Dispatches from the heart of the GPU cluster. Notes on mathematics and technology, and the ways that they interweave with and influence our culture and community.

What is a signalling NaN?

NaNs are a kind of outcome you might get from a machine learning operation if something has gone wrong with your computation. Literally meaning Not a Number, NaNs can arise from errors such as division by zero, producing a number that is too large or too small to be represented by the machine you’re operating on, or any one of a number of problems that are harder to uncover.

We can distinguish signalling NaNs from quiet NaNs. If you generate a signalling NaN in your code, the resulting exception will be flagged as soon as its encountered during execution. Quiet NaNs are helpful in a different way: they propagate through the chain of computation, producing further NaNs, but allowing the end result to reached.

I started this blog to get clear on some of the debates around AI that struck me as involving a kind of category error, or at the very least, as deserving closer examination than they typically got. This blog is raising the flag: hey, what’s all this about? Conscious machines? Existential risk? What’s going on here?

Why subscribe?

Subscribe to get full access to the newsletter and publication archives.

Stay up-to-date

Never miss an update—every new post is sent directly to your email inbox. For a spam-free, ad-free reading experience, plus audio and community features, get the Substack app.