The world of artificial intelligence has conquered text, images, and video. Now, it’s setting its sights on the physical world. In a recent episode of MTS, host Sophia sat down with Thijs Simonian, a 20-year-old builder, Stanford student, and former OpenAI intern who has been at the forefront of some of the most viral AI demos of the year. Their conversation touched on everything from the changing culture of tech to the impending revolution of AI-controlled robotics.

Here’s a breakdown of why general Large Language Models (LLMs) are poised to control physical robots—and how builder culture is shifting to make it happen.

The “Computer Use” Parallel

Thijs draws a direct and fascinating parallel between the early days of AI “computer use” (models learning to click, type, and navigate desktop interfaces) and the current state of robotics.

Just a few months ago, having an AI control a computer was a clunky, slow, and error-prone process. The models would click in the wrong places, get confused, and take forever to complete a simple task. However, the fact that it could succeed at all was the critical breakthrough. Once a model achieves a baseline level of success, developers can train toward that success, rapidly improving the model’s speed and accuracy.

Right now, physical robotics is in that exact same nascent, clunky stage.

“You plug a robot into your computer, ask it to do something like pick up a block and put it in a bowl, and it’s like… slow,” Thijs explains. “It moves like super slow, it keeps making mistakes like it’s hitting the block in the wrong way, but it does do it. It is able to succeed a lot of the time.”

The exciting part? This isn’t being done with highly specialized, niche robotics software. It’s being done with general-purpose LLMs, like OpenAI’s Astra or Anthropic’s Claude models.

Why General LLMs Change the Game

Historically, getting a robot to perform a task required rigid, specialized programming. If you wanted a robot arm to paint, you had to write code specifically for painting, calibrating every coordinate and joint angle.

The breakthrough of using general LLMs (like Astra) is that these models already possess a vast understanding of the world, physics, language, and logic. When you give an LLM control of a physical robot, you aren’t writing new code from scratch; you are leveraging the model’s existing reasoning capabilities and applying them to physical outputs.

During his internship at OpenAI, Thijs worked with researcher Tel to create a viral demo just before the Astra launch: a robot arm that successfully painted a picture. It wasn’t a specialized robotics model; it was a general-purpose LLM navigating the physical world.

As publicly released models like Claude 3.5 Sonnet or future iterations of GPT get better and faster, the barrier to entry for robotics will plummet. Anyone with access to an LLM will soon be able to program complex physical tasks simply by asking the model to do it.

The Builder’s Mindset: Ideas as a Muscle

Beyond robotics, Thijs’s story is a testament to the shifting landscape for young tech talent. He started building Discord bots at age 14, scaling one to 1.6 million communities. Despite being rejected twice from traditional intern pipelines at OpenAI, he landed a job by networking through his niche work on video model evaluation and coming to the table with a clear, self-directed roadmap of what he wanted to build.

His advice to aspiring builders—whether they are students or seasoned professionals—is simple but profound: Treat idea generation as a muscle.

If you don’t actively flex your creative muscles and act on the ideas you have, that capability will slowly fade. You don’t need a formal lab, a startup incubator, or even a full-time job to innovate. Thijs emphasizes the importance of always having a short-term (e.g., three-month) project or goal to focus on in your free time.

“Creating stuff and bringing it into the world is one of the most beautiful and awesome and fun and rewarding things you can possibly do,” Thijs notes.

The Physical Frontier

We have largely conquered the digital frontier. As we look ahead, the integration of general LLMs with physical robotics is the next massive leap.

Whether it’s an LLM navigating a car or a robotic arm painting a canvas, the clunky, slow demos of today are the foundation for the rapid, fluid automation of tomorrow. And as builders like Thijs prove, the people who will shape this future aren’t just the ones passing standardized tests—they are the ones tinkering, building, and flexing their idea muscles in their free time.