Just did a quick analysis to illustrate how skewed the competitiveness of batches is when you look at averages vs quantiles. Unfortunately I can’t easily share it in a reproducible or self-contained way, and the data is not available offline, so if someone from the DAO could run it on the ground truth might be nice.
I took recent batches with at least 5 solutions, and ranked the top solutions by USD value of the objective function (using a fixed $1650 eth price). Then I looked at the average utility for each rank, and the fraction of the time each of the ranks are within x$ of the top rank.
Out of ~1000 batches, the breakdown looked roughly like this:
Rank : Avg utility | % within $1 of r1 | % within $5 of r1 | % with $10 of r10
- Rank 1 : 221$ | 100% within $1 | 100% with $5 | 100% with $10
- Rank 2 : 209$ | 33% within $1 | 81% with $5 | 89% with $10
- Rank 3 : 190$ | 6% within $1 | 64% with $5 | 81% with $10
- Rank 4 : 178$ | 2% within $1 | 54% with $5 | 74% with $10
- Rank 5 : 165$ | 1.5% within $1 | 46% with $5 | 66% with $10
The average utility difference between the ranks is quite big, 12$ for best to second-best, which is close to the average reward but doesn’t factor in failed tx cost and the fact that 25% of rewards can’t be sold. Based on average gaps there should be no point in pennying. However, this gap is dominated by outliers. In most batches the top solvers are all very close and they are equally rewarded for winning in those settings.
Let’s say we somehow manage to eliminate pennying altogether. Let’s say we also plug all other ways to burn reward to gain a couple of bucks in utility. We still have an incentive problem. It’s way more profitable to try to close a tiny gap for easy batches where everyone is close, than it is to find a large improvement in rare cases, which has more impact on the user.
In long term, we will be paying out the majority of subsidies to a large number of nearly identical solvers before it becomes competitive to build a long-tail solver, and those solvers then compete for a much smaller share of the pie.