Complementary Inputs to Utopia

one way to think about the work you could do

In economics¹, we often articulate production functions as follows:

Y = F(K, L)

where Y is output, F is the production function, K is capital, and L is labor.

Using this frame, we’re able to think about cases where we’re ‘capital-constrained’ or ‘labor-constrained’. It’s no good having 10 units of capital and 1 unit of labor if you saturate at a ratio of 5:1, and can’t use the marginal capital productively beyond that.

One thing we can do is suppose Y is Utopia, model our current production function F(A, B, C, …), and think: what is the gap? Does Y = F(A, B, C, …)? In which factors of production do we saturate fastest?

F(A) might put us in a bad position, where A = 5000 people working on building superintelligence ASAP,

whereas

F(A, B, C) might put us in a better position,

where

A = 5000 people people working on building superintelligence ASAP

B = 100 people proactively anticipating problems with A’s approach

C = 1000 people solving problems B identify and feeding the solutions into A’s design

You might already look at this and go “F(A, B, C) won’t bring utopia”. Great! The ‘production function’ frame allows you to think about what inputs are needed. It gives us a way to talk about the coarse-level bets various groups (e.g. ‘80 000 Hours’) are making when they inject some input (e.g. ‘technical AI safety researchers’) into the global production function. In what ways do they think that input will be complementary to the existing inputs?

Remember, there are some inputs that do nothing in the absence of complementary inputs, and require certain quantities of complementary inputs to work well. If you’re injecting an input like that into the global production function, think about articulating those complementary inputs and checking how realistic their co-occurrence is.

I view condu.it, a startup building thought-to-text brain-computer interfaces, as one group providing a complementary input: high-bandwidth tools for human expression. It’s possible fast, agentic AI systems might be regrettable in the absence of this input, endorseable when paired with it.

Combining this thinking with ‘backchaining’

Bidirectional search

I view this approach as a crucial complement to ‘backchaining’. My friend Nick has previously suggested to “think about the future you want, and work backwards from there to figure out what needs to be done.”

To me at this particular instant, backchaining-from-utopia feels nebulous and muddy. I more natively forward-chain-from-status-quo. And combining the two — “ah, this production function seems like it gets / doesn’t us where we want to go” — is a good way to incorporate the best of both approaches. Check that the forward-chain and back-chain meet up.

It’s possible F(A, B) may produce extinction where F(A, B, C) does not. Should you then work on C? What does F(current_world, C) look like in the absence of A and B?

You can start to ask yourself questions like these.

Related:

Why you shouldn’t build your career around existential risk (Guzey)

1. Similar thinking applies with chemical reactions.