There are the facts, and there is the story. Both can be true, but one without the other is incomplete, as it is the story that carries meaning. As the AI ChatGPT consumes my social media feeds, it’s enough to make a blogger worry. One day soon, will an AI put the words and the melodies together and write the songs – and the stories too? As Liam Neeson said, “what I do have are a very particular set of skills, skills I have acquired over a very long career.” This results from experiences, not only with projects but with people. Circumstances, contexts, and people add meaning to my knowledge, thoughts, or misconceptions. We are our stories. Among these experiences is trying to do what an AI might do one day, using massive amounts of information to illuminate the road ahead.
My story today begins back when an AI fit on a floppy disk. For those unfamiliar with the term, these were little plastic disks (I’m showing my age). Think vinyl for your retro record player from the fad resurrecting LPs. (OK, vinyl is not dead, and it’s still a thing everyone does with their latest album, from “Best of” oldies to Dua Lipa. But I digress, which is very human. Will an AI ramble one day? Come back to that.) So, these floppies stored data and programs, and this 5+¼ inch floppy disk promised an AI for searching your files. I’m reminded of a lost weekend on this one. Of course, Google eventually created the ultimate AI for search, courtesy of selling ads, as a front for the actual business of trafficking in personal information.
Years later, still on dial-up, I discovered a program that would archive web pages crawling around all by itself. (Digressing again, “dial-up” was when we had a thing called “landlines”…OK, I’ll stop here.) Of course, I set my computer to run all weekend – till it froze up nicely. The program couldn’t tackle such an endless task, no matter my storage space. Right after, The Internet Archive came along and did the job, operating to this day.
Today, large language models are taking in massive amounts of information indexed and accumulated over decades. Our thoughts are now the bits and bytes these models consume whole. Until recently, chat boxes took about 15 minutes of my time before I moved on. Now I’ve renamed my ChatGPT shortcut “Time Suck” to remind me to set a time limit. To judge by social media feeds, for many, this has been a thing for much more than a weekend.
It’s easy to be alarmed about what may come of all this. If you are an artist, the news before the ChatGPT buzz already sounded like your end-days are near. Multiple AIs have also been digesting vast image libraries and regurgitating rarefied fusions of these as new images when asked to render a kitten joining the styles of Warhol and Mehretu. It’s looking bad if you’re a web artist hustling for commissions.
In all these stories, what we don’t have is a good picture of what an AI might do with the rest of its time.
Yet perhaps we are getting ahead of ourselves. I had the good fortune to watch 2001: A Space Odyssey and Colossus: The Forbin Project when I was still a child. This was a curious combination, my first exposure to AIs – the first sympathetic if dangerous, the latter unfeeling and cruel. Ten years later came along my favorite evil AI of all time – Skynet. More recently, children are growing up with stories of super-intelligences that put us all in vats (The Matrix), TV series with battling AIs (Person of Interest), or an AI sending people across time (Travelers). “The Director” in Travelers, never referred to as “it,” has a grand plan. Choose your favorite dystopian future with an AI making mischief – there’s plenty to choose from. In all these stories, what we don’t have is a good picture of what an AI might do with the rest of its time.
In NASA, we were fond of saying, “all models are wrong, but some are useful.” I’m no stranger to models and a poor-engineers version of an AI – genetic algorithms. (These algorithms also take all night to run, now on my laptop, successfully and mostly unattended. See a common theme here?) In the late 1990s, as we looked to design new space launch vehicles, a fundamental problem was everyone’s models were asocial. Your model did not “talk” to my model. So much for “One NASA.” When we first attacked the problem, not everyone thought it could be fixed anytime soon. I saw plenty of “one day, far, far, away – but not now.” The arm waving included showing off your complex model, too complex to connect to anything else without the expert in the loop. Alternately, some argued their models were not complicated enough, being so narrowly crafted. How would anyone, except the maker, know which unique tool to pick from in the garage – or even less, how to use it?
More likely, much of this was about control, the desire not to lose it. If anyone could spin up models across NASA centers and design a spaceplane, what were all the experts to do? My retort was “build better models” – to create ever better spaceplanes. This included building new user-friendly interfaces for the non-experts. (Sounds familiar, non-artists and non-writers who are now generating art and essays?) My naivety was apparent. Yet the project carried on just long enough to see the possibilities.
We gathered at Marshall Space Flight Center for an all-day workshop, cranking up our models and writing real-time scripts to connect them. This marked a moment when the end-to-end connection of models created a mega-model. Albeit, this quiet hurrah soon met its nemesis in a combination of sticker shock and firewalls. To boot, the vibe persisted about the tools being misused, with automation leading to incorrect conclusions. True, the technology was not ready, but it didn’t help that neither were many of the model owners. Bespoke analysis, experts in the loop, workers unite!
Now we know, you can’t hold back a good thing, or an AI. I wrote about this in May of 2021 – “I’m with the AI, and I’m here to help” There, I showed the use of a genetic algorithm to make reusable launch vehicle design decisions. Genetic Algorithms are a low-level of AI, rather basic. These programs seek solutions by experimenting with a model at first to see what happens and “seeding” subsequent results with only the better ones from previous generations.
To appreciate the reason we need such techniques, we can count possibilities. Say you have a bunch of design choices to make, perhaps a few hundred. The number of unique combinations of these choices approaches 10 to the power of 47. Give it a try looking for the best combinations manually – but only for a while. That’s till you realize that adjusting one variable to do better near term is causing trouble later, and vice-versa. It’s then you call the AI and make peace. You may not understand how it gets its results, but it works. (As the Travelers said, “Trust the Director.” Some excellent pieces about the inability to trace an AI’s answers, with valid concerns, were recently published in Vice and Noahpinion.)
For NASA, it’s an obvious leap to see how space transportation and space applications could also be helped by an AI.
A lot has happened in the world of AI in the last 18 months. How appropriate – the AI field moving so fast it’s as if fact is following fiction. Fiction aside, we see AI applications in materials, discovering new invar alloys with ever lower thermal expansion coefficients. Or imagine finding new quantum materials with an AI assist.
For NASA, it’s an obvious leap to see how space transportation and space applications could also be helped by an AI. If you have a complex problem where improving in one spot causes problems in another, oh and by the way, you’re missing some technology that does not yet exist, you have a perfect application. Here, Northrop Grumman still uses the model-of-models approach we experimented with in the late 1990s, but to optimize the design of a stealth aircraft. If we want to see regular passenger service to low Earth orbit, spaceplanes will fit the bill – except we have only the parts and pieces for figuring these out right now. But we know an AI (and a supercomputer) loves to gobble up such complex problems. Similarly, as a reason for going to orbit at all, AIs will soon be finding new drugs for medical treatments or predicting drug response. (Think genetic algorithms, pun intended.)
Soon we may have an AI that generates cat videos, or better yet kittens. The cats and kittens we see won’t be real, but if you can’t tell the difference, will you laugh less? Or we will have other possibilities besides unemployed cats. Spaceplanes that breathe in air, more flying engines than aircraft, could use a dose of AI. The scramjet and hypersonic data are out there, and there’s no lack of prior work in NASA and the DOD. Among more of my stories would be ones where all the numbers said “go down this path” but agencies had neither the funding nor the interest. Too often, agency leadership questioned the technology and the numbers. Perhaps it’s time we trust the AI.