Wiki Contributions


I have an idea for a possible utility function combination method. It basically normalizes based on how much utility is at stake in a random dictatorship. The combined utility function has these nice properties:

Pareto-optimality wrt all input utilities on all lotteries

Adding Pareto-dominated options (threats) does not change players' utilities

Invariance to utility scaling

Invariance to cloning every utility function

Threat resistance


The combination method goes like this:

X=list of utility functions to combine

dist(U)=worlds where random utility function is argmaxed with ties broken to argmax U

Step 1: For each U in X, U=U-expect(argmax U)

Step 2: For each U in X,

U=-U/U(null universe) if (expected value of U on dist(U))==0

U=-U/(expected value of U on dist(U)) otherwise

Step 3: U(final)=sum(U in X)


I designed this to be a CEV of voting utility maximizers, but the combination has multiple discontinuities. It does not need transferrable utility like the ROSE values, but it does not have nearly have as many nice properties as the ROSE values.

For the examples in this article, for each option only take the monetary value that goes last. log(amount after year)~0.79*log(amount now)+0.79 is the indifference curve. If U(now)=log(amount now), U(year)=(log(amount after year)-0.79)/0.79.

There is a hypothetical example of simulating a ridiculous number of humans typing text and seeing what fraction of those people that type out the current text type out each next token. In the limit, this approaches the best possible text predictor. This would simulate a lot of consciousness.

What if most people would develop superhuman intelligences in their brains without school but, because they have to write essays in school, these superhuman intelligences become aligned with writing essays fast? And no doomsday scenario has happened because they mostly cancel out each others' attempted manipulations and they couldn't program nanobots with their complicated utility functions. ChatGPT writes faster than us and has 20B parameters where humans have 100T parameters, but our neural activations are more noisy than floating-point arithmetic.

This is what I am wondering: Does this algorithm, when run, instantiate a subjective experience with the same moral relevance as the subjective experience that happens when mu opioids are released in biological brains?

‘By 'obvious to the algorithm' I mean that, to the algorithm, A is referenced with no intermediate computation. This is how pleasure and pain feel to me.  I do not believe all reinforcement learning algorithms feel pleasure/pain. A simple example that does not suffer is the Simpleton iterated prisoner’s dilemma strategy. I believe pain and pleasure are effective ways to implement reinforcement learning. In animals, reinforcement learning is called operant conditioning. See Reinforcement learning on a chicken  for a chicken that has experienced it. I do not know any algorithms to determine whether there is anything to be like a given program. I suspected this program experienced pleasure/pain because of its paralells to the neuroscience of pleasure and pain.

As this algorithm executes, the last and 2last variables become the program's last 2 outputs. L1's even indexes become the average input(reward?) given the number of ones the program outputted the last 2 times. I called L1's odd indexes 'confidence' because, as they get higher, the corresponding average reward changes less based on evidence. When L1 becomes entangled with the input generation process, the algorithm chooses which outputs make the inputs higher on average. That is why I called the input 'reward'. L2 reads off the average reward given the last 2 outputs. The algorithm chooses outputs that make the number of ones outputted closer to the number that has yielded the highest inputs in the past. This makes L2 analogous to 'wanting'.