AI’s Dirty Little Secret

"Alexa, I need emergency medical treatment" "I've added emergency medical treatment to your shopping list"

Stop all trains to prevent train crashes is the same logic like cancelled trains are not delayed. I think the AI learned from Deutsche Bahn (German railway company).

As someone who works in machine learning research, I find this video a bit surprising, since 90% of what we are doing is developing approaches to fight overfitting when using big models. So we do very well know why NNs don’t overfit: stochastic/mini batch gradient descent, momentum based optimizers, norm-regularization, early stopping, batch normalization, dropout, gradient clipping, data augmentation, model pruning, and many, many more very clever ideas…

"It's like a teenager, but without the eye-rolling." 🤣

This is a story I read from a magazine long time ago: In distant future, scientists create a super complex AI computer to solve energy crisis that is plaguing mankind. So much time, resources and money was put into creating this super AI computer. Then the machine is complete and the scientists nervously turn on the machine for the first time. Then the lead scientist asks, "Almighty Super Computer, how do we resolve our current energy crisis?" Computer replies, "Turn me off."

Human: Stop all Wars AI: Are you sure?

One of my favorites is that in skin cancer pictures, an AI came to the conclusion that rulers cause cancer (because the malignant ones were measured in the majority of pictures)

Man, I went out with a model. I never could predict what was going to happen next

And people who come to emergency medical departments by car tend toward better outcomes than those who arrive by ambulance. We should likely stop using ambulances.

1:38 "A strange game. The only winning move is not to play."

The “you can’t crash a train that never leaves the station” answer sounded kinda like a glorious StackOverflow response.

This sounds like the Dunning–Kruger effect for AI.

I come here every day just to listen to how Sabine says: "No one knows"

It occurs when a model is too specialized to the training data and performs poorly on new, unseen data. This can happen when a model is too complex, has too many parameters relative to the amount of training data, or when the training data itself contains a lot of noise or irrelevant information "The man with a hammer analogy perfectly captures the essence of the overfitting issue in AI. Just as the man with a hammer sees every problem as a nail, an overfitting model sees every pattern in the training data as crucial, even if it's just noise. It becomes so specialized to the training data that it loses sight of the bigger picture, much like the man who tries to hammer every problem into submission. As a result, the model performs exceptionally well on the training data but fails miserably when faced with new, unseen data. This is because it has become too good at fitting the noise and irrelevant details in the training data, rather than learning the underlying patterns that truly matter. Just as the man with a hammer needs to learn to put down his trusty tool and approach problems with a more nuanced perspective, an overfitting model needs to be reined in through regularization and other techniques to prevent it from becoming too specialized and losing its ability to generalize.

My boss always gave me the correct answer to any question I ever asked him.......when I questioned a decision he made the answer was always "you can do it or get another job."

This really hits home for me, having done a lot of multi-variable regression back in the 80’s

Haha. The "stop all the trains" solution is a mirror of the old movie "Colossus, the Forbin Project." To prevent human race from hurting itself, enslave it.

You’ve just put your finger on the main research topic of my career, Sabine. The “reason” they work unexpectedly well is because at their core they are doing weak constraint relaxation, and WCR just has this behavior as an emergent property. I know, that sounds circular. But it’s a tremendously subtle issue, and I’ve written papers about it (just search for my name and ‘publications’) and I’ve also been trying to get people to understand it since around 1989, with virtually zero success.

Human: Stop All Wars! AI: I did, but nobody is reading the directions.

That graph is actually very easily explainable.