Before anyone says "ugh, AI!" it's important to look past the headlines to understand what's actually being discussed here.

I wish we could get rid of the stupid term "artificial intelligence" and separate the technologies themselves from the companies abusing them. It's not "intelligent" and not every form of machine learning takes the form of sycophantic chatbots being forced down everyone's throats by an Orwellian tech oligopoly.

Weather forecasting is an area where learning models actually make sense. We already have weather models which is exactly what the meteorologists already use, however these are very manual and static models which an enormous amount of manual work went into creating, and we only have a handful of reliable models to use, and even those reliable models are far from perfect and often disagree in profound ways. And just like today's learning models, they take large amounts of processing power to compute. They are tools in the meteorologist's toolbox, used to guide them without replacing their own intelligence and interpretations. Learning models for weather are not just a buzzword, they take advantage of modern technologies to actually potentially be significantly more efficient than the older, manual models and the fact that they are capable of both learning and rapidly iterating is potentially very helpful in a world where the climate is rapidly changing thanks to greenhouse gas emissions. They don't replace our existing weather models, but they may be able to adapt faster or provide alternative projections that meteorologists may find genuinely useful when the traditional models are not working well.

Environment Canada is a highly data-driven organization of capable experts and have what I would call a very strong track record. They are very unlikely to be using this technology irresponsibly. Give them some credit.

I thought AI was only for entertainment?!

It plans to launch a hybrid model this spring that uses both AI and traditional forecasting and says the combination of the two will lead to more accurate predictions.

Scientists for years have been using "AI" neural network models for approximating differential equations and parts of complex models. The transformer architecture is just the next step in that, which allows us to scale up training and get more capabilities out of neural nets.

As an example, Google just released TimesFM, which is an AI model for timeseries which is pretty cool

The reality is, when you're working with large scale physical models there are differential equations involved which no human can properly parameterize and implement based on raw theory alone. These models let us use large amounts of data to infill that step.