NVIDIA CEO says 'trustworthy AI' is still several years away
NVIDIA's CEO Jensen Huang has expressed skepticism over the current state of artificial intelligence (AI) in an interview at the Hong Kong University of Science & Technology. He said today's AI doesn't always give reliable answers and hinted it could be several more years before we have an AI system that can be largely trusted. "In the meantime, we have to keep increasing our computation," noted Huang, the leader of the world's most valuable company.
Huang highlights the issue of 'hallucination' in AI
Huang emphasized that users shouldn't have to question if an AI's response is "hallucinated or not hallucinated" or "sensible or not sensible." He used "hallucination" to refer to a case where AI gives false or fictitious information. This has been a major problem for advanced language models such as ChatGPT, despite their incredible progress in responding to complex queries over the past few years.
OpenAI faced legal action over AI's 'hallucination'
The issue of hallucination in AI has already resulted in legal troubles. Last year, a radio host sued OpenAI after its ChatGPT model produced a fake legal complaint against him. The incident highlights the real-world ramifications of AI's shortcomings and the difficulties tech companies encounter in making them reliable.
Huang questions the effectiveness of pre-training AI models
Huang also questioned the effectiveness of pre-training AI models on large, diverse datasets before they're developed for specific tasks. He argued that this approach is insufficient for creating reliable AI systems. His comments highlight an ongoing debate in the tech industry about how to improve large language models (LLMs) without solely relying on vast amounts of data, which is a limited resource.
