Generative AI applications have generated many questions about how enablers and users of the technology make these apps comply with privacy regulations, ethics guidelines, and other rules for corporate usage.
That all of this was not apparent earlier reflects just how quickly Generative AI emerged into wide usage. Companies like Nvidia for years have been behind the cause to ensure ethical AI behavior and policies, but the overnight success of ChatGPT did not leave much time for technology companies to demonstrate safety-first and privacy-above-all approaches to Generative AI usage.
Now, Nvidia is rectifying that with new software for its NeMo large language model (LLM) framework, which helps companies create and train their own Generative AI models and applications. The appropriately-named NeMo Guardrails aims to make sure that emerging apps powered by LLMs “are accurate, appropriate, on topic and secure,” an Nvidia blog post stated. “The software includes all the code, examples and documentation businesses need to add safety to AI apps that generate text.”
Jonathan Cohen, vice president of applied research at Nvidia, said during a media briefing about NeMo Guardrails, “We think that every enterprise really in every industry will be able to take advantage of Generative AI to support their businesses. But in order to use these models in production, it's important that they're deployed in a way that is safe and secure. A guardrail is a guide that helps keep the conversation between a human and an AI on track.”
He added that Nvidia is making its offering open source so that the broader community of Generative AI technology enablers and developers can quickly start cooperating on creating and maintaining the best possible systems and capabilities for Generative AI usage as the technology continues to mature.
Guardrails works with all LLMs, including OpenAI’s wildly popular ChatGPT, and also can run on top of other open source toolkits like LangChain, which helps developers plug third-party applications into the power of LLMs.
During the briefing, Cohen discussed and demonstrated three kinds of boundaries that NeMo Guardrails will allow developers to create:
-
Topical guardrails, which prevent apps from veering off into undesired areas. For example, they keep customer service assistants from answering questions about the weather.
-
Safety guardrails, which ensure apps respond with accurate, appropriate information. They can filter out unwanted language and enforce that references are made only to credible sources.
-
Security guardrails, which restrict apps to making connections only to external third-party applications known to be safe.
Regarding security guardrails, Cohen said that this will be a growing area of concern as LLMs become more integrated with other services across the enterprise. “The concept of security is becoming more and more important, as large language models are allowed to connect to third party API's and applications,” he said. “This can become a very attractive attack surface for cyber security threats. Whenever you allow a language model to actually execute some action in the world, you want to monitor what requests are being sent to that language model.”
He added that Nvidia’s knows that while NeMo Guardrails probably is not the ultimate solution for privacy, accuracy, safety, and security for Generative AI apps, it is at least a step in the right direction to create a framework for future solutions. “I expect this is a very active research area, and people are going to come up with all kinds of improved techniques,” he said. “So we don't think that we've solved any one of these problems, but we can take the best practices, the state of the art today, put them in one place, and as that state of the art improves, we can continue integrating it.”