Dr. Yonatan Belinkov: "Why do we ask for a penguin and get a cat? We are trying to do brain surgery for AI."
"If we don't understand how artificial intelligence systems work, we won't be able to control them later. This also allows us to correct the biases that exist in language models," said Dr. Yonatan Belinkov, a senior lecturer at the faculty, at TheMarker and Technion's Technovation 2025 Conference.
Belinkov heads a research group that is trying to study and improve the way image generators work. "We are trying to understand what is happening inside tools in order to improve them. How do they work? The user writes a textual description, which is entered into a language model (LLM), and then there is a fairly complex and deep system built from layers of computation. After the textual description goes through several layers of computation until it reaches the final part and goes to the image generator," says Belinkov. "Today, through very targeted updates to the model, we are able to make the content relevant, introduce diverse representation, and delete specific styles," added Belinkov.
TheMarker's full article at the link
[Back to the news index]