Everyone is talking about “physical AI,” the idea that models will finally step out of the chatbot and into the real world of robots, factories, and labs. I ecently hosted Cyriac Roeding, founder of Earli, a company curing cancer with AI, for a Cornell keynote on exactly this topic. He said something I have not stopped thinking about. “Physical AI is only interesting when you train it on physical-world data.” That one sentence explains why so many AI strategies, in healthcare and far beyond, are aimed at the wrong target.
Here is the part that sounds like science fiction but is not. For fifty years, biology was something we could only read. We are now starting to write it. AlphaFold made proteins predictable and earned a Nobel Prize in 2024. DNA is next. Models like Evo 2, built by the Arc Institute and Nvidia and recently published in Nature, treat the genome as what it actually is, a language with a four-letter alphabet. And if DNA is text, then the same transformer architecture behind ChatGPT can learn to autocomplete a genetic sequence the way it autocompletes a sentence.
Missing Model Moat
This is where most people stop and assume the hard work is done. It is not. The best DNA model available, Roeding noted, returns roughly 99% useless output. A model that can write DNA is not a company. It is a starting point.
I have made this argument before, back in 2023, about how brittle OpenAI’s model moat is. My point then holds now. Selling a model does not create a moat. Training costs keep falling, open-source alternatives keep arriving, and Gartner now classifies foundation models as “strategic commodities.” The algorithm is necessary. It is almost never sufficient.
So if the model is not the advantage, what is?
The Moat Is The Loop
The answer is the one thing competitors cannot download. Earli’s process is the clearest illustration I have seen. A proprietary prediction model, what Roeding calls “Oracle”, proposes candidate genetic switches. A wet lab then tests up to 250,000 of them in a single batch, each tagged with a DNA barcode like a product at a checkout. The rare winners, the true outliers, get read out and fed back into the model. Every turn of the wheel makes the system smarter.
That is a data flywheel. Data to model to candidate to experiment, and back to data. I have written that this is exactly how AI works for healthcare in Healthcare’s AI Lesson: Autocomplete Isn’t Understanding. This approach will be identical in any other industry you care to name. The base model is rented from a handful of providers. The loop, fed by real-world data nobody else has, is owned. When everyone has the model, the moat is customization through data and workflows.
A world model is only as good as the physical reality it learns from, a point Yann LeCun keeps making. In biology, that reality is generated by pipettes and a vivarium full of mice, not scraped from the internet. The company that builds the machine to generate proprietary data, and the loop to learn from it, wins. The company that simply buys the model competes with everyone else who bought the same one.
Curing Cancer With AI, Responsibly
AI, Roeding said, is like suddenly having a bazooka in your toolbox. That means though, you still have to aim, and you still have to pull the trigger. The tool does not tell you which problem to solve. Earli’s genetic switches are not all machine-designed. They are a combination of human judgment and software, because pure brute force, in his words, would still be running a hundred years from now.
Curing cancer with AI is the clearest example of why that power demands responsibility. Earli is moving toward a Phase 1 human trial through monkey safety studies, manufacturing review, and the FDA, because in healthcare “seems right” is not good enough. I have said before that healthcare is one of the most thoughtfully regulated areas of our lives, and that is a feature, not a bug. As we make more of the physical world programmable, the discipline healthcare has been forced to build, safety first, real-world validation, and a human who is accountable for the outcome, is exactly the discipline every industry deploying AI will need to borrow. The model is the easy part. Owning your data, closing the loop, and aiming responsibly is the whole game.
