Thomas Wolf, a top scientist who embraces Face, said the current AI system is unlikely to make some leading labs hope for scientific discoveries.
Speech wealth In Viva Technology in Paris, Embrace Face co-founder says that while large language models (LLMSs) exhibit impressive abilities to problematic, their problems are short when trying to come up with the right language, Wolf sees it as a more complex part of the real scientific advancement.
“In science, asking questions is the difficult part, and no answer is found,” Wolf said. “Once the question is asked, the answer is usually obvious, but the hard part is really asking questions, and the model is very bad at raising good questions.”
Wolf said he came to the conclusion after reading a widely circulated blog post by human CEO Dario Amodei. The machine of grace. Among them, Amodei believes that as AI accelerates science sharply, the world is about to see the 21st century “compression” into a few years.
Wolf said he initially found the work inspiring, but began to doubt Amodi’s idealistic vision for the future after a second reading.
“It means it will solve cancer, it will solve mental health problems – it will even bring peace into the world, but then I read it again and realized it sounded like something wasn’t right, and I didn’t believe it.”
For Wolf, the problem is not AI’s lack of knowledge, but the lack of ability to challenge our existing knowledge framework. AI models are trained to predict possible continuities, for example, the next word in a sentence, while today’s models do well in mimicking human reasoning, but they don’t have any real primitive thinking.
“The model is just trying to predict the most likely thing,” Wolf explained. “But in almost all the big cases of discovery or art, it’s not the art you’re most likely to see, but it’s the most interesting one.”
Taking GO’s games as an example, board games became a milestone in AI history, when DeepMind’s Alphago defeated the world champion in 2016, but Wolf believes that while it’s impressive to master the rules of GO, the biggest challenge is to invent such a complex game in the first place. In terms of science, he said, the equivalent of invention of games is equivalent to these real primitive problems.
Wolf first wrote in an article titled ” Einstein AI modelpublished earlier this year. “To create an Einstein in a data center, we need not only a system that knows all the answers, but also a system that can ask questions that others think of or dare not ask.”
He believes that what we have is a model that acts like “Yes-Men on the server” – without a doubt, but unlikely to challenge assumptions or rethink the underlying ideas.
