Goldman Sachs' Summary After Silicon Valley Investigation: Agents Enter the Execution Era, AI Competition Shifts to Workflows, World Models Rise
Based on a recent field research in Silicon Valley, Goldman Sachs highlights a key shift in the AI industry: moving from systems that "answer questions" to autonomous AI agents that "execute tasks." Commercial models are transitioning from per-seat subscriptions to usage- and outcome-based pricing. The competition is shifting from raw model capability to mastery over specific business workflows, with value accruing to proprietary data, domain context, and operational expertise.
A major hurdle for enterprise Agent deployment is not technical ability but "controllability"—issues of accountability, auditability, and error correction, especially in regulated fields. Workflows with clear rules, verifiable outcomes, and reversible actions (e.g., invoice processing) are being automated first.
The model landscape is evolving toward a division of labor. Frontier models (like GPT-4) are expected to handle high-value, high-reliability core tasks, while improving open-source models will likely capture the majority (~90%) of inference tokens for standardized, high-volume tasks due to cost advantages.
Finally, attention is moving from Large Language Models (LLMs) to "World Models," which understand physical environments, causality, and dynamic interactions. This shift elevates the importance of proprietary, real-world data (from industrial, scientific, and robotic systems) and could drive a second wave of compute demand. Goldman Sachs projects compute needs could grow ~24x over five years, benefiting cloud and infrastructure providers.
marsbitHá 52m
