Training GPT-4 reportedly consumed enough energy to power thousands of homes for a year.

This isn't a fun fact — it's an engineering problem we need to solve.

The carbon footprint of AI is real and growing. And as engineers, we can make choices that reduce it without sacrificing capability.

Practical steps for more sustainable AI:

Choose efficient models. A quantized 7B model has a fraction of the carbon footprint of a 70B model. If the smaller model works, use it.

Optimize training. Mixed precision training, gradient checkpointing, and efficient data loading reduce training time (and energy) by 30-50%.

Use carbon-aware computing. Train during off-peak hours when the grid has more renewable energy. Some cloud providers show real-time carbon intensity.

Cache aggressively. Every cached response is an LLM call that didn't happen.

Evaluate before training. Make sure your experiment is well-designed before burning GPU hours. A 10-minute review of your training configuration can prevent 10 hours of wasted compute.

I'm not suggesting we stop building AI. I'm suggesting we build it efficiently. Efficiency and sustainability are often the same goal — both reduce waste.

The engineers who build powerful AI systems with minimal resource consumption aren't just environmentally responsible. They're better engineers. Because efficiency is a skill, and it matters.