How Salesforce eVerse uses digital twins to smooth AI jaggedness

One concern about using Large Language Models (LLMs) for enterprise tasks is the disconnect between simple benchmarks and real-world performance. This disparity in performance on edge cases, often requiring additional human attention, is called jagged intelligence. To address this gap, Salesforce has been working towards a vision of Enterprise General Intelligence (EGI) that combines LLMs, better digital twins & simulations, and reinforcement learning.

The firm recently announced the commercial implementation of this vision through its new eVerse branding. The first fruits of this support digital twins of end-user voice calls and business processes to improve performance on enterprise tasks. For example, early work improved performance on enterprise tasks from 19% to 88% success rates. This continuous loop enables the ability to synthesize more accurate and useful training environments, measure performance, and address gaps, promising to transform performance on real-world tasks.

Silvio Savarese, Chief Scientist at Salesforce, observes that pure LLM-based ...