Resolves to the first year where a model is released that:
1. Can be run locally on a reasonable consumer phone
2. Has comparable scores on most benchmarks to o3-mini (medium).
I will judge 2 based on benchmarks like Codeforces scores, AIME math, GPQA-Diamond, etc, but will use discretion if it seems obvious some benchmarks have been compromised and that the AI is clearly inferior to o3-mini. It will not need to surpass or match o3-mini's performance on all of these, but it will need to do so on the majority. Again, some discretion will be used here.
I won't bet in this market.
Update 2025-06-26 (PST) (AI summary of creator comment): To qualify for resolution, the model must meet a minimum performance standard:
It must run at at least 2 tokens per second (TPS).
This performance must be on a currently available consumer smartphone.
1,000People are also trading
@LeonLang Hmm, I don't want to gerrymander the rules. What do you think is a reasonable lower cutoff that would mean it's pragmatic to run on one's phone for purposes people would actually use? Is 5 tps too high for that?
@Bayesian That sounds reasonable to me! Let's say it has to run at at least 2 TPS on some currently available consumer smartphone.