There is gain saying that the LLMs are the flavour of the day. The human-like capabilities and conversational skills of LLMs are widely admired. However, with the rapid growth of agentic AI platforms, ironically, LLMs are being used for repetitive, specialised tasks. Is it businesslike to be dependent on LLMs for Agentic AI requirements? Latest is that the researchers from NVIDIA and Georgia Tech argue that the small language models, SLMs, are powerful enough for many agent tasks and provide for a more cost-effective solution than the LLMs. Hence, we find a shift to SLMs.
The shift from LLMs to SLMs marks a pivotal shift in how Agentic AI systems – autonomous, goal-driven software agents – are designed and deployed. What does SLMs have to offer? SLMs (Small Language Models) offer a lightweight, efficient alternative better aligned with real-world agentic needs. The Agentic AI work is believed to be repetitive and the nature of operations simple, though at times gigantic, seems to be ideally suited to small language models. The proponents offer a framework for transitioning from LLMs to SLMs. They are ready to engage in open discussion to encourage more resource-conscious AI deployment.
Unlike generalised LLMs, SLMs can be fine-tuned for narrow domains, tasks and workflows. In the process these SLMs become more predictable, controllable, and aligned with industrial requirements. The researchers define SLMs as models that run efficiently on consumer devices, highlighting their strengths – lower latency, reduced energy consumption and easier customisation. Given the nature of work Agent AI performs the move is from monolithic to modular architecture. Agentic systems require modular intelligence- coordinated smaller models working in tandem instead of an all knowing LLM.
To switch slip over to SLMs in agent-based systems, the process starts by securely collecting usage data while ensuring privacy. In the later stages based on the task needs, more SLMs are chosen and fine-tuned with tailored datasets, often utilizing efficient techniques such as LoRa, the shift to SLMs reflects a maturation of AI .design thinking. The journey has been from brute force intelligence to agile, efficient and specialised agentic intelligence. The future of scalable, real world AI systems most likely lies in SLMs orchestrated within an intelligent, memory rich agentic framework.
THE SHIFT TO SLMs NOT ONLY DEMOCRATISES AI BUT ALSO ENSURES ITS RESPONSIBLE AND SUSTAINABLE USE.
Sanjay Sahay
Have a nice evening.