Reining in AI: The Path to Controlled Interactions

Foundation models, particularly large language models (LLMs), are driving a new wave of human-in-the-loop (HITL) cyber-physical systems. These systems, powered by AI agents, promise dynamic interactions across physical spaces and human interfaces. But as these agents become more autonomous, their unpredictability increases. How do we harness their potential without losing control?

The Challenge of Unpredictability

Human unpredictability combined with the evolving dynamics of physical environments complicates the behavior of AI agents. This agentic autonomy, while powerful, introduces a level of nondeterminism that can veer systems into chaos. Here lies the crux of the problem: How do we ensure these systems remain reliable?

The solution? A reactor-model-of-computation (MoC) approach, brought to life through the open-source Lingua Franca framework. By anchoring AI interactions within this structured environment, developers aim to restore determinism without stifling innovation.

Case Study: The Agentic Driving Coach

A practical application of this theory is demonstrated through an agentic driving coach. This scenario highlights the potential and pitfalls of HITL CPS. While the driving coach can dynamically adjust to real-time changes on the road and in user inputs, it also exemplifies the difficulties in predicting every possible variable.

Evaluations of the Lingua Franca-based framework reveal both progress and challenges. Reintroducing determinism isn't straightforward. It requires careful orchestration of agent behavior, environmental feedback, and user interaction.

Why It Matters

The AI-AI Venn diagram is getting thicker. With AI agents interacting more intricately with the real world, ensuring their reliability becomes critical. The compute layer needs a payment rail, and without structured control, system failures could escalate into real-world dangers.

So, what's the takeaway for industry stakeholders? The pursuit of agentic autonomy shouldn't sacrifice predictability. As AI continues to blur the lines between human and machine, the infrastructure supporting these interactions must evolve accordingly. This isn't a partnership announcement. It's a convergence.

Foundation models show promise, but the industry's challenge is clear: harness this power safely and effectively. If agents have wallets, who holds the keys?