Deep state, SJWs, discrimination, and reinforcement learning

Or; How to optimize a Facebook and win at poker.

AltDeep is a newsletter focused on mental models and microtrend-spotting in machine learning, and AI.

The following thought experiment would put you in mind for some of the things that are coming down the pipeline of AI (particularly agent models like reinforcement learning) and their effects on society.

The Deep State

Alice tells John are colleagues in a "deep state" executive government agency. Alice and John are debating what new government interventions to make. Alice firmly believes that government interventions that actively discriminate against a group of people are unjust and should be avoided. Further still, Alice says their government agency should undertake interventions should that alleviate systematic discrimination in society.

Later, the conversation switches the topic to an episode Alice had at work. She describes an incident where a superior overtly maltreated her because of her gender and ethnicity. The superior went unpunished and is likely to continue this behavior.

John says that Alice should take it up with HR, even go on social media and publicly shame the company for this behavior.

Alice demurs. She says that to her career goals, she must play the game with strategic nuance. Alice says that the world is full of people who behave this way and to survive and thrive, one must learn to deal with them with subtly and strategy. She says she must not allow herself to think of herself as a victim. She says that a good poker player doesn't fold and complain when she gets dealt a bad hand, instead she employs tactics that will get her the win despite that hand.

On agents, their actions, and games

Here, "policy" means a mechanism that tells you what action to undertake, given the state of the world, and some outcome you want to achieve.

At first, Alice seems to advocate for actions that directly counter discrimination - call this the liberal left "SJW policy." But in the next subject of conversation, Alice appears to argue for a more Sun Tzu-style "subdue the enemy without fighting" strategy - call this the conservative right "bootstrap policy" (as in finger-wagging with "take personal responsibility and pull yourself up by the bootstraps!").

Is there a contradiction here? How can Alice argue both for an SJW policy and the Bootstrap policy?

I suspect the answer is no, because these are two different "games," with different game dynamics.

In the first example, the agent is treating society much as a data scientist on Facebook's experimentation team treats the Facebook ecosystem. Both society and Facebook's ecosystem are phenomena. The agent has a set of actions (social policies, possible changes to Facebook's features) that can change the behavior of that phenomenon. The policymaker/data scientist is playing a game where the objective is to drive some desired outcome, like social justice/time spent on Facebook.

There is a certain dynamic missing from the Facebook optimization game. The Facebook ecosystem does not judge the agent's intentions in taking actions that change the app's features and then respond with behavior intended to affect that agent's future actions.

However, that is what is happening in the second example. Borrowing Alice's poker analogy, the agent is more like a player in a game of cards, trying to select actions that maximize winnings. However, the opponents know Alice is trying to maximize winnings and undertake actions intended to minimize her winnings. Alice knows that her opponents are doing this. So Alice's "Bootstrap policy" factors this into the actions she undertakes.

“Opposition is all you need”

Do you agree? One thing missing from my explanation is scale. In the first example, the actions are macro-scale (affecting society), while in the second, the actions are micro-scale (affecting a small group of competing agents in the workplace). I think scale doesn't matter; it's only a proxy for whether or not there are opponents. One might also reason that in the second case, Alice has less power. I don't think this matters either. It only matters that, given the actions Alice has available to her, the probability of Alice winning the game is high enough to make it worth it for her to play.

Let me know what you think. If you have a colleague who thinks they know a thing or two, forward this along.