Despite the hype, GPT-4 and other text-generating AI models from OpenAI have significant shortcomings. James Vincent previously described such a model as an “emotionally manipulative liar,” which fairly accurately describes the present climate.
Companies behind these models have said they are working to address these concerns by introducing filters and human moderation teams to address issues when they are reported. However, no one answer will suffice. Biases, toxicity, and intentional assaults may affect even the most advanced models.
Nvidia Brings A Solution
Nvidia has published NeMo Guardrails, an open-source toolkit for creating “safer” text-generating models that will make AI-powered applications more “accurate, appropriate, on topic, and secure.”
Nvidia’s vice president of applied research, Jonathan Cohen, says the business has spent “many years” developing the technology behind Guardrails, but only in the last year did they discover it was a suitable match for models like GPT-4 and ChatGPT.
“We’ve been developing toward this release of NeMo Guardrails ever since,” Cohen told TechCrunch in an email. To successfully deploy models for corporate use cases, “AI model safety tools are essential.”
Text and voice may both be generated by AI applications, and Guardrails provides code, examples, and documentation to “add safety” to these apps. According to Nvidia, the toolkit can be used with a wide variety of generative language models, and it just takes a few lines of code to implement new rules.
How Does Gaurdrails Work?
To be more specific, Guardrails may be used to stop (or at least try to stop) models from going off-topic, replying with false information or poisonous language, or forming links to “unsafe” other sources. Examples include preventing a search engine chatbot from directing users to questionable academic papers and preventing a customer care assistant from answering queries about the weather.
While using Guardrails, “developers possess the final say regarding what is off-limits within their application,” as stated by Cohen. It’s possible that “they may develop guardrails that are either too broad or, conversely, too narrow for their use case.”
However, a global solution to the problems that plague language models is too good to be true. Firms like Zapier are using guardrails to make their generative models safer, although Nvidia admits that the toolkit isn’t flawless and won’t spot every issue.
Using the famous LangChain architecture, Cohen explains that Guardrails is most effective with models that are “sufficiently good at instruction-following,” similar to ChatGPT. Because of this, many open-source alternatives are eliminated.
Is The Tool Paid or Free?
It’s important to note that Nvidia isn’t necessarily providing Guardrails out of the kindness of its heart, regardless of how great the technology is. Included in the Nvidia Enterprise AI Software Suite and the NeMo Fully Managed Cloud Service, it is part of the company’s NeMo architecture. The Guardrails open source distribution is available for any business, although Nvidia prefers to have the business pay for the hosted version.
While Guardrails is unlikely to do any damage, it’s important to remember that it’s not a panacea and to be careful of Nvidia’s assurances.
Continue reading iTMunch for the latest social media news, articles, and stories.
Feature Image Source: Nvidia