Opinion: It is time to virtualize foundation models with a self‑evolving OS layer
A new arXiv position paper argues that AI development has moved from single, monolithic foundation models to complex, agentic systems, but the supporting software stacks remain fragmented. The authors note that each framework now embeds its own runtime for state, memory, budgeting, and guardrails, making behavior hard to port and governance brittle—much like early computing before operating…
Key points
- Current AI stacks embed separate runtimes for state, memory, budgets, and guardrails, hindering portability.
- Proposed FMOS would virtualize foundation model interactions, handling model selection, resource allocation, and policy enforcement.
- FMOS would learn when to intervene in inference, adapting policies from operational experience.
The paper proposes a Foundation Model Operating System (FMOS) that would virtualize interactions with foundation models, similar to how virtual machines abstract physical hardware. FMOS would manage memory tiers, select appropriate models, allocate resources, and enforce verification and policy rules. It would also learn when to intervene in inference versus letting it run directly, continuously adapting its policies based on operational experience. The authors liken this to the brain’s fast intuition versus slow deliberation, suggesting a self‑evolving layer that gives applications the illusion of dedicated, trustworthy model instances with effectively unbounded capabilities.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.
More in Research
All →- New framework optimizes LLM inference costs via adaptive model activation · 4 src
- Qwen3.5-4B outperforms larger LLMs on new user-side conflict benchmark · 1 src
- Neo-Classic benchmark evaluates linguistic-aesthetic reasoning in Classical Chinese poetry · 1 src
- Study finds trust and friction issues in major generative AI app reviews · 1 src
- Study finds PCA can detect stylistic axes in LLM activations without training · 1 src
Comments
via GitHub Discussions