OpenAI just launched a page called "ChatGPT Futures".
There are 26 young people (or teams) in total, each receiving a $10,000 prize, plus access to cutting-edge models.
Among them, the most eye-catching name is Matteo Paz.
Last March, he was an 18-year-old high school student. He developed machine learning algorithms to process nearly 200TB of data and about 200 billion records accumulated by NEOWISE over more than ten years of infrared sky surveys. From this data, he marked and classified 1.9 million infrared variable source objects, of which about 1.5 million were potential new discoveries that had not been recorded before.
His paper was published in the Astronomical Journal.
In March of this year, he also won the top prize in the Regeneron Science Talent Search.
According to Caltech, this is "a local high school student achieving a breakthrough at Caltech."
Paz was just one of the 26 selected.
On March 11, 2025, 18-year-old Matteo Paz holds the Regeneron Science Talent Search trophy at the awards ceremony. He won the award for discovering 1.5 million unknown celestial objects using AI algorithms.
On the same list, there are also—
18-year-old Crystal Yang: She developed "listening instead of seeing" learning games for 200,000 visually impaired students;
19-year-old Anshi Bhatt: Her anti-fraud system has helped 18,000 people avoid online scams;
25-year-old Amrita Bhasin: Her logistics system has allowed more than 5 million pounds of unsold inventory to bypass landfills.
...
The 26 projects range from astronomy to disaster relief, from medicine to agriculture, from education for blind children to financial management of street vendors in South America. None of them are "writing papers using ChatGPT". They are all focusing on tough issues that used to require qualifications, institutions, and funding to tackle.
AI empowers them to think and act, something the previous generation of young people could hardly imagine.
The "first generation of ChatGPT natives" have graduated.
The Class of 2026 will be the first cohort of graduates to have ChatGPT “available at any time” throughout their university experience.
While "always available" does not mean "full reliance," it is enough for AI to reshape the learning and lifestyle of a generation.
About three and a half years ago, in the fall of 2022, the class of 2026 enrolled. A little over two months later, on November 30th, ChatGPT was born. Their university life became inextricably linked with ChatGPT, and the "first generation of ChatGPT natives" were born.
Before their first semester of freshman year was even over, an AI appeared on their desks that could write code, find literature, and chat about any topic.
Among these 26 individuals (or teams), there are 18-year-old high school students and research groups formed across schools. They are not all labeled as "recent graduates," but they are all samples of this generation of young people.
OpenAI's "ChatGPT Futures" initiative is not just about awarding prizes, but also about setting an example for "outstanding young people in the AI era."
They "use AI to see things that humans cannot see".
What are the "first-generation ChatGPT natives" doing with AI?
Let's look at three of the most representative projects first.
The first one is Matteo Paz's project.
He was facing NEOWISE: all the data collected over ten years by a retired NASA infrared survey telescope.
In the words of Paz's mentor, Davy Kirkpatrick, "This table is approaching 200 billion rows, recording every probe we've made over the past decade."
With 200 billion lines and nearly 200TB of data, it's impossible to process it all by human means alone. This is the kind of work that AI can do but humans find very difficult to do.
In 2023, Matteo Paz presented early results of his AI astronomy project at the Caltech Summer Research Connection workshop.
Paz wrote a machine learning algorithm called VARnet, which went through the entire table and identified 1.9 million infrared variable source objects, of which 1.5 million were completely new discoveries that had never been recorded before: supermassive black holes, newborn stars, supernovae...
Kirkpatrick originally only hoped to "find a few variable stars and tell the astronomical community that there is treasure in this data" with this work.
As a result, Paz generated a complete catalog of the entire dataset: 1.9 million variable-source objects, divided into ten categories, all archived.
The second project is called AION-Search, and it is managed by Nolan Koblischke.
His goal is to make 140 million galaxy maps "searchable by natural language".
Traditional astronomical image retrieval relies either on image similarity or predefined categories. Want to find "spiral galaxies showing signs of merger" or "suspected gravitational lensing"? Sorry, you'll need to train a specialized classifier first.
AION-Search has released a demo interface that supports natural language search. The paper claims the system can scale to 140 million galaxy images. https://huggingface.co/spaces/astronolan/AION-Search
Koblischke's approach was to first have GPT-4.1-mini automatically write text descriptions for 275,000 galaxy images (at a cost of $150); then use contrastive learning to train a shared image-text retrieval space; and finally, expand to 140 million images.
How hardcore is the result?
Gravitational lensing is the rarest type of object in galaxy data, accounting for only 0.1% of the entire database: equivalent to finding 1 in 1000 images.
When traditional image similarity algorithms are used, almost all of the first 10 results are wrong. However, with AION-Search, a significant portion of the first 10 results are correct.
The industry uses a metric called nDCG@10 to measure how accurate the ranking of the top 10 results is.
AION-Search achieved a score of 0.180, while traditional methods only scored 0.015: this represents a more than 10-fold improvement in search performance.
Rare phenomena that used to require astronomers to manually search through hundreds of thousands of images can now be found using natural language.
The third project is called WiFind.
The project, WiFind, was created by Nayel Rehman, Arhan Menta, Rushil Kukreja, and Aayush Tendulkar. It uses AI to process WiFi signals, attempting to penetrate walls and rubble to find survivors in disaster areas.
WiFind project team members
WiFind is currently a Springer conference paper and a Conrad Challenge award-winning project. It is still in the prototype stage and is not an already deployed disaster relief system.
But its idea is very novel: WiFi routers are everywhere in the world, and each one is a potential "life detector".
There's also Zeyneb Kaya using AI to protect endangered languages; Amrita Bhasin's project has diverted over 5 million pounds of unsold inventory from landfills for reuse...
What these 26 projects have in common is not "using AI to write academic papers," but rather "using AI to tackle things that humans cannot."
26 names, not just celestial bodies and rescue efforts.
If you lay out this list in its entirety, you will see a more comprehensive picture:
The 26 selected individuals (teams) come from more than 20 universities and institutions, including MIT, Stanford, Harvard, Oxford, Berkeley, Yale, etc. The list basically covers the top tier of research in North America and the UK.
OpenAI categorizes them into three groups: Creators (who create products), Explorers (who conduct research), and Advocates (who promote and popularize products).
Celestial body discovery, galaxy search, and disaster relief are just the three most concentrated areas.
Among the remaining projects, some are creating learning aids to reduce stress for their peers; some are translating mental health materials into minority languages so that psychological counseling is no longer limited to the English-speaking world; some are developing accessibility features for students with disabilities so that classrooms no longer exclude people; and others are using AI to identify fraudulent messages and prevent the elderly from being scammed.
Kyle Scenna, 24, from Waterloo, is an entrepreneur. Speaking about ChatGPT, he said, "I never imagined the distance from identifying a problem to making it work could be so short."
Michelle Lawson, 20, is a student at Smith College. She says, "I've always believed that with the right support and resources, you can achieve anything you imagine. AI has made that possible, for myself and for thousands of others."
Nolan Windham, 23, is already the head of AI at a well-known hedge fund. He says, "The exciting thing is, this is just the beginning."
When it comes to AI, they all have one thing in common: AI has enabled them to do more things.
This is the biggest difference between this generation of "AI natives" and the previous generation:
They have come to regard AI as the default infrastructure, an indispensable part of their learning and life, just like the previous generation of digital natives viewed "Wi-Fi".
The threshold hasn't disappeared, it's just moved.
The fact that even high school students can make astronomical discoveries may give many people an optimistic illusion: that AI has truly lowered the barriers to scientific research.
However, it is too early to make such a judgment. Let's take a look at Paz's complete resume first.
In the summer of 2022, while he was still in high school, he entered Caltech's Planet Finder Academy.
In 2023, he participated in Caltech’s six-week Summer Research Connection program, with IPAC senior astronomer Davy Kirkpatrick serving as his research mentor.
Paz completed the Pasadena School District's "Math Academy" program in high school: he completed AP Calculus BC in eighth grade, a subject that most high school students don't encounter until 12th grade, but he mastered before the age of 14.
In other words, Paz is not "an ordinary high school student plus ChatGPT", he is "a high school student who has advanced to college level in mathematics, has been mentored by top Caltech tutors for two years, and can directly access IPAC computing resources", plus AI.
https://arxiv.org/pdf/2512.11982
The paper also mentions the limitations of AION-Search, which allows 140 million galaxy images to be searched using natural language:
VLM misses out on subtle astronomical structures and introduces the biases inherent in GPT-4.1-mini into the system. The fact that the whole method works in the field of astronomy is also due to the fact that manually labeled data such as Galaxy Zoo has been used as training corpus by GPT.
The phenomena found by AI are mainly those that astronomers already knew how to label.
WiFind, which uses WiFi signals to penetrate rubble to find survivors, is currently only a prototype and not a rescue system already in operation in earthquake-stricken areas.
AI lowers the "threshold of repetitive labor," but it does not lower the "taste, judgment, and long-term training."
The key to Paz's story is not that AI enables any high school student to do astronomy, but that a high school student who was destined to make an astronomical discovery brought it forward by ten years.
The threshold hasn't disappeared; it's just shifted from "whether it can be done" to "whether it can be thought of."
References:
https://x.com/OpenAI/status/2052086313797705954
This article is from the WeChat official account "New Zhiyuan" , author: New Zhiyuan, and published with authorization from 36Kr.
