Be a part of our every day and weekly newsletters for the most recent updates and unique content material on industry-leading AI protection. Study Extra
Open-source AI platform supplier H2O.ai believes a mix of generative and predictive AI fashions makes for extra constant responses which enterprises need from an AI agent.
H2O.ai launched its new multi-agent platform that blends generative and predictive AI and is now usually obtainable.
The platform, h2oGPTe, makes use of the corporate’s AI fashions Mississippi and Danube, however also can entry different massive and small language fashions obtainable. The corporate stated h2oGPTe works in air-gapped, on-premise and cloud techniques.
Sri Ambati, founder and CEO of H2O, informed VentureBeat that having each generative and predictive AI provides enterprises extra confidence that the brokers will work precisely as they want with out compromising safety.
“The primary downside with brokers is consistency. Can I get a constant response from an [large language model] LLM for a similar immediate? I believe you get two completely different, like a number of responses proper now,” Ambati stated. “However you may carry a number of fashions that negotiate, plan and ship an consequence. Consider it as people can have a little bit of variability with one another, however you continue to count on a constant response, and that’s the area of predictive AI mixed with generative AI.”
Ambati defined that generative AI fashions are “respectable at content material era and excellent at code era,” however predictive fashions carry extra situation simulation to the desk. He stated the predictive fashions carry consistency to agentic responses as a result of these don’t simply generate responses however be taught from patterns in information.
The platform is constructed for finance, telecommunications, healthcare and authorities enterprises that have to handle multi-step duties. H2O.ai’s agent works finest for organizations that need to get insights into their enterprise and never only a information that runs by means of their workflows. It is because brokers throughout the h2oGPTe platform can learn multimodal information like charts and craft solutions to questions like “Ought to my firm promote extra dolls this yr?” that think about the enterprise’s historic monetary information or market development info they retailer.
Multimodal brokers
Like different AI brokers, h2oGPTe automates workflow duties so human workers don’t must do these actions themselves. Ambati stated the multimodal capabilities of H2O.ai’s brokers open up extra info that it may possibly be taught from to supply the perfect, most constant solutions to customers.
The corporate stated the brokers also can create PDF paperwork with charts and tables grounded in enterprise information to visualise info for the human consumer. H2O.ai ensured that the brokers cite their sources for information traceability and provide customizable guardrails.
H2O.ai’s agentic platform builds in mannequin testing, together with automated query era, the place an AI mannequin will create variations of a immediate and barrage the agent with inquiries to see if it persistently responds. It additionally has a dashboard the place folks can determine which sort of database, mannequin, or a part of the workflow the brokers tapped.
Consistency and accuracy in brokers
With the hype round AI brokers predicted to proceed to the next yr, there’s a want to make sure brokers present worth to enterprises, together with performing persistently, reliably and precisely.
Reliability is important as a result of AI brokers are supposed to automate a big portion of an enterprise’s workflow with out human intervention.
H2O.ai’s strategy of mixing generative and predictive fashions is a method, however different firms are additionally taking a look at methods to make sure AI brokers don’t trigger hassle for enterprises. The startup xpander.ai launched its Agent Graph System for multi-step brokers. Salesforce additionally launched to a restricted preview its Agentforce Testing Middle to check agent response consistency.
