Over the weekend, while playing LiverLive with friends, I discovered an AI blind spot. They can’t differentiate between 3/4 and 4/4 time signatures. I’m guessing there wasn’t relevant training data during pre-training, or at least there’s a lack of Chinese song data. So they just start making stuff up, and it’s basically all wrong (both GPT and Grok are like this, and they always give 4/4). **GPT Chinese Songs** Later, I got curious and tested it out. Turns out it’s indeed due to the lack of Chinese training data. For English songs, the recommendations are all correct, which shows they can actually tell the difference. **GPT English Songs** Then I went on to test Deepseek, a domestic large language model, to see if it included this kind of data for Chinese songs during pre-training. The results were indeed better than GPT and Grok, but only by half. Half of the results were 3/4 (or 6/8), while the other half were still 4/4… But hey, it’s still better than those two foreign models that got everything wrong. **Deepseek Chinese Songs** (didn’t bother testing English + limited to 4 images) Lastly, a quick recommendation for the LiberLive entertainment electric guitar. It’s fun and affordable, perfect for people who haven’t really learned music but want to experience the thrill of playing and singing with a guitar. You can get started in 10 minutes—it’s like how it used to take a year to launch on the mainnet, but now it’s just a 5-minute one-click process. :)