Aardvark Daily

Please visit the sponsor!

What if AI develops psychosis?

One of the biggest users of supercomputers are weather forecasters.

It's only by using some incredibly complex computer models that forecasters can predict the weather days or even weeks ahead.

In predicting the weather there are a huge number of variables that must be combined through modeling software that makes a massive demand of the computers on which that code runs.

These models are constantly being revised and updated as our understanding of exactly how "weather" works and they're still prone to errors and the unexpected effects of chaos theory.

This week however, a breakthrough was announced.

It seems that artificial intelligence (AI) systems have proven themselves to be more effective and more efficient than traditional forecasting models at predicting extreme weather events.

Apparently Google's DeepMind AI system has proven to be significantly better than traditional forecasting models according to this statement from Google themselves.

Not only are its forecasts more accurate but they are generated in a tiny fraction of the time required by the traditional supercomputer-based modeling methods.

Google doesn't quite tell the whole story however, and once you dig a bit deeper it seems that while the Graphcast system running on DeepMind may be very impressive at forecasting the behaviour of extreme weather systems such as cyclones, tropical storms etc -- but it's not quite ready to replace regular forecasting systems for general (more benign) weather patterns.

Forecasters are saying that their supercomputers and weather models are not about to be replaced... just yet.

However, Google seems very optimistic that Graphcast will eventually replace the standard computer modeling and provide a superior forecasting service across the board.

In my opinion, this is perhaps one of the most eye-opening aspects of the power of AI.

While it's fun and entertaining to use AI to create images and even videos or to maybe interact in a way that passes the Touring Test, the practical application of the technology in a way that exceeds the abilities of decades of hard work by "real people" is impressive.

Not only does this have the potential to save lives through more accurate forecasting but because the results are generated so much more quickly I'm pretty sure it's only a matter of time before Google starts touting the "evironmental benefits" of decomissioning power-hungry supercomputers in favour of its own AI framework.

Given that Graphcast's AI has been trained on decades of previous weather patterns it will be very interesting to see how it handles the effects of accelerated climate change and whether it'll adapt quickly enough to remain accurate.

It's looking increasingly as if the AI is going to be an incredibly valuable tool in future and perhaps the only thing we should worry about is whether maybe we will become too reliant on it. What do we do if/when those traditional weather models are decommissioned and, perhaps years later, we discover that AI systems are vulnerable to something as simple as a cyber-psychosis that effectively renders them useless overnight?

I certainly hope there are people considering these possibilities. After all, even though we have built these systems we still don't totally understand what's going on in those neural networks once they are "trained". The possibility of human-like mental illness must surely be a possibility?

Carpe Diem folks!

Please visit the sponsor!

PERMALINK to this column

Rank This Aardvark Page

Change Font