OpenAI just launched a page called “ChatGPT Futures”.
There are a total of 26 young people (or teams). Each individual (or team) receives a $10,000 bonus and access to cutting-edge models.
The most eye-catching name among them is Mateo Paz.
Last March, he was still an 18-year-old high school student. He developed machine learning algorithms to process approximately 200 TB of data and approximately 200 billion rows of records accumulated from over 10 years of infrared sky surveys by NEOWISE. He marked and classified 1.9 million infrared metamorphic objects. Approximately 1.5 million of these were potentially new discoveries that had not been previously recorded.
His paper was published in “Astronomical Journal”.
In March of this year, he won the top award in the Regeneron Science Talent Search.
According to Caltech, this is a “local high school student who made a breakthrough at Caltech.”
And Paz is just one of the 26 chosen.
On March 11, 2025, 18-year-old Matteo Paz picked up the Regeneron Science Talent Search’s top prize trophy at an awards ceremony. He won the award for discovering 1.5 million unknown celestial objects using AI algorithms.
The same list also includes –
Crystal Yang, 18: She developed a “listen instead of look” learning game for 200,000 visually impaired students.
Ansi Bhatt, 19: Her fraud prevention system has helped 18,000 people avoid online scams.
Amrita Bhasin, 25: The logistics system she built diverted more than 5 million pounds of unsold inventory from landfills.
…
From astronomy to disaster relief, medicine to agriculture, educating blind children to managing the finances of street vendors in South America, none of the 26 projects are “writing a paper on ChatGPT.” All of them are tackling difficult problems that previously could only be tackled with seniority, institutional support, and sufficient funding.
AI enables bold thinking and action that was unimaginable to previous generations of young people.
“ChatGPT native first generation” has graduated
The Class of 2026 is the first group of graduates to have “ready access” to ChatGPT throughout their college career.
“Ready access” does not mean “total dependence,” but it is enough for AI to reshape the learning and living styles of a generation.
Approximately three and a half years ago, in the fall of 2022, new students from the Class of 2026 entered the university. Over two months later, on November 30th, ChatGPT was launched. Since then, their college days have been linked to ChatGPT, creating the “first generation of ChatGPT natives.”
Before the end of their first semester, their desks had AIs that could write code, search literature, and talk about any topic.
These 26 individuals (or teams) include 18-year-old high school students and research groups formed across different schools. Although not all of them are labeled as “graduates”, they are all samples of this generation of young people.
“ChatGPT Futures” launched by OpenAI this time aims not only to give bonuses, but also to set an example of “talented young people in the age of AI.”
They use AI to see things that humans cannot see.
What does “first generation ChatGPT native” do with AI?
First, let’s take a look at three representative projects.
The first is Matteo Paz’s project.
He works with all the data accumulated from 10 years of sky surveys by NEOWISE, NASA’s retired infrared sky surveying telescope.
In the words of Paz’s tutor, Davy Kirkpatrick, “There are nearly 200 billion rows in this table, recording every detection made over the past 10 years.”
With 200 billion rows and nearly 200 TB of data, it’s impossible to process manually. This is the kind of work that AI can do but is difficult for humans.
In 2023, Matteo Paz presented the initial results of his AI astronomy project at the Caltech Summer Research Connections Seminar.
Pass created a machine learning algorithm called VARnet that went through the entire table and marked 1.9 million infrared variable objects. Of these, 1.5 million were new discoveries never before recorded, such as supermassive black holes, newborn stars, and supernovae.
For this study, Kirkpatrick initially only wanted to “find some variable stars and tell the astronomical community that there is still treasure in this data.”
As a result, Paz has created a complete catalog of the entire dataset. 1.9 million source objects are classified into 10 major categories and are all archived.
The second project is called AION – Search and its manager is Nolan Koblischke.
His goal is to make 140 million galactic maps “searchable in natural language.”
Traditional astronomical image search relies on image similarity or predefined categories. What if you want to find “spiral galaxies with signs of merging” or “suspected gravitational lensing”? Sorry, you need to train a special classifier first.
AION – Search’s public demo interface supports natural language search. The paper says the system can scale to 140 million images of galaxies. https://huggingface.co/spaces/astronolan/AION – Search
Koblischke’s approach is as follows. First, let GPT – 4.1 – mini automatically create text descriptions for 275,000 galactic maps (cost $150). Next, use contrastive learning to train a shared search space for images and text. Eventually, we’ll scale that to 140 million maps.
How effective is it?
Gravitational lenses are the rarest targets in galaxy data, accounting for only 0.1% of the entire database. It’s like finding one photo out of 1,000.
When searching using traditional image similarity algorithms, almost all of the top 10 results are wrong. In AION – Search, a significant portion of the top 10 results are correct.
The industry uses a metric called nDCG@10 to measure “how accurately the top 10 results are ranked.”
AION – Search gives 0.180, but traditional method only gives 0.015. This is more than a 10x improvement in search efficiency.
Phenomena that previously required astronomers to manually search through hundreds of thousands of images can now be found using natural language.
The third project is called WiFind.
The WiFind project was developed by Nayel Rehman, Arhan Menta, Rushil Kukreja, and Aayush Tendulkar. It uses AI to process WiFi signals and penetrate walls and debris to try to find survivors in a disaster area.
WiFind project team members
WiFind is currently an award-winning project at the Conrad Challenge and presented in a paper at the Springer conference. This is still in the prototype stage and is not a deployed disaster relief system.
However, the concept is very innovative. Wi-Fi routers are everywhere in the world, and each one is a potential “life detector.”
There’s Zeyneb Kaya, who uses AI to protect endangered languages. Amrita Bhasin’s project diverts more than 5 million pounds of unsold inventory from landfill for reuse…
What these 26 projects have in common is not “writing papers using AI,” but “using AI to tackle things that humans cannot do.”
26 names that are not just celestial bodies or rescues
If you look at the entire list, you’ll get a more three-dimensional picture.
The 26 selected individuals (teams) come from more than 20 universities and institutions, including MIT, Stanford, Harvard, Oxford, Berkeley, and Yale, a list that essentially includes top research institutions in North America and the United Kingdom.
OpenAI divides people into three categories: creators who develop products, explorers who conduct research, and advocates who promote and disseminate products.
Astronomical discovery, galaxy exploration, and disaster area relief are just the three most concentrated directions.
Some of the remaining projects are developing learning support tools to reduce peer pressure. Some are also translating mental health materials into the native languages of ethnic minorities so that psychological counseling is no longer limited to English-speaking countries. Some are developing barrier-free features for students with disabilities to make classrooms less exclusive. Some companies are using AI to identify fraudulent information to prevent seniors from being scammed.
Kyle Schena, 24, of Waterloo, is an entrepreneur. When talking about ChatGPT, he said: “I never thought the distance between discovering a problem and implementing a solution could be so short.”
Michelle Lawson, 20, is studying at Smith College. She said, “I have always believed that with the right support and resources, you can achieve anything you imagine. AI has made this a reality for me and thousands of others.”
At 23 years old, Nolan Windham is already a well-known hedge fund AI director. “What’s interesting is that this is just the beginning,” he said.
The common thing about AI is that it can do more things.
This is the biggest difference between this generation of “AI natives” and the previous generation.
They have seen AI as a default infrastructure, an essential part of learning and living, just as previous generations of Internet natives thought of “Wi-Fi” as “Wi-Fi.”
The threshold hasn’t disappeared, it’s just changed.
Since even high school students can make astronomical discoveries, many people may have the optimistic illusion that AI has truly lowered the barrier to scientific research.
However, it is too early to make such a judgment. First, let’s take a look at Mr. Paz’s complete resume.
In the summer of 2022, while still in high school, I enrolled in the Planet Finder Academy at Caltech.
In 2023, he participated in a six-week Summer Research Connections project at the California Institute of Technology, with IPAC senior astronomer Davey Kirkpatrick as his research instructor.
Paz completed the Pasadena School District’s “Math Academy” program in middle school and AP Calculus BC in 8th grade. He mastered calculus before the age of 14, which most high school students encounter in the 12th grade.
In other words, Paz is not an “ordinary high school student + ChatGPT”. He’s a “high school student who went on to take college-level math, had a top-notch tutor at Caltech for two years, and had direct access to IPAC’s computing resources,” and is equipped with AI.
https://arxiv.org/pdf/2512.11982
A paper on AION – Search, which allows users to search 140 million galactic maps in natural language, also notes its limitations.
VLM can miss subtle astronomical structures and introduce GPT – 4.1 – mini bias into the system. This entire method works in the field of astronomy because manually labeled data, such as data from the Galaxy Zoo, is used as training material for GPT.
AI primarily discovers phenomena that astronomers already knew how to label previously.
Additionally, WiFind, which uses WiFi signals to penetrate rubble and find survivors, is still only a prototype and not a rescue system deployed in earthquake-affected areas.
AI has flattened the “repetitive labor threshold,” but not “taste, judgment, and long-term training.”
The point of Paz’s story is not that AI made it possible for any high school student to do astronomy, but that high school students who were just trying to make astronomical discoveries advanced it by a decade.
The threshold has not disappeared. It just changed from “Can you do it?” to “Can you think of it?”
Reference materials:
https://x.com/OpenAI/status/2052086313797705954
This article is from the WeChat official account “New Source”, written by New Source, and published by 36Kr with permission.
