I am currently researching how large-language models reason in normal-form games. As part of this research I am empirically evaluating their capability to reason practically about simple normal-form-games such as Prisoner's Dilemma and Paper-Rock-Scissors. The wider motivation for this effort is to gain a better understanding of their reasoning and...
Epistemic status: My Best Guess (some preliminary empirical evidence- more experiments and critique required) This post introduces an early-draft working paper investigating the propensity of LLMs to cooperate and reciprocate in the Prisoner's Dilemma: https://sphelps.net/papers/llm-cooperation. I plan on submitting to arXiv after incorporating feedback from the community. An edited version...
@Dagon perhaps I should have place the emphasis on "transfer". The key thing is that we are able to reliably transfer ownership in exchange for renumeration and that the resource on which on goals are contingent at least needs to be excludable. If we cannot prevent arbitrary counter-parties consuming the resource in question without paying for it then we can't have a market for it.
@johnswentworth ok but we can achieve Pareto optimal allocations using central planning, but one wouldn't normally call this a market?
"And that’s the core concept of a market: a bunch of subsystems at pareto optimality with respect to a bunch of goals."
The other key property is that the subsystems are able to reliably and voluntarily exchange the resources that relate to their goals. This is not always the case, especially in biological settings, because there is not always a way to enforce contracts- e.g. there needs to be a mechanism to prevent counter-parties from reneging on deals.
The anonymous referees for our paper Economic Drivers of Biological Complexity came up with this concise summary:
"Markets can arise spontaneously whenever individuals are able to engage in voluntary exchange and when
they differ in their preferences and... (read more)
What is the epistemic status of this claim? e.g is it based on well established evidence beyond hearsay and personal opinion? How does it relate to other well-established cognitive biases such as the sunk-costs fallacy? Alternatively, are you conjecturing that there is a new kind of cognitive bias (that psychologists have not yet hypothesized) that causes people to persist with failing projects when it is irrational to do so? If so, how could it be experimentally tested?
If somebody is finding it difficult to move on from a failed project I would tend to suggest to them to "be mindful of the sunk-costs fallacy" rather then to "stare into the abyss".
https://www.lesswrong.com/tag/sunk-cost-fallacy
The main problems are the number of contracts and the relationship management problem. Once upon a time drawing up and enforcing the required number of contracts would have been prohibitevly expensive in terms of fees for lawyers. In the modern era Web 3.0 promised smart contracts to solve this kind of problem. But smart contracts don't solve the problem of incomplete contracts https://en.m.wikipedia.org/wiki/Incomplete_contracts, and this in itself can be seen as a transaction cost in the form of a risk premium. and so we are stuck with companies. In the theory of the firm companies are indeed a bit like socialist enclaves. Individuals give up some of... (read more)
The point is that all the people making cars in the company could, in principle, do the same job as self employed freelance contractors rather than as employees. Instead of a company you have lots of contracts between eg assembly line workers, salespeople and end customers, without any companies. The same number of cars could in theory be built by the same number of people in each case. The physical scenario would be identical in each case. The machinery would be identical in each case. But in the freelancer case you still have lots of people building cars but there is no invisible company to coordinate this activity instead you are relying the market.
Perhaps 'The Theory of the Firm'? The very existence of large companies is a puzzle if you are a naive believer in the power of free markets, because if the market is efficient then individuals can simply contract with other individuals through the market to achieve their desired inputs and outputs and there is no economic advantage to amassing individuals into higher-level entities called companies. The reason this doesn't this work is because of transaction costs. An example transaction cost could be the time invested in finding partners to make contracts with for all your inputs and then forming and enforcing contracts. In the 'the theory of firm' the insight is that... (read more)
This problem has been studied extensively by economists within the field of organizational economics, and is called the principal-agent problem (Jensen and Meckling, 1976). In a principal-agent problem a principal (e.g. firm) hires an agent to perform some task. Both the principal and the agent are assumed to be rational expected utility maximisers, but the utility function of the agent and that of the principal are not necessarily aligned, and there is an asymmetry in the information available to each party. This situation can lead the agent into taking actions that are not in the principal's interests.
As the OP suggests, incentive schemes, such as performance-conditional payments, can be used to... (read more)
I am currently researching how large-language models reason in normal-form games. As part of this research I am empirically evaluating their capability to reason practically about simple normal-form-games such as Prisoner's Dilemma and Paper-Rock-Scissors. The wider motivation for this effort is to gain a better understanding of their reasoning and behavior in a wide range of social dilemmas for the purposes of empirically evaluating their safety (arXiv:2305.07970).
OpenAI have a github repo - evals - for benchmarks for their models. There are some existing benchmarks for games such as chess, but as far as I can tell there are no benchmarks for normal-form games, and so I am submitting a pull-request.
The submission process... (read 2278 more words →)
Epistemic status: My Best Guess (some preliminary empirical evidence- more experiments and critique required)
This post introduces an early-draft working paper investigating the propensity of LLMs to cooperate and reciprocate in the Prisoner's Dilemma: https://sphelps.net/papers/llm-cooperation. I plan on submitting to arXiv after incorporating feedback from the community. An edited version of the paper's introduction is given below.
TLDR; there has been some discussion on applying ideas from game theory to bring the incentives of AI agents into alignment with human goals, but as far as I know there has been little empirical investigation of how actual AIs behave in cooperative tasks when faced with e.g. conditional altruism (though happy to be corrected). We... (read 1007 more words →)
yes if you take a particular side in the socialist calculation debate then a centrally-planned economy is isomorphic with "a market". And yes, if you ignore the Myerson–Satterthwaite theorem (and other impossibility results) then we can sweep aside the fact that most real-world "market" mechanisms do not yield Pareto-optimal allocations in practice :-)