Nvidia CEO Jensen Huang said that today’s artificial intelligence does not provide the best answers and that the world is still “several years” away from an AI we can “broadly trust.”
CEO Jensen Huang says companies need more computing power to improve artificial intelligence.
- Nvidia CEO Jensen Huang says today’s AI does not yet provide the best answers and is still several years away from being a trustworthy AI.
- One problem with AI chatbots is hallucination, i.e. providing false or fictitious answers, as in the case of ChatGPT.
- Huang emphasizes that pre-training on large data sets alone is insufficient and AI companies need to find alternative solutions to further develop LLMs.
"The answers we have today are not the best we can give," Huang said in an interview at the Hong Kong University of Science & Technology on Saturday. The CEO said people should no longer have to question the answers of an AI and wonder whether they are “hallucinated or not hallucinated” or “rational or not rational”“We need to get to a point where you have a lot of confidence in the answer you get,” he said, “and I think we’re still a few years away from getting there, and in the meantime we need to keep increasing our computing power.”
Chatbots hallucinate
Large language models like ChatGPT have made exponential progress in answering complex questions in recent years, but they still have their limitations. A persistent problem with AI chatbots is hallucination, that is, providing false or fictitious answers.
OpenAI , widely considered to be a leader in the AI race, was sued last year by a radio host after ChatGPT created a fake lawsuit about him.
A spokesperson for OpenAI did not respond to a request for comment.
Some AI companies also face the dilemma of how to advance LLMs without relying exclusively on large amounts of data – an already finite resource.
Training AI models with data alone is not enough
During the interview on Saturday, Huang said that pre-training, or training a model on a large, diverse dataset before developing it to perform a specific task, is not enough.
“Pre-training — just taking all the data in the world and automatically discovering knowledge from it — is not enough,” he said. “Just like going to college and graduating is a very important milestone, but it’s not enough.”
