Higher Model Benchmarks Don't Necessarily Mean Better Results on Your Task

On the slides at a model launch event, the leaderboard has been refreshed again. In the second episode of Technically Speaking, Databricks Chief AI Scientist Jonathan Frankle offered a sobering caveat: just because a model looks better on benchmarks doesn't mean it has actually gotten better for your specific use case.

This episode of the podcast is titled What actually is AI research?, hosted by Megan Reynolds. The full version is on YouTube, and the show notes and transcript are at technically.dev. Frankle is Databricks' Chief AI Scientist and a co-founder of MosaicML, which was acquired by Databricks.

The episode doesn't start by talking directly about models; instead, it talks about what AI research actually does. Frankle's analogy is cosmology and biology: no one hands researchers the fundamental laws; everyone faces a system so complex that it cannot be derived directly from first principles, so they can only observe, break things down, and look for simplifications that can predict. Generative AI is similar. Parameters, data, compute, and network architecture are like a set of initial conditions; mixed together and run for a long time, they eventually produce a model. How those intermediate abstractions work is still not fully understood.

So the daily life of an AI researcher is not romantic. Frankle says the basic questions he asks every day are very plain: What experiment did we run? How were the initial conditions set? Did the metrics move? By how much? What should we change next? Do these metrics actually mean anything for what customers care about? A large amount of the remaining time is spent fixing GPUs, making jobs run a bit faster, checking why the data is wrong, and tracking down one small bug after another.

His way of leading a team is also direct. The episode mentions a catchphrase: If I don't have numbers, I don't believe you. No numbers, no belief. When hiring, he looks for people who are never quite satisfied—people who are always a little dissatisfied with the status quo. On release cadence, his advice is to ship the best of what you have at any given time, and don't wait for everything to be perfect.

The episode also discusses fine-tuning specifically. In most cases, Frankle does not recommend fine-tuning; the specific reasons are in the original video. On evaluation, the episode talks about a trickier problem: many tasks have no standard answer, so how do you score AI? And in agent scenarios, how do you measure whether it is doing a good job?

One reply put it very accurately: your use case was never on the leaderboard. The leaderboard is other people's questions; your task isn't on it. Another reply pointed out that this episode has a visual segment explaining RL with hot dogs that makes a complex concept funny.

Before upgrading a model, run a set of numbers on your own task first. The leaderboard can only prove that it performs well on other people's questions. Your use case is up to you to test.