<aside> 💪🏻

Current version is published here

https://meaningalignment.substack.com/p/market-intermediaries-a-post-agi

</aside>

Older text and notes

Not only do these approaches centralize power, they also don’t actually re-align markets: markets continue to pull the wrong way, patched by pledges, regulations, or redistributions.

• Cheap assessments. Outcome-based contracting requires one or both parties to estimate outcomes, assess and audit those outcomes regularly, and insure against errors in their estimates. In the airline industry, these costs are small compared to the overall expense of owning and operating an aircraft. For outcome-based contracting to be affordable, these transaction costs must be relatively low.
• Symmetric information. For both sides to converge toward an optimal contract, neither can have a decisive advantage. With cloud compute, for instance, providers often have an information advantage. That drives towards metered, but not outcome-based contracts.

Powerful AI changes all three of these factors: it can extend the assessment of benefit into quantitative domains; it can make assessment much cheaper; and it can occupy a third-party, auditable position to eliminate information asymmetry in contract design. Outcome-based contracting can thus be extended to many more markets—even in domains with fuzzy assessment needs, where measuring ‘good outcomes’ requires qualitative interviews.

Components of a Market Intermediary

To illustrate the components of a market intermediary, let's imagine building one for car repair, a classic market with high information asymmetry.

A market intermediary needs to cover four functions:

• Assessment. First, it needs to track the kinds of benefit that consumers are asking for and which providers think they can provide, and then track whether that benefit works out in the end. In the case of car repair, the intermediary would need to know how long somebody wants their car to continue running well, whether they care about visual polish or just engine health, etc. And it would need to either interview the car owner periodically, to see how the car‘s doing, or otherwise access data about ongoing car health.

  There are two important challenges here. First, we must avoid using proxy measures that diverge from the real benefit the market provides. All important elements of benefit must be captured, or they will be stripped from the market. For instance, if repair shops turn out to provide an important role in educating drivers, that information needs to be gathered and tracked by the intermediary.

  Secondly, the intermediary must be robust against incentives that providers will have to game the system. For instance, auto repair shops may try to pay their customers to say their car is running better than it is. Intermediaries will need to detect and avoid this kind of fraud.

• An intermediary’s matching function finds provider/consumer pairs which it believes would provide benefit with an acceptable price. This requires additional data: wait times, locations, and hours of different providers, how good they are at visual polish vs engine health, etc.

  We can consider the intermediary's job as finding the actual cost of each marginal unit of benefit, for each consumer. In our car repair case, how much would it cost to make each car run one year longer, or look a bit better, etc? To make this possible, providers can publish a price list and consumers can say what they’d pay for a unit of benefit[3].

• In optimization, this list of matches can be further ranked and pruned. In most markets, it doesn’t make sense to pick the cheapest proposed transaction, but rather to find something optimal on the frontier between benefit/cost trade-offs.

  We are excited to try using ML to do this without making benefit (and benefit uncertainty) into a single number.

Markets for early tests

1. Local events and meaning. Our earliest test will be with small-scale local events in Berlin. The idea is to connect consumers with yoga teachers and studios, dance events, etc based on their values in the relevant domain. We hypotheses that (1) this is an easy market in which to reorganize suppliers, interview users, assess contracts, and shift incentives; (2) the results will be relevant to issues of labor irrelevancy and over-dependence on AI social simulations.
2. Compute for science / values of science. A good second test would be to find a compute provider like https://sfcompute.com/ and get them to agree to set aside a portion of their cloud for scientific applications, and then get a science lab like https://arcinstitute.org/ to agree to have the MI interview all their scientists about the values of open science. Then the compute provider is paid according to improvements in these scientific ‘immeasurables’. Proof that such a contract worked out could be leveraged to get larger compute providers and other suppliers to commit to similar terms, eventually leading to shifts in provisioning for government-funded research.
3. AI assistants for flourishing. A reseller experiment for AI assistants like Claude and ChatGPT where the supplier is paid according to users’ flourishing.
4. Local public infra bids. In many jurisdictions, real estate developers are incentivized to pitch building projects like playgrounds and parks which they claim will have various social outcomes. These social outcomes don’t make it into the contracts. This seems like a ripe domain for MIs.

List of hypotheses

Initial Hypotheses - to test in our first run (see below)