Create a better incentive system to discourage model copies

[donaldknoller]

Although current systems in place (3% time penalty, model safetensors hash checking) help guard against blatant model copying, there are still potential avenues of exploitation.
The requirements for a solution that can accurately detect model copying are the following:

1. The solution can distinguish between "copying" and fine tunes
2. The solution is a relatively "fast" computation (ie. does not cost more execution time than the current scope of evaluation)

Further notes on tuning other subnet properties:

1. It is desirable to keep the current temperature at roughly the same range. Although the current number of participants in the subnet may be small, this value can change in the future. Furthermore, there are existing discussions with validators regarding the emissions curve and the current level

2. It is desirable to keep a relatively low registration cost. A low registration cost can encourage newer participants, and there is already an implied cost for miners who are training and evaluating their models offline.

Additional discussion points are welcome and it is expected that this may not result in a concrete solution, but at the very least provide discourse to find a way forward.

[CharewiczH]

In my opinion, the biggest problem is that every model gets emission, which encourages cloning... And the best miners are not properly rewarded. Therefore, I propose that only about 10% of the miners should earn, rather than everyone.

[donaldknoller]

Thanks for your input.
Curious as to why it would be 10% specifically?
Given that the subnet is limited to 256 entrants, 10% would be roughly 25 entries. Implicitly, this means that 231 participants will never receive any emission at all. It also means that the worst performing entrant may immediately lose emissions with a slight change in scoring. Keep in mind that there may be future changes in the landscape where many more miners will want to participate, and each entry in itself may be a valuable submission in and of itself. Would this 10% be optimal in that scenario as well?

[CharewiczH]

I'm not convinced that 10% is the optimal solution, but I haven't come across any other subnet where every miner gets emission. For example, in SN37, only the top 1 and 2 receive emissions, which isn't necessarily ideal, but... On the other hand, if every miner earns something, it could encourage cloning and discourage the top miners, for whom the emissions are relatively small.

[donaldknoller]

I think in the case of managing emissions, this may be better addressed by adjusting temperature, which other miners have requested as well.
I don't think that there is a worthwhile attempt for a "copying" miner to pay a registration fee to receive 0.001 tao in emissions vs no emissions, as long as the "top" miners are happy.
We will consider adjusting the temperature in a future update, and go from there in this case. Will keep this open for continued discussion