[jerpint]

Transformers were used on text to train models without needing labeled data. People realized that simply scaling the data and models meant better performance. when they scaled even further, emergent intelligence started appearing and the models were dominating every known task. Now everyone wants an LLM

[amelius]

> and the models were dominating every known task

I don't think this is very accurate. How well does LLM perform on image segmentation, for example?

[jerpint]

I meant more transformers than LLMs, models trained in self supervised way like DINO actually are pretty good at image segmentation