Photo credit: venturebeat.com
Stay updated with the latest insights and developments in the AI industry through our newsletters.
H2O.ai, a prominent provider of open-source AI platforms, is advocating for a hybrid approach that intertwines generative and predictive AI models, which the company asserts leads to more reliable and consistent responses—qualities that enterprises are increasingly seeking in AI agents.
The launch of H2O.ai’s new multi-agent platform, named h2oGPTe, marks a significant advancement in the company’s capabilities. This platform is not only based on H2O.ai’s own AI models, Mississippi and Danube, but it also has the flexibility to utilize a variety of large and small language models. H2O.ai has stated that the h2oGPTe platform functions effectively across various environments, including air-gapped, on-premise, and cloud systems.
Sri Ambati, the founder and CEO of H2O.ai, emphasized in an interview with VentureBeat the importance of combining generative and predictive AI to instill greater confidence in enterprises regarding their AI agents’ performance while upholding security measures. He pointed out that a common issue faced by businesses is the inconsistency of responses from large language models (LLMs). “The primary challenge with these agents is ensuring consistency. Often, you’ll receive different responses for the same prompt,” Ambati explained. He likened the operation of multiple models to human interactions, where variability exists but consistent expectations are maintained, underscoring the potential of integrating predictive AI with generative AI.
Ambati described generative AI as competent in content and code generation, while predictive AI excels in scenario simulations. The latter is crucial for providing consistency in responses as it learns from patterns in historical data.
The h2oGPTe platform is tailored for sectors such as finance, telecommunications, healthcare, and government—areas that require effective handling of complex, multi-step tasks. Organizations using H2O.ai’s agents can extract insightful analyses rather than just passively following predefined workflows. The platform’s agents have the capability to interpret multimodal data like charts, enabling them to address business questions, such as assessing the need to increase doll sales based on historical performance and market trends.
Leveraging Multimodal Agents
Similar to numerous AI agents, the h2oGPTe platform automates various workflow tasks, limiting the need for human intervention. Ambati remarked that the multimodal capabilities of H2O.ai’s agents enhance the breadth of information they can assimilate, which in turn allows for more precise and consistent responses for users.
Furthermore, H2O.ai’s agents can generate PDF documents containing relevant charts and tables derived from organizational data, thus providing a visual representation of information. The platform assures data traceability by citing sources and includes customizable safeguards to enhance user confidence.
The architecture of H2O.ai’s platform incorporates mechanisms for model testing, including automated question generation. This feature allows an AI model to formulate prompt variations and rigorously test the agent’s consistency in responding. Additionally, the platform provides a dashboard to track the types of databases, models, and workflow segments utilized by the agents.
Ensuring Consistency and Accuracy in AI Agents
As the enthusiasm surrounding AI agents continues into the next year, it is imperative to ensure that these agents deliver tangible value to enterprises through reliable, consistent, and accurate performance.
Dependability is essential since AI agents aim to automate considerable segments of an enterprise’s operations with minimal human oversight.
H2O.ai’s methodology of merging generative and predictive models is one innovative approach in a landscape where several companies are exploring alternatives to enhance the reliability of AI agents. For instance, startups like xpander.ai have introduced their Agent Graph System designed for multi-step agent tasks. In a parallel development, Salesforce has unveiled a limited preview of their Agentforce Testing Center, aimed at evaluating the consistency of agent responses.
Source
venturebeat.com