lmarena/lmsys is beyond useless, looking at prior rankings of models vs formal benchmarks or testing for accuracy + correctness on batches of real world data. It's a bit like using a poll of Fox News to discern the opinions of every American; the audience voting is consistently found wanting. Not even getting into how easily a bad actor with means + motivation (in this "hypothetical" instance wanting to show that a certain model is capable of running the entire US government) can manipulate votes which has been brought up in the past (yes I'm aware of the lmsys publication on how they defend against attacks using cloudflare + recaptcha, there are ways around that.)
So you're saying that either A: users interacting with models can't objectively rate what responses seem better to humans, B: xAi as a newcomer has somehow managed to game the leaderboard better than all those other companies, or C: all those other companies are not doing it. By those standards every test ever devised for anything is beyond useless. But simply not having the model creator running the evaluation is already going a long way.
No I'm saying that some companies are doing it (OpenAI at the very least), the company in question has motive and capability to game the system (kudos to them for pushing the boundaries there), AND the userbases' rankings have been historically, statistically misaligned with data from evals (though flawed) and especially when it comes to testing for accuracy + precision on real world data (outside of their known or presumed dataset). Take a look at how well Qwen or Deepseek actually performed vs the counterparts that were out at the same time vs their corresponding rankings.
In the nicest way possible I'm saying this form of preference testing is ultimately useless, primarily due to a base of dilettantes with more free time than knowledge parading around as subject matter experts and secondarily due to presumed malfeasance. The latter is more apparent to more of the masses (that don't blindly believe any leaderboard they see) now that access to the model itself is more widespread and people are seeing the performance doesn't match the "revolution" promised [0]. If you're still confused why selecting a model based on a glorified Hot or Not application is flawed, perhaps ask yourself why other evals exist in the first place (hint: some tests are harder than others.)
At work, developed our own suite of benchmarks. Every company with a serious investment in AI-powered platforms needs to do the same. Comparing our results to the Arena turns up some pleasant surprises, like DBRX hitting way above its weight for some reason.
You say no, but then go on and explain why you believe a combination of both option A and option B. That's fine I guess, I just don't consider it particularly likely given the currently available information.
