Artificial intelligence is maturing into a new phase — one powered by customizable frontier models built for real-world domain expertise.

Look no further for evidence of this shift than with Nova, Amazon Web Services Inc.’s in-house family of frontier AI models, which saw multiple updates rolled out on Tuesday. But to best understand the motivation behind these updates starts with acknowledging a persistent industry challenge, according to Rohit Prasad (pictured), senior vice president and head scientist of artificial general intelligence at AWS.

“A frontier model comes out. Public benchmarks look great,” Prasad told theCUBE. “And then the production reality sets in that when you try to build your applications, your workflows — it doesn’t meet your expectations quite a few times. There’s a fundamental reason for that, [and it’s] because the knowledge you have of your domain, your use cases, is not in the frontier model.”

Prasad spoke with John Furrier at AWS re:Invent, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed Nova Forge’s approach to customizable frontier models and the growing role of agents in next-generation AI systems.

Frontier models come into focus

Frontier models often perform well on benchmarks, but fall short in real production because of a lack of domain-specific knowledge. Until now, there were only imperfect solutions, including starting with a closed model and fine-tuning it using approaches such as reinforcement learning, according to Prasad. Alternatively, companies could start with an open-weight model and make some progress, but once they add their own domain data without access to the original training data, the model becomes unstable, he added.

“The third [option] is you go with a very expensive route … which [is] you build your own, but that requires deep expertise, massive computing [and a] massive amount of data collection. All of that is prohibitive for most enterprises. This is why we challenged the team and said, ‘Go invent this.'”

That invention turned into Nova Forge, introduced as a way to let users build deeply customized variants of Nova models, according to Prasad. It was made possible by providing multiple checkpoints that users can adapt with their own data: pre-trained, mid-training and post-training.

“[At] each of these model training stages, you can add your frontier data — your proprietary data — with one unique thing, which is [that] we are giving you access to Amazon-curated data to blend in,” he said. “Then when you do that, the model maintains its general intelligence and becomes an expert in your domain.”

As the AWS team developed the system, customers and partners could collaborate and test it in real use cases, with promise showing itself early. For example, Reddit Inc. used it to build a sophisticated model that deeply understand the company’s content moderation needs, according to Prasad.

“They found that they used to use six bespoke models … and then they got access to Forge and were able to build [one],” he said, pointing to Reddit dramatically simplifying its moderation stack.

Here’s the complete video interview, part of SiliconANGLE’s and theCUBE’s coverage of AWS re:Invent:

Photo: SiliconANGLE