Face’s top scientist Thomas Wolf says it’s unlikely that current AI systems will be imposed on what major labs want to discover.
I’ll talk luck At Viva Technology in Paris, the co-founder of Hugging Face said that large-scale language models (LLMs) show an impressive ability to find answers to questions, but are lacking when they try to ask the right one.
“In science, asking questions is a difficult part and we don’t find the answer,” Wolf said. “When a question is asked, the answer is very obvious, but the hard part is that they are really asking questions and models are very bad at asking great questions.”
Wolf said he came to a conclusion after reading a blog post widely distributed by humanity CEO Dario Amodei The machine of beloved blessing. In it, Amodei argues that the 21st century is about to be “compressed” in years as AI accelerates science significantly.
Wolf said that initially found the film to be exciting, but after a second read, he began to doubt Amodei’s idealistic vision of the future.
“I said that AI is trying to solve cancer and solve mental health issues. It can even bring peace to the world, but after reading it again I realized there was something that sounded very wrong about it.
For wolves, the problem is not that AI lacks knowledge, but that it is not that it has the ability to challenge existing knowledge frameworks. AI models are trained to predict the likelihood of continuity, for example, the next word in a sentence, and while today’s models excel at mimicking human reasoning, they fall short of real, original thinking.
“The model is just trying to predict what’s most likely,” Wolf explained. “But in almost every big case of discovery and art, it’s not the most likely artwork you want to see, but it’s the most interesting one.”
Using the GO example, Wolf, a board game that became a milestone in AI history when Deepmind’s Alphago beat the world champion in 2016, claimed to be impressive while mastering the rules of GO. Science said that the equivalent of inventing a game is asking these truly original questions.
Wolf first proposed this idea in a blog post titled Einstein AI ModelIt was released earlier this year. In it, he wrote: “To create Einstein in a data center, you need a system that knows all the answers, but also a system that allows you to ask questions that no one else thinks or dares to ask.”
He argues that what we have is a model that behaves like “yes on the server” instead.
