Stephen Casper, stephencasper.com

Summary

Adversaries have been effective ways of making deep learning systems fail. For example, in image-classifying networks, human-imperceptible perturbations to an input can result in a completely incorrect classification. Most research into adversaries, however, has not focused on challenging settings. Typically, they are created by making queries to a model whose internal parameters are known. However, these attacks can be easily precluded by making models black boxes.

Certain black box attack algorithms have been developed for these situations which work in practice, but there is a lack of theoretical models giving insight into how well they work. Toward…


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

Previously: Decision Theory II: Going meta

Remember: CDT=causal decision theory, EDT=evidential decision theory, and FDT=functional decision theory.

Now let’s shift gears and consider decision theory in context of a class of problems related to prisoners’ dilemmas. (Make sure you’re familiar with prisoners’ dilemmas before reading this.) I’ll call them threat dilemmas. A classic prisoners’ dilemma is symmetric between the agents, and they act simultaneously. But threat dilemmas are asymmetric and nonsimultaneous.

We’ll call one player the threatener and the other a threatenee. Both are self-interested. Suppose that the threatener has threatened to do something…


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

Previously: Decision Theory I: Understanding functional decision theory.

Remember: CDT=causal decision theory, EDT=evidential decision theory, and FDT=functional decision theory.

I hope that my first post on decision theory made you as excited about FDT as I am. It’s of great interest to people working in AI theory and practice because intelligent software may indeed live inside a world in which its code is open source or it has twins. FDT is the key to making such an agent fare well in Newcombian situations. That’s extremely powerful, but sadly, it’s not universally optimal. But wait…


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

What is decision theory?

Decision theory is what helps us make decisions. If this seems boring or pedantic, hold onto your horses.

An assumption that might seem out of place

I won’t focus on free will in this post, partially because I don’t want to, and partially because I think that this debate is unproductive and mostly semantic anyway. …


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

Jean Piaget, a University of Geneva psychologist, conducted a pretty interesting experiment in 1952. He showed children a diorama of three mountains with a doll sitting next to it. Then he showed these children pictures of the mountains from different perspectives and asked them to identify the one that showed them from the doll’s point of view. …


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

It took me a few years of being in the Effective Altruism community to understand this rift. It’s subtle, but it gets right at the heart of what morality even is. First though, let me give some background. I’ll give a few operational definitions here, and you may be able to find contexts where people use them differently, but indulge me.

Moral Realism

In a nutshell, moral realism says that morality is something to be discovered while moral nonrealism says that it is something to be invented or defined. As an additional bit of nuance, as…


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

Something really interesting to me is that utilitarianism can be summed up very well in 4 words as “the greatest happiness principle,” but at the same time, there’s a trove of different flavors, twists, variants, and deviants to it. One of those types (actually more than one) is negative utilitarianism. Different negative utilitarian theories in one way or another place more emphasis on the moral importance of reducing suffering than promoting happiness. I think it’s hard to defend, but a lot of modern utilitarians and Effective Altruists (EAs) believe in it. …


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

I’m not planning on dying anytime soon, but I thought I’d make a list of some requests in case I don’t end up achieving trans/post-human immortality. If you are someone responsible for me after my death who reads this and ignores these requests, and if I can become a ghost, I will haunt you. The severity of which will be disproportional (I practice updateless decision theory).

If I am ever in a vegetative or minimally conscious state and unable to make choices for myself, and if there is less than a 10% chance that…


Stephen Casper, thestephencasper@gmail.com

Euconoclastic blog series

The famous checker-shadow illusion.

This is a classic optical illusion. We see a very unsuspecting-looking photo of a green cylinder casting a shadow on a checkerboard. And interestingly, squares A and B are the same color.

Surprising! But it’s not too hard to understand. First, we can think for a moment and maybe block out part of the screen with our thumbs to confirm that they are indeed the same shade. We can also make a version of the photo with a grey bridge between A and B. …


Euconoclastic blog series

Stephen Casper, thestephencasper@gmail.com

Out of everything that I’ve ever come across, here are the six things that I think are the best (all are short). For each, I’ll share a few thoughts which are NOT meant to be a substitute for actually reading/watching them. One quick note: I know that the authors/speakers featured here aren’t very diverse, and I wish they were more so, but I think their messages speak for themselves. Let’s go!

1. Same Deere: Four Ideas You Already Agree with (That Mean You’re Probably on Board with Effective Altruism)

Spoiler alert: the four ideas are that (1) it’s important to help others, (2) people are equal, (3) helping more is better than…

Stephen Casper

Euconoclastic (ajd.) \yu̇-ˈkä-nə,-klast-ic\: iconoclastic in a good and virtuous way. Find me at stephencasper.com.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store