Machine-learning

Most approaches to AI alignment incur upfront costs to the creator (an “alignment tax”). In expectation, startups will skimp on alignment in order to avoid the tax. “Alignment windfalls” are strategies which tend towards long-term public good and short-term benefits to a company. One suggestion for a high impact career is to discover, validate, and advocate for these alignment windfalls so as to accelerate their adoption.

Want to make Artificial Intelligence less likely to result in doom? Don't have a machine learning PhD? You can still help!

At one level: yes. At another: maybe.

A machine learning approach called CycleGAN has been shown to be effective at transforming an image from one style to another. Perhaps turning a photo into a Monet-style painting, or changing a picture of a horse into that of a zebra. However, I wanted to try to use the approach on a more challenging transformation: turning audio data into an aesthetically-pleasing image to illustrate the latent structure of music.

In The Moral Landscape, Sam Harris describes a framework which allows for strictly rational conversations about morality. In this post we explore the possibility of applying this framework to serve as the basis of ethical artificial intelligence, and show that although The Moral Landscape might work well for humans, it's a poor choice for machines.

Existential risk is hard to weigh up versus other priorities we see around us in the world. When we look at the data, however, I think it's clear that this deserves more of our attention, and is generally under-funded when compared to shorter-term charitable causes.

There are striking similarities between the training of a neural network and the way that we – as humans – form habits. How should that inform our thinking about the potential for this type of artificial intelligence, and what we should and shouldn't be asking of it?