From Prediction to Preparation: How AI World Models Could Cook Up Dinner

Why a Tech Conference Got Me Thinking About Dinner

I was scrolling through my feed the other day, half-watching a webinar about AI and world models, when something clicked. The speaker was talking about how AI is moving from predicting the next word to predicting the next state of the world. And I thought: that's exactly what I do when I'm standing in my kitchen, staring at a half-empty fridge.

You see, cooking is all about prediction. You predict how long the onions need to caramelize. You predict whether that chicken breast is done. You predict that if you add too much salt, there's no going back. And if AI can learn to predict physical states—like what a pan looks like when it's perfectly seared—then maybe it can help us become better cooks.

This came from the AICon 2025 conference in Shenzhen, where HiDream.ai scientist Dr. Pan Yingwei talked about building a "native multimodal" model that treats images, text, video, and even actions as one unified stream. Instead of stitching together separate models for vision and language, this architecture learns from raw signals together. The goal is to predict "next states" rather than just "next tokens."

Now, I'm not a data scientist. But I am a home cook who burns garlic more often than I'd like to admit. And I couldn't help but wonder: what would a world model do in my kitchen?

The Kitchen Is a Physical System

Think about it. Cooking is one of the most physical things we do. It's not just about following a recipe—it's about feel, sight, smell, and timing. When you flip a pancake, you're predicting the trajectory of that pancake through the air. When you judge if a steak is medium-rare, you're reading the resistance of the meat.

Traditional AI models, the kind that power chatbots, are like cooks who only read recipes out loud. They know the words but can't actually taste or see the food. They might tell you to "cook for five minutes," but they can't tell you what the pan should look like at that moment.

Dr. Pan's talk focused on moving from token prediction to state prediction. That means teaching AI to understand the physical world—not just language. For cooking, that would be a game-changer (sorry, I know I'm not supposed to say that, but it's true).

What a World Model Could Do in Your Kitchen

Let's get concrete. Imagine you're making a stir-fry. You have a wok, some vegetables, a protein, and a sauce. A world model could watch the wok and predict when the vegetables are perfectly crisp-tender. It could notice that your stove runs hot and adjust the heat accordingly. It could even suggest when to add the garlic so it doesn't burn.

Here's a few things a cooking world model might help with:

• Real-time doneness detection: Point your phone at the steak, and the model predicts the internal temperature based on color and sizzle.
• Adaptive recipes: The recipe adjusts based on what you actually have, your equipment, and even the humidity that day.
• Skill coaching: It watches your knife skills and gives feedback on your grip and angle.
• Meal planning from leftovers: Scan your fridge, and the model suggests meals based on what's about to go bad.

These aren't just fantasy. The underlying tech—multimodal models that understand images, video, and actions—is already being built. The challenge is making it work in the messy, variable world of a home kitchen.

Why Current AI Fails at Cooking

Most AI cooking assistants today are glorified search engines. They pull up recipes and maybe answer questions, but they don't see your pan. They don't know that your oven runs 25 degrees hot. They don't understand that "medium heat" on your stove is different from mine.

That's because they're trained on text, not on physical states. They know the word "sizzle" but not the sound. They know "golden brown" but not the exact shade.

Dr. Pan's point is that we need AI to perceive the world directly—through raw pixels, audio, and even action data—rather than through language alone. A native multimodal model, like HiDream's UiT architecture, maps all these signals into one shared space. That means it can learn relationships between a visual change and a physical state change, which is exactly what cooking is about.

From Prediction to Action: The Next Big Leap

But predicting is only half the battle. The other half is acting. That's where robotics comes in. At AICon, there were talks about embodied AI and robot AGI—machines that can move and manipulate objects in the real world. If we ever get a robot that can cook, it will need a world model to predict the outcomes of its actions.

Picture a robot arm that can flip a pancake. To do that, it needs to know the pancake's trajectory, the pan's angle, and the right force. That's a state prediction problem. The same model that helps a robot navigate a room could help it navigate a recipe.

We're not there yet, but the pieces are coming together. The conference had sessions on AI infrastructure, agent safety, and high-value business scenarios. The tech is moving fast.

What This Means for Home Cooks (and Why You Should Care)

You might be thinking: "I just want to make a decent lasagna, not launch a Mars rover." Fair enough. But the same AI that powers autonomous vehicles could one day power your smart oven. Imagine a countertop appliance that knows exactly when your roast is done, adjusts the temperature automatically, and even videos the final plated dish to share with friends.

As these models get better at predicting physical states, they'll become more useful in everyday situations. Cooking is a perfect testing ground because it's universal, safe, and full of feedback. If AI can help you cook a perfect omelet, it can probably help you do a lot of other things.

The Human Touch Still Matters

I'll be honest: I don't want a robot to cook for me. I enjoy the process—the chopping, the sizzling, the tasting. But I'm all for having an intelligent assistant that helps me avoid mistakes and learn new techniques. A world model that can predict "the sauce is about to break" would have saved me many times.

AI won't replace your love of food, but it can make you a better cook. And that's something worth raising a glass to.

Final Thoughts: Keep an Eye on the Kitchen Tech

Next time you hear about AI conferences or "world models," don't just think about chatbots or self-driving cars. Think about your kitchen. The same technology that helps AI understand the physical world could help you understand your dinner.

So, the next time you're staring at a sad, wilted bunch of cilantro, imagine a future where your AI assistant says, "Let's make a cilantro-lime marinade for the chicken, and I'll watch the pan so you don't burn it." That's the future I want to cook in.

Until then, I'll keep practicing my flip—and hoping my own internal prediction model gets a little better.