Anthropic’s revamp of its constitution document for the cloud is an effort to strengthen its position as a security-first, responsible AI model maker and a move that reflects the continued value enterprises place in model transparency and openness.
Generative AI model maker introduced a new cloud constitution on January 21, which is different from the original Constitutional AI Document It was released in 2023. The Basic Constitution The Cloud Foundational Model provides several rules for the family to follow.
The amended Constitution provides general principles, focus on logic and a 4-tier priority system that establishes a hierarchy Security, Ethics, ComplianceAnd helpfulness. The document explains why the cloud follows certain rules and hints that there may be some consciousness behind the models.
Cloud Constitution emphasizes that, while much remains unknown about how AI models work, enterprises are right to assume that each model has a bias based on its training and the principles that guide it.
how vs what
With Cloud Constitution, Anthropic aims to provide greater transparency, giving enterprises confidence that the vendor cares about keeping its models within bounds, especially given that some model providers, notably Elon Musk’s XAI, have not stopped their models from doing inappropriate things, e.g. pictures of women taking off their clothes.
“(Anthropic) is generally interested in providing AI with a set of principles,” said Bradley Shimin, an analyst at Futurum Group. “This is something that companies can have some degree of confidence in when they build their software.”
Arun Chandrasekaran, an analyst at Gartner, said the changes Anthropic made to its new constitution are designed to give the cloud a reason to act a certain way, not just tell it what to do.
“The goal is to help models make good decisions in new and unpredictable situations by applying broad principles rather than following specific rules,” he said.
The emphasis Anthropic has placed on teaching models to reason about principles means it can lead to “more reliable behavior in edge cases,” Chandrasekaran said, referring to extreme and rare instances where the output the models produce is not predictable, such as when using models in new applications they have not been trained for.
“This is important for enterprise deployments where unexpected scenarios are inevitable,” he said. The unexpected scenario may be applying the technology to a new experience that was not previously thought of.
“What we’re talking about here is something that is more akin to philosophy and ethics and less a strictly engineering-oriented approach to AI, and with alignment And trust these models,” Shimin said. He said the emphasis on trust is related to the idea that models can have consciousness or can think in the same way humans reason.
A value on transparency
It also shows that enterprises are increasingly emphasizing transparency in model training. Anthropic is not the only AI model provider trying to meet this need of enterprises.
Open source model vendors such as IBM, Nvidia, Meta, and AI2 aim to provide transparency about their models. Training Data and Recipes.
“This idea of transparency and alignment and thinking about ethics is important,” Shimin said. Even enterprises are grappling with these concepts as they design around their data, he said. However, it is important for enterprises not to view the guidance and principles that Anthropic provides as a sense of security that the model will never go astray. Despite the principles of the model, domain expertise is still needed, Shimin said.
Additionally, anthropic principles can also limit creative freedom, causing enterprises to feel stuck with a cloud approach, Chandrasekaran said.
