Team Orlando News later asked Brown (after his presentation) about possible national security threats from a near-peer adversary gaining an edge in AI exploration, he cited developments he’s observed in the civilian AI market.

 

“If the USA falls behind that would be a problem,” Brown said. “I don’t think we are behind right now given the leadership positions of OpenAI, Anthropic, Microsoft and NVIDIA, but anything can happen in the future.”

 

When one thinks of the MS&T products used in training scenarios, often flight simulators, virtual-reality headsets, or marksmanship simulators will come to mind. LLM-generated content may also be eventually capable of presenting in-depth scenarios for training simulations that many not have yet imagined.

 

Brown cited a “Star Trek”-themed experiment he conducted two years ago. Taking the role of a U.S.S. Enterprise captain, he directed ChatGPT to respond like the computer from the iconic spaceship, with all the knowledge of the various “Enterprises” from all the different “Star Trek” TV shows. He also instructed it to provide scenarios consistent with the show’s episodes, and to respond as the franchise’s crew members would if Brown mentioned them by name.

 

He said his description worked “amazingly well.” It resulted in an open-ended scenario that ChatGPT generated, wherein it presented a simulation involving the Enterprise investigating an anomaly in space. After Brown ordered the characters Data and Geordi La Forge to scan the anomaly, which they did unsuccessfully, he had Lieutenant Worf hale the anomaly itself. “Worf” told Brown that the anomaly was – unexpectedly – responding. In a plot twist in the scenario, “the anomaly” turned out to be aliens who were also explorers, and they were also investigating an anomaly in the same region of space.

 

“This scenario kept me engaged for weeks, and once in a while I come back to it,” Brown said. “As a world simulator, these things are going to be amazing.”

 

Although AIs can give responses that can make them seem sentient or “alive,” Brown emphasized that “they have no soul, and no goals,” but paradoxically, they also seem “eager to please.” He explained that most hallucinations occur when an LLM is trying to give humans what they ask for, but it does not understand the request. This phenomenon is indicative of the iterative nature of collaborating with AI, especially when it does not fulfill a request correctly on its first attempt (or after multiple, subsequent attempts) and the desired outcome requires human judgment, for example, in fixing the nonword, “landscooks.”

 

“These new beings are happy to redo things many times,” Brown said. “Sometimes it takes many times to get what you want. It’s really different from working with a human because I’ll ask it to redo the whole thing each time [based on] one little bit of input.”

 

Cheatwood said that his most valuable takeaways from the evening involved the tips emphasizing different techniques that work with the iterative nature of collaborating with generative AI, and providing additional clarification and guidance.

 

“I hadn’t thought much about the art and science of doing prompt engineering in the context of using large language model-type applications like CHATGPT or GPT-4o and others,” Cheatwood said after the INCOSE meeting. “I’ve thought about how I could use them to do generative things to make my job easier as a model-based systems engineer, but I never thought about the patterns I should be using to prompt [using those tools]. Now, I feel like I’m better equipped to do some more research and start practicing [those aspects] better.”