Artificial intelligence Trustworthy AI that does not lie is still many years away

Related Vendors

NVIDIA GmbH

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

Wuxi InfiMotion Technology Co., Ltd.

When artificial intelligence lacks information, it sometimes fills these "knowledge gaps" with lies without being prompted—the artificial intelligence then hallucinates. Nvidia's CEO Jensen Huang believes that the issue of hallucinating AI will persist for many years.

Artificial intelligence is currently a helpful tool in everyday life, but only if you know how to handle it. You cannot blindly trust AI as long as the problem of hallucinations is not resolved. When artificial intelligence hallucinates, it basically lies to the user. This happens when the AI lacks information. It then fills in the gaps with misinformation or completely fabricated knowledge. That's not trustworthy.

Jensen Huang, the CEO of the world's largest AI chip provider, Nvidia, shares this view. In an interview at the Hong Kong University of Science and Technology, Huang spoke about this and other topics, such as the increase in demand for computing power over the past ten years. On that occasion, Huang, along with others, was awarded an honorary doctorate from the university, specifically the Doctor of Engineering honoris causa.

In discussing the developments achieved in AI training, Huang emphasized that the progress to date is impressive, but AI cannot yet be widely trusted.

Fight against hallucinations—with computing power

Each year, the demand for AI computing power increases about fourfold, according to Huang's host Professor Harry Shum at the fireside chat at the university. "When you gaze into your crystal ball and look to the future, are we going to see a million-fold increase in demand over the next ten years?" was the subsequent question.

Huang did not provide a direct answer but explained how artificial intelligence is currently being developed. In pre-training, AI essentially learns all the data in the world and thereby discovers knowledge; Huang compares this to a college phase. However, this is not sufficient. The next phase is post-training, in which different learning techniques are applied. These include reinforcement learning through human feedback, reinforcement learning through AI feedback, synthetic data generation, and multipath learning to specialize the AI.

The final and most complex step is test-time scaling, also referred to as "thinking." In this process, the AI breaks down the problem step by step to find a solution and may need to iterate and simulate different outcomes if the answer is not predictable. Huang: "We call this thinking, and the longer you think, the higher quality the answer could become."

Despite all these efforts, according to Huang, we still cannot fully trust the best answers that AI can provide. "Today, the answers we have are the best we can get. But we need to reach a point where the answer is not only the best, but where we no longer have to decide whether it is hallucinated, whether the answer makes sense, whether it is reasonable or not," Huang explained. "We need to get to a point where we can largely trust the answer we receive. I believe we are still a few years away from this point, and in the meantime, we need to continue increasing our computing power."

Without Nvidia, AI would be expensive

In this field, Nvidia has particularly distinguished itself in recent years, Huang is convinced. "What Nvidia has contributed is that we have reduced the marginal cost of computing by a millionfold." This has fundamentally changed the habits regarding computing with enormous amounts of data in business and science, which is why machine learning has taken off. Huang countered the claim that Nvidia's AI GPUs are still relatively expensive by saying they would be much more expensive without Nvidia. "I have given you a millionfold discount over the past ten years. It's virtually free," Huang said. You can watch the interview in the video below.

(sb)