According to Jensen Huang, AI hallucinations are going to take several years to eliminate. A huge amount of computing power is needed for that, he thinks.
Nvidia CEO Jensen Huang thinks the problem of AI hallucinations – situations in which AI generates incorrect information – will remain a challenge for years to come. According to Huang, solving them is a complex process that requires a significant increase in computing power. He said this during an interview at the University of Hong Kong that Tom’s Hardware attended.
Three stages of development
Huang emphasized that artificial intelligence is currently going through three main stages of development: pre-training, post-training and Test Time Scaling. At the latter stage, AI attempts to solve problems step by step. Although these techniques lead to increasingly better answers, confidence in AI answers remains limited. Users often have to judge for themselves whether AI’s output is meaningful and reliable.
He predicts that the industry still needs years to reach a level where AI answers are fully reliable. Until then, the focus will be on further improving computing power and learning techniques.
Despite these limitations, Huang sees Nvidia’s role as crucial: “We have reduced the marginal cost of computing power by a factor of one million. As a result, machine learning is more accessible than ever.” This is a remarkable achievement that Huang says has enabled the explosive growth of AI technology.
read also