

Is there more detail on the process beyond what’s in the blog post? I could see a scenario in which the training data was just generated by running multiple playthroughs on a $500,000 GPU at impossible quality, creating a copy of what that would look like on a mid-range GPU, and then training a model. I’m not sure I would object to that.














Natural gas is used to produce hydrogen, which is then used in the Haber-Bosch process to produce ammonia from nitrogen in the atmosphere. Only about 6% of natural gas is used to produce hydrogen, so even if the price were to rise substantially, we could divert natural gas from other uses and have plenty for making ammonia. We also have other ways of producing hydrogen, it’s just that natural gas is more established.
PEM electrolyzers paired with cheap solar in countries with high insolation can now produce hydrogen for less than the cost of natural gas, but we’re only recently starting to see the construction of the large-scale green ammonia plants needed to accomplish this. Egypt is currently constructing a 100-MW green ammonia plant powered by solar energy. Even if you didn’t have enough PEM eletrolyzers you could still just pass current through some salt water and produce hydrogen, albeit much less efficiently.
It’s not going to be a catastrophic issue.