Read the full transcript of bioengineering researcher Eric Nguyen’s talk titled “How AI Could Generate New Life-Forms”, recorded at TED2025 on April 9, 2025.
Listen to the audio version here:
The Engineer’s Approach to Biology
ERIC NGUYEN: I’ve always found it curious how biologists study life. There’s this saying: a biologist will learn how a car works by poking at it, removing one part at a time and seeing how it affects the rest of the car. On the other hand, an engineer will learn how a car works by taking it completely apart and rebuilding it.
Take for example the Human Genome Project, one of the biggest breakthroughs of the last century. We spent over a decade mapping out all 3 billion letters of our genome, the complete set of DNA. We thought that once we could read DNA and apply that same principle, poking and dissecting one letter at a time, that we could start eradicating all human diseases. But instead, we began to realize just how little we actually understood about the true function of DNA.
As a researcher, I work on artificial intelligence and trained as an engineer. I learn by building things and understand by creating. Today I want to share an idea that can fundamentally change how we study biology and life itself. Instead of just reading and dissecting DNA, we should be generating it. And we can do this by treating DNA as a language, one that AI can learn to read, write, and ultimately build.
A Moonshot: Generating Life from Scratch
This idea led myself and a team of researchers at Stanford and the Arc Institute with a sort of moonshot. Can we generate an entire genome from scratch using AI? Build life from the ground up? Now I understand the thought of feeding the code of life into a generative AI is both thrilling and perhaps unsettling, but I came to realize that if this was possible, it could unlock some of the most powerful breakthroughs in science and medicine.
But I’ll be honest, we had no idea if AI could actually generate DNA. In many ways, DNA is like a language. It has grammar, structure, sort of like sentences and paragraphs that group together to form a story. And these stories are passed down through evolution, generation by generation.
The Scale Challenge of DNA
For humans, it’s been hard to comprehend these stories written in DNA, in large part because of its scale. DNA is extremely long, and yet at the same time sensitive to the smallest mistakes. Imagine trying to write something the length of 30,000 books in a foreign language, and that when you’re off by a single letter, one of billions, this can mean the difference between a healthy person and a person with a life-threatening disease.
And so to tackle these challenges, together with my colleague Michael Pauly, we developed an AI that could generate extremely long sequences of DNA 500 times longer than previous AI models at high levels of detail. We assembled a team of scientists and AI experts and gathered the largest collection of DNA used to train AI: 80,000 whole genomes fed into a model that we called Evo.
Introducing Evo: ChatGPT for DNA
And our goal was to create something like a ChatGPT for DNA, where you can prompt Evo and describe the DNA you want, and it would generate new sequences one letter at a time. But there is one key difference. With chatbots, you can just read what it writes, and you can decide if it makes sense to you. With DNA, it’s not so simple. It’s not an intuitive human language. How do you know if it’s any real or good? What does that even mean?
What we needed was a test, a way to verify that what it wrote would actually function. And so we started with a familiar tool in biology called CRISPR. CRISPR is like a pair of molecular scissors that can edit DNA used for things like gene therapy. We asked Evo to make its own and generate its own version of CRISPR from scratch, which had never been done before. It’s got proteins and RNA inside. It’s a complex system.
Testing AI-Generated CRISPR
And so our biologists would take that generated DNA and analyze it to see how realistic does it look? Does it resemble something in nature? How do its proteins fold? Which all give us a sense of its function? But ultimately, we have to test their ability to cut DNA by actually building them in the lab. And so that’s what we did.
Now, waiting for lab results can sometimes be a nerve-wracking experience. Honestly, it’s kind of like waiting for the results of a pregnancy test. You’re excited, anxious, and hopeful for a positive outcome. And then all of a sudden, these two little lines appear. Just to be crystal, crystal clear, these two lines are a good thing. That’s what we want.
The First AI-Generated CRISPR System
It means that our CRISPRs cut a single strand of DNA into two in the exact right spot, just like natural CRISPRs in the lab. And so that’s when we knew it worked. What you’re seeing here is the world’s first CRISPR system designed entirely by AI. Evo-generated DNA that not only looked realistic, but that actually functioned.
And so next, we decided to go for that moonshot, to try and generate a whole genome from scratch. And Evo was able to generate hundreds of synthetic proteins in a genome that resembled those in nature. But ultimately, he was missing a few parts. It wasn’t yet complete, more like a rough sketch of the genome. However, this is just the first version. That rough sketch of the genome will become more detailed over time. In fact, within years, we anticipate AI will be able to generate whole, functional genomes. In other words, AI will be able to generate new life.
From Discovery to Design
As this technology improves, biology will shift from discovery to design.