carado.moe

tammy's blog about
AI alignment, utopia, anthropics, and more;

also on twitter, lesswrong, rss, and at

alignment is an optimization processes problem

i like to talk about AI alignment a lot, but the matter of alignment is really general to optimization processes in general.

here are some ways it applies to some other areas:

natural selection

natural selection is an optimization process that improves the survival and duplication of inheritable traits (genes) in living beings.

it is not intelligent: there is no agent involved in this process which is able to make decisions by looking ahead into the future at what consequences those decisions will have; with the possible exception of humans making rational decisions about what will maximize their amount offspring.

it is completely multipolar: basically no agents in this process (either genes themselves, or individuals or populations carrying those genes) have the ability to coordinate decisions with one another, since they're not even intelligent.

the default of natural selection is genes whose only purpose is to be better at duplicating themseles.

one way in which we've aligned this process is by breeding: by selecting the individuals we like best among, for example, crops, cattle, or dogs, we've been able to align the process of gene selection to respond to what we value rather than the default.

economics

the economy is an optimization process that improves economic efficiency.

it is intelligent: actors in the economy, ranging from individuals to states and giant conglomerates, have the ability to make intelligent decisions about the long term.

it is fairly multipolar: while they don't use it much, states do have overriding power over companies (they determine what's legal or not, after all), and also economic agents are able to coordinate to an extent using contracts and trusts. nevertheless, it is still largely multipolar, with agents overall competing with one another.

the default of economics is the optimization out of anything that doesn't generate maximally much resources: the optimizing out of people when they become the unoptimal form of labor because of automation, and the strip-mining of the universe to acquire ever more resources with which to create more machines to mine even more resources, and so on.

the way we align economics is through taxes, redistribution, and the like. redistribution like UBI aligns the economy to serves the demand of people, while tax externalities can align economic agents to take steps to preserve nice things, such as avoiding pollution.

electoralism

electoral representative democracy is an optimization process that improves voter satisfaction.

it is intelligent: the agents competing for the reward, here political parties, are able to make decisions about the future. some organisms even plan for the very long term, taking steps to improve their chances when they become parties, long before they do.

it is fairly multipolar: like economics, while parties can coordinate and ally with one another, they are still competing agents at the end of the day, with no central authority to guide them and solve coordination.

the default of electoralism is parties throwing all values under the bus to do whatever gets them and keeps them in office for as long as possible.

the way we align electoralism is by having universal sufferage on the one hand, which makes it that it is the population that parties must try to satisfy; and the various apparatii of liberal democracies (journalism and free press, public debate, education of the voting public, etc), which we'd hope would help that voting population determine which parties do indeed implement policies that satisfy their demand.