Dropped out of high school, still got into OpenAI: He changed his way of learning, using AI to learn about AI.

He is 23 years old and didn't finish high school.

On his resume, he can't present a diploma from any university.

But in OpenAI's internal system, his title is: Gabriel Petersson, Research Scientist, Sora Team.

After dropping out of high school, Gabriel Petersson worked at Depict.ai, Dataland, and Midjourney successively. He officially joined OpenAI in December 2024 to conduct research on video generation.

No degree. But a long list of projects.

On November 28, 2025, on the podcast Extraordinary, he shared his learning method: When encountering a new field, he never starts by reading thick books. Instead, he directly looks for real - world problems, asks ChatGPT to break down the steps, write code, and debug, and then learns backwards from the problems, filling in knowledge such as mathematics, algorithms, and papers item by item.

Do first, then learn. This is what he calls: Learning AI with AI.

And this is not an isolated case.

Last month, Sam Altman, the CEO of OpenAI and also a dropout, said that he was a bit envious of the current generation of dropout youth. With the support of AI tools, they can create truly useful products faster, and the advantage of traditional diplomas is being compressed.

A Swedish high - school dropout, moving from the classroom to the Sora laboratory, has seized the great opportunity of this era:

Diplomas are depreciating, abilities are being re - evaluated, and the way of learning is being rewritten by AI.

How did he do it?

Section 1 | Reselling Pokémon cards at 14, joining OpenAI at 23

Many people think that he is a technical elite who has been programming since childhood, has always been a top student, and holds competition medals.

In fact, he has no academic qualifications, no famous teachers, and no background.

But he has always been doing one thing: Starting projects, making mistakes, and then making improvements.

The first turning point was at 14

That year, at his home in the Swedish countryside, he earned his first pot of gold, over $20,000, by reselling Pokémon cards. He wanted to build a price - comparison website, so he opened YouTube to search for how to make a webpage. Without anyone teaching or forcing him, he created his first small tool by himself.

"I'm not from a programming background. I just want something and then build it."

The second turning point was during the COVID year

At 18, he built a hand - sanitizer price - comparison website in a week and earned $22,000 in the first week. The pandemic amplified the demand, and he solved a real problem with the simplest crawlers and front - end technology.

A few months later, he was recruited as the temporary CTO of Curb Food, the largest cloud kitchen in Sweden with 80 employees. He built a seven - person engineering team from scratch, developed a kitchen management system, and slept on the sofa while writing code and fixing bugs.

The project was launched, and then he left. But he understood one thing: The success or failure of entrepreneurship is not important. What matters is what works you can leave behind.

The third turning point was when he encountered Midjourney

In 2023, he saw AI image - generation for the first time and was shocked. From that day on, he plunged into AI.

He is not a master at image - generation, but he created one of the most useful UI tools on Midjourney, called fast - grid. He explored everything by himself: How to arrange the interface? How to organize prompts? How to run model parameters? He tried one problem after another.

The project was copied, used, recommended, and spread. Some people call him a hacker, but his goal is simple: To make Midjourney more user - friendly.

These experiences didn't earn him credits, but they left real outputs. It's not about what books he read or what courses he took, but about the products he wrote, the models he fixed, and his ability to debug.

This is what OpenAI ultimately wants.

• A runnable GitHub project.
• A product demonstration that can solve real problems.
• A person who is actually doing things.

He couldn't enter a university, but he gradually made his way into the world's top AI laboratory.

Section 2 | Learn while doing, not after learning

Normally, a doctoral degree is required to conduct research on AI video generation.

But Gabriel relies not on any diploma, but on a bunch of real and executable GitHub projects.

He admitted that at the beginning, he couldn't understand what a diffusion model was at all. He even didn't understand the attention mechanism of the Transformer.

But instead of opening a textbook, he opened ChatGPT.

He asked very specifically: If I want to build a small - scale video generation model myself, what should I understand first?

After ChatGPT gave an answer, he continued to ask: Can you give an example of this concept? Is there corresponding code? Can you explain what each line does? I still don't understand this sentence. Can you explain it again assuming I'm only 12 years old?

Then he copied the code and tried it in the project. If it didn't work, he took a screenshot and provided feedback: What does this error mean? One cycle after another until the bug was solved.

There were no fixed steps and no standard answers in the whole process, only one goal: To solve the bug in front of him.

Gabriel called this method:

Recursive learning from the task (Learning recursively from the task).

Start from the problem, work backwards through the knowledge chain, and fill in the gaps layer by layer.

He doesn't learn from scratch. Instead, he learns backwards from the thing he wants to do.

The traditional route is bottom - up:

Learn mathematics → Then learn probability → Then learn neural networks → Only then can you understand the diffusion model

Gabriel's route is top - down:

A bug occurs in the diffusion model → Ask about attention → Need mathematics → Fill in concepts → Continue to modify the code

What's the result? He said:

"Using this method, it only takes three days to learn the diffusion model top - down, while the traditional bottom - up path takes six years."

It's not just theoretical understanding, but the kind of understanding that can be applied in parameter adjustment and practical use.

This ChatGPT - driven reverse learning has given him the ability of a doctor. But it only takes three days instead of six years.

He said: Many people use AI and leave after asking once. I keep asking until I truly and fully understand.

Start from the task, start from intuition, and learn knowledge backwards from mistakes.

AI is the top - notch teacher, but you have to keep asking questions.

Section 3 | He doesn't just use ChatGPT; he conducts research with ChatGPT

Sora is one of the most complex multimodal models at OpenAI.

Generating each frame of video requires scheduling billions of parameters, coordinating multiple modules, and performing dual reasoning in time and space.

This is not as simple as asking questions in a chat box and generating a picture. It's real AI R & D: Training models, adjusting parameters, writing code, and solving bugs. Gabriel is the person doing this.

He said that his daily work process is probably like this:

Observe the videos generated by the model and find out the unreasonable parts

Assume where the problem lies in the architecture and write prompts for GPT - 4 to help with analysis

After getting the suggestions, read/rewrite the code of the core module

Then use ChatGPT to debug, or refer to papers to understand the mechanism - After modification, retrain and evaluate the effect with visual results

If it's not right, ask again, try again, and retrain

Many times, my conversation with GPT is not to get an answer, but to make my thinking more complete.

He will explain the situation in detail as if talking to a human colleague: What parameters I adjusted in this module, where the generated video is unreasonable, I think it may be that the model doesn't pay enough attention to details, and maybe some time - related information needs to be added...

After GPT understands, it will give him two directions:

Quick trial - and - error type: Adjust the weights in a different position and retrain directly

Structural adjustment type: Introduce other mechanisms, such as depth - wise conv

Then he decides which one to try.

It's Gabriel who really makes the judgment. And it's also him who really executes the code.

But GPT gives him a second brain: Not afraid to try, not tired, not annoyed, with a wide coverage, and able to respond at any time.

In the traditional scientific research system, a doctoral student needs to consult a supervisor, find reference code, and wait for group discussions to continue the experiment.

Gabriel has handed over these steps to GPT. It's a 24 - hour online research partner.

He said: I didn't defeat doctors with ChatGPT. I use it as a researcher.

The real difference, is not whether the model is smart, but whether you know how to use it.

Section 4 | Diplomas are no longer important. These three things are

Gabriel himself doesn't emphasize the label of academic counter - attack. He clearly stated in an interview:

I didn't get into OpenAI by proving that I'm extraordinary. Instead, I used projects to show them that I'm already doing what they want to do.

So, the three things he really did right are:

First: Start from projects, not from books

He never asks which concept to learn first. Instead, he asks which problem he needs to solve now.

When he found Midjourney not user - friendly, he wrote a UI tool.

When he didn't understand the diffusion model, he started from a video generation demo.

When he wanted to study Sora, he built a simplified version and adjusted it himself.

He doesn't start doing after learning. Instead, he learns while doing. Each time he completes a project, he fills in a piece of the knowledge puzzle.

Second: Use AI to accelerate understanding, not to skip it

He doesn't use GPT to write code for him. Instead, he uses it to help with understanding.

He asks repeatedly about the parts he doesn't understand until he can explain them clearly.

He asks GPT to analyze the logic of each line of code during debugging.

He asks GPT to summarize the key changes in papers when a new model is released.

Many people use AI to avoid thinking, while he uses AI to promote his own thinking.

Third: Let your works speak, don't wait for academic certification

When applying for a job at OpenAI, he couldn't present a diploma, a scholarship certificate, or a recommendation letter. But he submitted three other things:

A usable GitHub project (fast - grid)

Engineering experience accumulated at Midjourney

A small - scale video generation pipeline project he replicated himself

These three things are not about what he learned, but about what he created.

OpenAI hired him not because he could recite "transformer", but because he really built a usable transformer.

What they want are not people who can memorize terms, but people who can understand problems, drive experiments, and produce results.

Many people are amazed at his successful counter - attack.

But behind this is a bigger trend: Academic degrees can only prove that you've taken courses, while projects can prove that you've done real work. Learning according to the curriculum is step - by - step, while learning according to problems is real skill.

AI has broken down the knowledge barrier. With AI, you can learn at any time. Whether you can learn depends on whether you know how to ask questions.

In the AI era, what's scarce is not knowledge, but the ability to ask questions and the ability of autonomous learning.

In future recruitment, it's not about where you graduated from, but about whether you can use AI to produce real results.

Conclusion | It's not exams that determine life, but works that speak

Gabriel's story is not about an unorthodox approach defeating the regular army.

It's a replicable learning path in the AI era:

Start from projects, use AI to accelerate understanding, and prove your ability with works.

This is not an exception, but a paradigm shift:

No longer learn and then do, but learn while doing.

No longer wait for academic certification, but let projects speak.

No longer use AI to write on your behalf, but use AI for in - depth understanding.

He said: I'm not a genius. I just changed my learning method and learned AI with AI.

Now, works are more important than academic degrees.

📮Reference materials

https://www.youtube.com/watch?v=vq5WhoPCWQ8&t=1547s

https://www.businessinsider.com/high - school - dropout - openai - chatgpt - learn - ai - gabriel - petersson - 2025 - 11

https://www.theatlantic.com/technology/2025/11/openai - lawsuit - subpoenas/684861/

https://x.com/GabrielPeterss4/status/1977907158504133118?referrer=grok - com

https://www.plymouthstreet.com/stories/gabriel - petersson

https://github.com/gabrielpetersson

This article is from the WeChat official account "AI Deep Researcher", author: AI Deep Researcher, published by 36Kr with authorization.