Co-founder and Chief Science Officer at Hugging Face, Thomas Wolf, speaks at the opening ceremony of the Web Summit, in Lisbon, Portugal, November 11, 2024.
Current artificial intelligence models from labs like OpenAI are unlikely to lead to major scientific breakthroughs, a tech co-founder said, pouring cold water on some of the hype around the technology and claims by major figures in the field.
The comments by Thomas Wolf, co-founder of $4.5 billion AI startup Hugging Face, are in stake contrast to those by major names in AI including OpenAI boss Sam Altman and Anthropic CEO Dario Amodei.
When Wolf talks about scientific breakthroughs, he means novel ideas like those at a Nobel Prize level. Examples including Nicolaus Copernicus who theorized the sun was at the center of the universe and other planets move round it.
Wolf explained a couple of issues with chatbots right now. The first is that these products like ChatGPT and others often agree or align with the person prompting it. Think back to if you've asked a chatbot a prompt and it will tell you how interesting or great that question is.
The second is that the models underpinning these chatbots are designed to "predict the most likely next token" or "word" in a sentence.
However, he noted two key traits of scientists. The first is that scientists who make major breakthroughs are often contrarian and question what others are saying.
"The scientist is not trying to predict the most likely next word. He's trying to predict this very novel thing that's actually surprisingly unlikely, but actually is true," Wolf said.
The Hugging Face co-founder has been thinking about this topic for the last few months. His interest was sparked after he read an essay penned by Anthropic's Amodei, who posited that "AI-enabled biology and medicine will allow us to compress the progress that human biologists would have achieved over the next 50-100 years into 5-10 years."
That got Wolf thinking about the state of AI and how this won't be possible, in his view, with the current crop of models.
... continue reading