Nvidia announced on Tuesday "NeMo Guardrails", a new AI-powered software that could stop ChatGPT from "hallucinating".
The tech manufacturer giant Nvidia announced on Tuesday that they might have come up with a way of preventing AI softwares from “hallucinating”. Nvidia is a crucial company in the technological sector due to their supply of microchips. According to CNBC, Nvidia accounts for 95% of the AI chips market share.
On Tuesday, Nvidia announced a software that could solve one of the AI’s biggest issues. Text generators powered by artificial intelligence like ChatGPT have groundbreaking potential of disrupting the technological sector, but also have one fatal flaw.
ChatGPT often “hallucinates”, meaning it often confidently makes up answers just to avoid admitting ignorance. Any user that tries to extract an objective fact from ChatGPT knows it.
A similar mistake caused a $100 billion loss in Google when they announced their early stage AI software.
Bard was supposed to be Google’s answer to ChatGPT but upon announcement it answered with such a blatantly wrong fact that investors immediately lost interest. Bard was asked what the James Webb telescope is famous for, and the software replied that it had taken the first picture of a planet outside the solar system. The problem was the James Webb telescope was launched 20 years after such a picture had been taken.
NeMo Guardrails
Nvidia’s solution to AI’s hallucinations is called “NeMo Guardrails”. It’s essentially another layer of artificial intelligence capable of “guiding” the target software.
NeMo Guardrails is a Large Language Model (LLM), meaning it uses large amounts of data to “learn” its purpose. ChatGPT and Bard are LLMs as well, but NeMo would work differently.
While ChatGPT uses its data to predict what outcome it will “spit out”, NeMo would use it to double check the software’s answer. In other words, ChatGPT will use the data to develop an answer, while NeMo will run it through the same data to check its plausibility.
If NeMo does not find matching data, the answer will simply be “I don’t know”.
Nvidia claims that NeMo can also be used to prevent the software from giving certain answers. If, for example, an AI chatbot is used as customer service, NeMo can prevent it from mentioning the competition.
“You can write a script that says, if someone talks about this topic, no matter what, respond this way,” said Nvidia vice-president Jonathan Cohen.
Nvidia has an established position in Silicon Valley as one of its biggest manufacturers. Now, it wants to retain this role in the AI race. During the California gold rush of 1848, those who truly profited were the shovel sellers. Today, the shovels became microchips and the gold became artificial intelligence.
