Experts Identified Key Hurdles for Agentic AI

Daniela Rus and Ion Stoica outlined technical infrastructure requirements to stabilize AI scaling.

Updated on Sept. 21, 2026 in Artificial Intelligence

Isometric editorial illustration of silicon wafers and memory modules in a sterile, modern laboratory setting, highlighting AI hardware architecture. — Researchers Daniela Rus and Ion Stoica warned that current AI infrastructure faces critical bottlenecks in memory capacity and compute efficiency, threatening the scaling of future autonomous systems. AI Illustration. Upload story photo >

Live Poll

Do you trust artificial intelligence to handle your personal data securely on your local devices?

At the Imagination in Action event, researchers Daniela Rus and Ion Stoica analyzed how current AI infrastructure fails to support the growth of agentic AI systems. They identified exponential cost increases and memory capacity limitations as primary barriers to future development.

Why it matters

The discussion underscores a growing disconnect between compute power and system efficiency that threatens to stall autonomous AI capabilities. Addressing these structural gaps is essential for aligning advanced models with human intent and deployment safety.

Modern storage requires one transistor to maintain 1 bit of information, a physical constraint contributing to the widening gap between memory capacity and compute performance. This divergence forces developers to rethink traditional system modularity to sustain scaling.

The players

Daniela Rus

Director of the MIT Computer Science and Artificial Intelligence Laboratory focusing on robotics and autonomous systems.

Ion Stoica

Professor at UC Berkeley and co-founder of Databricks specializing in distributed systems and large-scale data processing.

The details

Efficiency bottlenecks arise because current technology stacks struggle to maintain modularity as AI complexity increases. Researchers propose using formal methods—mathematical techniques for verifying software and hardware designs—to bridge the gap between human intent and the underlying model logic. Integrating human feedback loops is presented as a mechanism to ensure user intent remains central as AI systems transition toward recursive self-improvement, a process where an AI system redesigns its own code.

September 14-15, 2026: Daniela Rus and Ion Stoica spoke at the Imagination in Action event in Mountain View.

The Tech Race

This dialogue updates the discourse around the AI Alignment Problem by shifting focus from theoretical safety to infrastructure reality. It highlights the urgent need to reconcile hardware limitations with the ambitious goals of agentic system development.

Improved on-device AI will eventually offer users greater data privacy by reducing the need for cloud-based processing. The transition remains in the research stage, with no current timeline for commercial deployment of these infrastructure-optimized systems.

The takeaway

The trajectory of autonomous AI depends heavily on solving physical bottlenecks in memory and power efficiency. Observers should track the adoption of formal verification methods in model development as a primary indicator of progress toward reliable agentic systems.