Identifying a simple topic in text isn't the same as detecting what people think about it.
There are several language models out there that can get at the rough top-level 'what'. But it's when we want to get at the 'why' that things get really interesting.
As a simple example, we gave 20 short statements to Qwen models of different sizes. 10 statements were about enjoying sport, 5 were general facts about sport and 5 were about cooking.
As you can see below, all models could correctly answer how many statements were in this short numbered list. But out-of-the-box models needed to be larger to correctly identify which statements related to sport, and even larger to detect which were about enjoyment of sport.
The difference between detecting a topic and understanding an outlook may seem small, but in practice, it's the difference between knowing roughly what people are mentioning and properly understanding why they're mentioning it.
When it comes to AI, it's not just about how a given model performs on a popular benchmark – it's also about whether a system can capture the layer of meaning that you care about, and do so consistently at scale.