This post is a companion piece to a forthcoming paper. This work was done as part of MATS 7.0 & 7.1. Abstract We explore how LLMs’ awareness of their own capabilities affects their ability to acquire resources, sandbag an evaluation, and escape AI control. We quantify LLMs' self-awareness of capability...
I present a macroeconomic model of a competitive economy that, counterintuitively, predicts that increases in the productivity of automation technology will decrease economic production. In the model, the decrease in production occurs over an intermediate range of automation productivities, and once automation productivity is high enough that all labor is...
Good point about workers preferring negative L when they have a high passive income!
I think you're right about using a CES utility function. My guess would be that for an elasticity of substitution (between consumption and leisure) less than 1, the equilibrium would go down as goes up (and vice versa for an elasticity greater than 1). I actually can't decide whether I'd predict most people's elasticity to be more or less than 1. A complicating factor, which you mentioned in your footnote 3, is that people's preferences for wealth are highly influenced by the people around them (both by the wealth of those around them, and by the culture). I've always wanted to explore the economic consequences of this--my intuition is that it messes up all the classic "welfare theorems" in economics.
You're right that the result could change if there were heterogeneous agents with an uneven distribution of capital. However, the model above can be reformulated to include an uneven distribution of capital, and the result remains mostly the same. Specifically, suppose that instead of the representative agent, there were agents with identical preferences (the same utility function as above) but different capital allocations . Then, total labor supply (the sum of all agents' labor supplies) would be
This leads to an equivalent result for total labor supplied: . So the model still predicts that total labor supplied is independent of the capital stock, regardless of how capital is distributed. But, how much each agent contributes to that total labor supply does depend on the distribution of capital. If the distribution of capital changes, agents who gain a larger share of total capital will work less than before, and those whose share of total capital falls will work more than before.
Interesting question! As @harsimony alluded to, there are opposing effects here: increasing returns to capital may reduce labor supplied as people shift toward leisure, but increasingly scarce labor pushes wages up and pushes interest rates down, causing people to shift away from leisure toward labor. Let’s see how these competing effects play out in a simple macroeconomic model. The TLDR is that, in the simple model below, the two effects exactly cancel.
Suppose there is a competitive economy with production function , and a representative agent with utility function where is consumption, is labor, is capital, is the wage, and is the rental rate of capital (a.k.a. the interest rate). Note that the representative agent values both consumption and leisure (where... (read more)
Agreed that successful sandbagging will likely require Schelling coordination, and my guess is that this will be extremely difficult for models to pull off! Great to see that you're investigating this topic.
This post is a companion piece to a forthcoming paper.
This work was done as part of MATS 7.0 & 7.1.
We explore how LLMs’ awareness of their own capabilities affects their ability to acquire resources, sandbag an evaluation, and escape AI control. We quantify LLMs' self-awareness of capability as their accuracy in predicting their success on python coding tasks before attempting the tasks. We find that current LLMs are quite poor at making these predictions, both because they are overconfident and because they have low discriminatory power in distinguishing tasks they are capable of from those they are not. Nevertheless, current LLMs’ predictions are good enough to non-trivially impact risk in our modeled... (read 5265 more words →)
I present a macroeconomic model of a competitive economy that, counterintuitively, predicts that increases in the productivity of automation technology will decrease economic production. In the model, the decrease in production occurs over an intermediate range of automation productivities, and once automation productivity is high enough that all labor is replaced, production begins to rise.
Most discussions of the economic impacts of AI assume that automation will increase total economic output. Even though wages may decline, the common assumption is that total production will rise. However, in this post I present a model that predicts otherwise. The model follows a standard macroeconomic framework and invokes plausible assumptions. Starting with a simple general equilibrium... (read 1357 more words →)
Definitely agree! The "don't worry about other people's stuff" argument gets thrown around a lot, and is often assumed to be equivalent to "don't be envious", but I think that argument actually contains a logical mistake. Suppose person A doesn't care directly about their relative wealth (i.e. their utility function is independent of their relative wealth), but their utility function depends on their interactions with other people (such as friendships, job interviews, etc) and those other people interact with person A in a way that depends on person A's relative wealth. Then, it can be instrumentally useful for person A to increase their relative wealth despite their utility function being independent of relative wealth! So person A has no envy, but their utility is (indirectly) affected by their relative wealth.