There are some broad conceptual similarities between the following: free market economy vs command economy letting a student find an answer on their own vs teaching them the answer directly letting employees do their thing vs micromanagement reinforcement learning vs fine tuning plasticity vs stability doing something naturally vs doing something via willpower
Notice how in each comparison, the second method privileges already-known solutions over emergent (i.e. mysteriously appearing) solutions. I don’t know a name for these, so I’ll call them **bottom-up** vs **top-down** methods respectively. I (w/help of Claude) managed to find some recurring patterns when analyzing bottom-up vs top-down methods:
1) Bottom-up methods tend to be better at handling system growth. Examples: Children’s brains tend to be more plastic, which I would guess helps them adjust to bigger brains and learning new things. A city that grows in a decentralized way is better at adapting to population growth than one with rigid central planning.
2) Top-down methods become infeasible when the ability of a central system is limited, and bottom-up methods become infeasible when stakes are high. Examples: A government doesn’t have all the knowledge a market does, but you can’t hand responsibility of AI x-risk to a market. Social skills are very hard to replicate via reasoning and willpower, and most people are better off doing things naturally, but in a crisis, sticking to whatever feels right is a terrible idea.
3) Bottom-up methods tend to give rise to clever but less stable proxy gaming, while top-down methods tend to give rise to powerful but less smart proxy gaming. Example: Companies in free markets can develop clever but constrained strategies, while command economies can wield a lot of power but in less sophisticated ways.
4) Bottom-up methods are more vulnerable to inappropriate system change, while top-down methods are more vulnerable to inappropriate system stability. Examples: Plastic neural networks are more vulnerable to inappropriate retroactive interference, while stable neural networks are more vulnerable to inappropriate proactive interference. Long-term democracies are more vulnerable to a new bad leader coming along, while long-term absolute governments are more vulnerable to sticking with bad leader.
5) Often, incentives for misalignment are different in bottom-up and top-down systems. (I won’t provide examples for this one.)
There are some broad conceptual similarities between the following:
free market economy vs command economy
letting a student find an answer on their own vs teaching them the answer directly
letting employees do their thing vs micromanagement
reinforcement learning vs fine tuning
plasticity vs stability
doing something naturally vs doing something via willpower
Notice how in each comparison, the second method privileges already-known solutions over emergent (i.e. mysteriously appearing) solutions. I don’t know a name for these, so I’ll call them **bottom-up** vs **top-down** methods respectively.
I (w/help of Claude) managed to find some recurring patterns when analyzing bottom-up vs top-down methods:
1) Bottom-up methods tend to be better at handling system growth.
Examples: Children’s brains tend to be more plastic, which I would guess helps them adjust to bigger brains and learning new things. A city that grows in a decentralized way is better at adapting to population growth than one with rigid central planning.
2) Top-down methods become infeasible when the ability of a central system is limited, and bottom-up methods become infeasible when stakes are high.
Examples: A government doesn’t have all the knowledge a market does, but you can’t hand responsibility of AI x-risk to a market. Social skills are very hard to replicate via reasoning and willpower, and most people are better off doing things naturally, but in a crisis, sticking to whatever feels right is a terrible idea.
3) Bottom-up methods tend to give rise to clever but less stable proxy gaming, while top-down methods tend to give rise to powerful but less smart proxy gaming.
Example: Companies in free markets can develop clever but constrained strategies, while command economies can wield a lot of power but in less sophisticated ways.
4) Bottom-up methods are more vulnerable to inappropriate system change, while top-down methods are more vulnerable to inappropriate system stability.
Examples: Plastic neural networks are more vulnerable to inappropriate retroactive interference, while stable neural networks are more vulnerable to inappropriate proactive interference. Long-term democracies are more vulnerable to a new bad leader coming along, while long-term absolute governments are more vulnerable to sticking with bad leader.
5) Often, incentives for misalignment are different in bottom-up and top-down systems.
(I won’t provide examples for this one.)