Tao Templar’s Laundry Experiment Exposes the Hardest Problem in Bittensor

Designing an incentive mechanism is as defining what participants should do, then rewarding them for doing it. In practice, getting those rules right can take months or years because every gap creates an opportunity to optimize around the system.

Tao Templar illustrates this with a household laundry experiment, teaching his 13-year-old son responsibility by having him handle the family’s laundry in exchange for computer time or money.

The first rule was to do the laundry and get rewarded. He, the 13-year-old, quickly found the easiest interpretation, gathering clothes, starting the washing machine, and stopping there. Drying and putting the clothes away were skipped.

The next version required washing or drying, which solved part of the problem. But clean clothes still ended up dumped around the house because the mechanism did not require them to reach their intended destination.

Adding “put the clothes away” improved the outcome, but created another problem which required constant checking to confirm and ensure the work had been done properly.

The mechanism was simplified again to one outcome, put clean clothes away. This removed some of the validation burden, but failed to specify quantity. The logical response was to put away the minimum number of items necessary to claim the reward.

The latest version pays five minutes of computer time for every article of clothing put away. Two days later, two giant piles of laundry appeared on the Tao Templar’s bed. Early results were promising.

The experiment captures a central challenge for Bittensor subnet builders. Participants respond to the incentives they are given, not necessarily to the outcome the designer has in mind. If a rule leaves room for an easier path, participants have a reason to find it.

That makes validation just as important as the reward itself. A mechanism that requires constant supervision may produce the desired result, but it becomes difficult to scale. Quantity and quality also need clear definitions so participants cannot maximize rewards by producing the cheapest acceptable output.

The challenge becomes even greater on a permissionless network. A family can rely on assumptions about how its only participant behaves. An open network, like Bittensor subnet, must account for anonymous miners actively searching for arbitrage and weaknesses in the rules.

This is why incentive mechanisms often require long periods of iteration. Each version can reveal behavior that was difficult to anticipate until real participants began optimizing against it.

When the mechanism finally works, the result can feel almost effortless. Participants perform useful work, validators can measure it, and rewards naturally follow the value produced.

That is the state some mature Bittensor subnets have reached after extensive refinement. For newer subnets, reaching that point can take considerably longer. The lesson from the laundry is knowing that writing the rules is only the beginning. The real test comes when participants start trying to beat them.