I spent six months training a neural network to design rocket fins. It was a disaster. The model converged. The loss curve looked beautiful. The output was complete aerodynamic garbage. Here's what I learned from that failure, and what it means for how we should think about Eric Schmidt's rocket company and the Mars mission. Step 1: Stop thinking of AI as a tool. Start thinking of it as a material. Most engineers approach AI like a hammer. You have a problem, you hit it with a model. This is wrong. When you're designing something for a Mars mission, you're not using AI to automate existing workflows. You're using it to discover properties of the design space that no human could find. The AI isn't replacing the engineer — it's terraforming the problem into something the engineer can work with. The practical tip: Before you write a single line of training code, ask yourself "What would I do if I had 10,000 years to brute-force this problem?" That's your target. The AI is just the time machine. Step 2: Build the data pipeline before you touch the model. This is the mistake I made with the rocket fins. I jumped straight to architecture search. I spent weeks tuning hyperparameters. The model learned to memorize my training data perfectly. Common pitfall: Your first dataset is always garbage. Not because you're bad at collecting data, but because you don't know what "good" looks like yet. The act of collecting and cleaning data teaches you the problem better than any model ever will. For the Mars mission work, the team likely spent 80% of their time on simulation environments and data curation. The actual model training was probably the easiest part. Step 3: Accept that your AI will be wrong in ways no human would be. This is the jagged intelligence problem. Your model will solve differential equations faster than any PhD, then fail to understand that a rocket needs to not explode on the launch pad. You need a human in the loop who understands both the physics and the model's failure modes. This is not a bug you can train away. It's a feature of the architecture. Practical tip: Build your validation set around edge cases, not average performance. The model will be fine on 95% of inputs. The 5% where it's catastrophically wrong are what kill missions. Step 4: Design for iteration speed, not initial accuracy. The eric-schmidt-rocket-company won the Mars mission because they could iterate faster than competitors, not because their initial models were better. Your first model should be terrible. If it's not, you're not learning fast enough. The goal is to get a working pipeline end-to-end, then improve each component one at a time. Common pitfall: Trying to optimize everything at once. You end up with a system where you can't tell which change caused which improvement. Isolate variables. Run controlled experiments. This is basic science, but everyone forgets when the AI is involved. Step 5: The real value is in the infrastructure, not the model. Everyone wants to talk about the AI breakthrough. Nobody wants to talk about the data pipeline, the simulation environment, the validation framework, the deployment infrastructure. But that's where the actual work happens. The model is the last 10% of the effort and the first 90% of the attention. For ai-in-space applications, the simulation environment is everything. If your simulator doesn't capture the right physics, your model will learn things that don't exist in the real world. This is how you get models that work perfectly in training and fail catastrophically at deployment. Step 6: Know when to stop. The hardest lesson: sometimes the model is good enough. I spent three months trying to squeeze another 1% accuracy out of my fin design model. In that time, I could have run 20 physical experiments that would have told me more about the actual problem. The AI is a tool for exploration. At some point, you have to stop exploring and start building. The Mars mission doesn't care about your SOTA benchmark. It cares about whether the rocket lands. The truth is, we're all still figuring this out. The people who win are the ones who treat AI as a material to shape, not a magic box to invoke. I'm still not great at this. But I'm better than I was before those six months of failure. And that's the whole point, isn't it?