Key Takeaways

  • AI for inorganic materials is fundamentally different from AI for biology: you can't text-encode a metal alloy's microstructure, processing, and supply chain like a DNA sequence. Joseph Krause, CEO of Radical AI, emphasizes that biology uses readily available text representations like “smile strings.”
  • Because of factors like alloy composition, supply chain, cost, and manufacturing processes (additive versus casting), a single AI model cannot “one-shot” predict a new material's real-world performance for applications like your iPhone or a Starship rocket.
  • The true “ground truth” for materials science exists only in the physical material itself. You have to make it, test it, and characterize it to see if it can withstand real-world applications, especially for high-stakes uses like a jet turbine on a 787.
  • Solving this means building “self-driving labs.” These are closed-loop systems that automate the entire process: running experiments, capturing novel data, and feeding that information directly back to an AI scientist.
  • This approach is CAPEX-intensive and demands AI models built specifically for sparse, high-stakes physical data, a challenge Radical AI is directly tackling to accelerate material discovery.

The Ground Truth Isn't Always in Your Dataset

For many founders, “AI” has become synonymous with large language models, image generation, or data analysis on existing datasets. Joseph Krause, CEO of Radical AI, argues this thinking hits a wall when you deal with physical materials. He points out a stark difference between fields like biology and materials science.

“If you look at bio or maybe small molecules as a more broad category, you look at selfies and smile strings, right? Which has been a big way to have those materials, those molecules in text. And then you can use that,” Krause explains. This textual representation allows AI models to learn from vast, structured databases.

But try applying that to a new metal alloy. How do you capture its microstructure, the specific manufacturing process (like additive manufacturing versus traditional casting), its supply chain, and its cost – all in a simple string? “You can’t. And this is what’s so hard is there is no one model that can one shot a new material that ends up in your iPhone or that ends up on Starship,” Krause asserts. The complexity of physical materials, and the endless variables that impact their real-world performance, simply defy easy digitization. The consequences of getting it wrong are also severe, especially for critical components like a jet turbine, where testing is not just rigorous but life-dependent.

Building Your Own Reality Engine

If the ground truth isn't found in a text file or an existing database, where is it? Krause’s answer is clear: “In materials, the ground truth is the material itself.” You have to actually make it. You have to test it. And then you need to see if it works in a real application. This isn't just about simulating; it's about physically creating and validating.

This insight led to Radical AI's core thesis: the necessity of building self-driving labs. These aren't just automated machines; they're intelligent, closed-loop systems. “You’re going to build this loop, this closed-loop system, what we call a self-driving lab, that can actually run those experiments, capture that data, and feed that information back to your AI scientist so that it can learn and actually predict materials that are relevant to industry,” Krause says.

This approach means investing in physical infrastructure – something many software-first founders might balk at. As Krause reflects, “When we started the company 2 and a half years ago, people would have thought we were crazy. That’s CAPEX intensive.” But for problems where the stakes are high, and the real-world variables are too complex for any existing dataset or simulation, building your own reality engine might be the only path to genuine innovation and competitive advantage.

What to Do With This

Take a hard look at your own domain: where does your ground truth truly lie? If you're building a physical product or dealing with complex real-world interactions (not just digital ones), identify a critical variable that can't be accurately modeled or predicted from existing data. Then, brainstorm what a “self-driving lab” equivalent would look like for your problem. Can you design a closed-loop system that automates real-world experimentation, captures the nuanced feedback, and feeds it directly into your decision-making AI? This week, map out three concrete steps you could take to build your own reality engine, however small, to generate true, actionable insight.