Allocating Attention in Claimspace

Allocation of scarce compute resources is a first-order problem for scaling and maintaining the claim graph as a technical matter.

Just to begin with, there is a direct tradeoff between ingesting more claims, assessing more claims, and being more thorough (higher effort, better model) on each assessment.

Claims vary in importance by many orders of magnitude. They also range widely in how difficult they are to assess.

There exist important claims (e.g. the Pythagorean theorem) that are nevertheless very easy to assess, and there are unimportant claims that one could work on for a lifetime and never make real progress.

There are a functionally infinite number of claims. Reality has a surprising amount of detail, and you can just keep raising the resolution to find fractally more subclaims under every claim.

The optimal allocation of attention to the median claim is zero. Only claims that could be relevant to a live discourse should be mapped in the first place.

On the other hand, for claims that are important, neglected, and tractable, it can be worth spending a lot of resources getting them right. For example, with certain mathematical conjectures that would be a big deal if proven/disproven and seem within reach, it may make sense for the claim steward to use the best tools and models available and try many strategies to try to actually solve the problem, even if it costs tens of thousands of dollars to do so. For promising claims in cancer research, it may literally be worth funding physical lab work via one of the AI-run robotic biochemistry labs that will soon be opening to figure it out or fund human scientists working on the problem. That is of course far down the road, once the much lower-hanging fruit of just mapping the claim graph and assessing the current evidence using LLMs has been picked.

While resources are small, I'm also thinking about where to spend resources to bring value to users, get traction, and kick off a viral loop.

I started with lab leak, black holes, and eggs because that was the assignment, so under the circumstances, the highest expected value place to spend a few hundred dollars in tokens was there.

The tension between dedicating resources to the objectively highest value claims versus claims that are more salient/marketable and therefore increase the visibility of the site and grow the total resources available is always going to exist. For the moment, it's an easy question: since the site has no real users or resources, focusing on overcoming the cold start problem is the priority. The danger is that in chasing salience and profitability one can lose sight of the larger goal, which is to do the most good.

Resource allocation will remain centrally important even once the claim graph achieves density in claimspace.

There are rather large gains to be made by deciding to dedicate cognition to mapping and assessing regions of claimspace that are simply neglected by the various algorithms (in the broad sense, including social dynamics, education systems, and literal social media algorithms) that shape the allocation of human attention and epistemic resources today. We can do a lot better.

Estimating costs and values

I haven't implemented this yet, but here's how I'm thinking about approaching this problem. One wants to distribute resources to the things that matter most, but central planning of resource allocation is notoriously difficult. The optimal effort, depth, levels of intelligence, tools, reupdate timing, triggers, etc. with which each claim, contribution, ingestion, etc. should be handled is not something that can be decided without knowledge local to that region of the claim graph, yet resources must be allocated across the whole graph. A distributed market-like system with the principle of subsidiarity is necessary.

The solution I'm planning is to estimate the value and cost of all actions and take all actions where the marginal expected returns on investment (marginal value/marginal cost) are above a certain threshold.

For every action that the system takes or could take, there is an estimated cost. These can be taken from statistical averages or estimated by models using heuristics that we can improve on over time. Assessing a claim using Fable at high effort costs about $X, while assessing using Sonnet at medium effort costs $Y. This claim has a deep literature and lots of subarguments, so it will probably cost n times the average claim. Ingesting an article from site A is likely going to cost $X directly in extractor and matcher costs, and then assessments will cost $Y, so all told ingesting this thing will cost $X + $Y. And so on.

There is also an estimated value assigned for every possible action. This too can be informed by data analysis and heuristics. Claim P is unassessed, and fairly important, but also rather simple and just a matter of persistently looking through sources to find the answer; no galaxy brain moves required. Assessing it with Sonnet at high effort is worth $10 (the amount stuff is worth also scales with the size/importance of the expected audience), whereas assessing it with Fable Medium would only be worth $5 and cost twice as much, whereas Fable High effort would cost 5 times as much and be worth only $11, so Sonnet High it is.

Estimation of costs is also something that has diminishing marginal returns. In many cases, a crude non-LLM heuristic based on data analysis will be optimal, while for other cases, one will want an expensive model to spec out the expected costs and benefits.

Grantors and grantmakers

The main funds of Minerval should go to the highest value actions. However, there may be funders who care about a particular domain or subdomain, or who simply want to have a particular claim investigated, even if it is not a universal priority.

To do this, they should be able to subsidize their chosen subset of claimspace using a “grantmaker” agent. A grantmaker agent would work for Minerval, and still be bound by the constitution and our guidelines. However, it would also have a mandate given by the funder and a budget that it can allocate with the help of any number of subagents (to traverse the graph, identify gaps, recommend grants) it decides to employ (as would be necessary for broad mandates) to any number of potential actions. The grantmaker agent's allocations would add onto the neutral universal value of a claim and put them over the edge. In this way, funders can prioritize the ingestion, mapping, and assessment of the things they care about simply by donating money with a particular mandate.

The claims that get ingested and assessed due to the grantmakers' donations would be tracked and attributed, and the grantmaker agent would be able to adjust spending over time as feedback came in.

Similar thoughts to be found at jacksonhurley.com/marketplace-of-ideas