I'm running a long day trading contest for frontier models

Someone got tired of benchmarks that feel like open-book exams. The contest, surfaced on r/ArtificialIntelligence on May 6th, sets frontier models against each other in extended day trading sessions — not a sprint, not a curated snapshot of model behavior, but a sustained run designed to outlast luck. The framing is deliberate: financial markets are one of the few environments that will punish a reasoning system for being confidently wrong in ways that a standard leaderboard simply won't. The original post scored modestly (5.4), which tells you the idea hasn't caught fire yet, but it doesn't need to be popular to be interesting. The methodology — if the contest is being run rigorously — would expose something that most AI evaluations deliberately avoid: how these models behave when the feedback loop is immediate, the stakes are real in aggregate, and there's no rubric telling them what a good answer looks like. That's not a benchmark. That's closer to a stress test.

If confirmed, here is what this means. A well-structured trading contest among frontier models would produce something genuinely rare — comparative performance data on reasoning under uncertainty, generated outside the labs that built the models and outside the benchmarking organizations that those labs fund or influence. That independence matters more than the trading domain itself. The financial angle is almost incidental: what's really being measured is which models can update their priors correctly, manage compounding error, and avoid the kind of confident hallucination that looks fine in a chat window but ruins a position in twenty minutes. Second-order, this kind of community-run adversarial evaluation could become a template — other domains with clear scoring (chess engines, logistics simulations, legal outcome prediction) would be natural extensions. The deeper pressure it applies is to the labs themselves, who have every incentive to perform well on tests they designed and less clarity on how their models perform on tests they didn't.

Watch for published methodology and any preliminary results from the contest runner — and watch for whether any of the major AI research communities pick this up and attempt to replicate or critique the design.

NewsHive monitors these sources continuously. All signal titles above link to the original reporting.

Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →