Robots, rather than the software that runs them, are what first appealed to Soham Batra.
“My interest in technology didn’t start with AI,” said the 17-year old high school senior. “It started with robotics. When I was a lot younger, around elementary school, I was very involved with Lego robotics. That sparked my interest. I was thrown into a class where we expected to make them move. We’re just a bunch of 6-year-olds and I was [thinking]: ‘How do I even do that?’ Then I started looking into it and found programming, coding. As I moved to high school the entire AI bubble started expanding, and got to everyone.”
High school for Batra being the Illinois Math and Science Academy (IMSA) in Aurora, a residential hive of some of the brightest kids in the state. There the senior distinguished himself… well, maybe we should let him explain, the way he did in his introductory email:
“You wrote last year that the AI genie is out of the bottle. I’m one of the people crawling into the bottle after it, and I’m in high school,” he began. “I do machine learning research on the side. Two of my papers were oral presentations at ICML and CVPR 2026 workshops (one on keeping a model aligned while it keeps learning, one on carrying 3D reconstruction uncertainty into physics simulations)…”
ICML is the International Conference on Machine Learning. CVPR refers to the Computer Vision and Pattern Recognition Conference. Before we wander too far into the weeds, I should explain why I was interested in Batra.
Most everything I see in the press about artificial intelligence is disaster-oriented. The leaders of AI, sounding the gong of doom. But kids like Batra don’t see it that way. To him, AI is just “a really cool tool.”
“The thing with AI, there’s nothing like it,” he said. “I got really lucky, to see the development of AI at a very young age. A lot of people ask if the AI bubble is going to pop… If you try to do your day-to-day tasks without AI, and with AI — you’ll really see a really big difference.”
At IMSA, students are given Wednesdays off to tend to side projects, and I could fill the column with Batra’s — such as doing research at the University of Chicago’s Booth School of Business.
The project that really impressed me, however, is a simulation he built called “Play the House,” where participants run a casino, adding new games, setting the house edge, all the way down to dangling drink specials. Young people are being seduced into gambling, particularly at colleges, which abandon their responsibility to look out for their charges by allowing betting sites to entice them. The point of Batra’s game, which he made in “an extremely basic version” before using AI to “glamour it up a bit so it was ready to put out there on a domain” is to drive home the insurmountable advantage casinos enjoy.
“I’ve had a couple family members succumb to gambling,” he said. “I’ve seen what it can do, first-hand.” He developed the game, “to better help people understand how the odds are stacked against them.”
Space doesn’t allow me to explain his AI work, not that I understand it myself.
“In my opinion, the most important [of his papers] he calls ‘the geometry of forgetting,'” said his IMSA adviser, Steve Goldblatt. “When you are working with AI, machine learning, it is an iterative process. What he has done with that is brilliant. He found a way for those parameters — for what you put in to stay in through the entire iterative process.”
Batra, who is from Naperville and whose parents are immigrants from India, is also the goaltender on the school’s varsity soccer team.
“He’s closer to your actual teenager than a lot of the students here,” said Goldblatt. “He jokes around. Very social, has a lot of friends here. At IMSA, we have a student body of introverts. Soham is definitely an extrovert, out there talking to people.”
When I asked why he reached out to me, his answer was surprising.
“Honestly, I always look for guidance in what I’m trying to do,” Batra said. “Often time, the advice people give me doesn’t exactly help me. What helps me is when people ask me questions about what I want to do. When I answer the questions, I figure out a lot more about what I actually want to do. It’s something I think about often, but when people ask me, I’m able to give them an answer. I’m trying to figure out what I’m trying to do. Steps like these are important.”
For instance, I asked about his college plans. He thinks he will go, but two former classmates at IMSA went to college then immediately dropped out and formed companies now worth $90 million. The straight-to-business route is tempting.
While that decision looms, he is ironing out the future of AI, which does veer toward malfunction.
“You have to train models, so they’re able to answer questions,” he said. “When these big companies develop open AI, they train them on text, images. They have to have certain guidelines, to refuse harmful requests. That’s what you call alignment. When you take that same AI and you fine tune it on something else, you try to make the AI specific to medical records, or legal documents, the alignment shifts slightly and the safety behavior degrades. People are able to find slight loopholes when you over-train a model. It has something called a ‘context window,’ the amount of information it’s able to handle. Once it exceeds that, it hallucinates, meaning it provides misinformation. My study was on how safety behavior degrades. how easy is it to degrade the safety model.”
“I get the impression it’s real easy,” I said.
“It is kinda scary how easy it is,” he said. “You can watch the alignment degrade. There’s no exceptions to it not degrading. It’s not possible to improve it. The only thing it can do is make it worse.”
That sounds bad.
“The thing I did find, no matter how you fine tune, you could never get it to do something so terrible,” Batra said. “The contributions, the strides we’ve made in AI, there’s no way that it’s not going to be part of our lives for a really, really long time. Almost forever. It is a very useful tool and should be consistently used, but more research should make it safer. It will be a lot more useful.”