Exploring smaller and more efficient LLMs
LLMs show incredible capabilities but are also associated with large costs beyond training costs. Expensive infrastructure is also required for deployment, not to mention the costs associated with simple inference that grows with the number of parameters. These large LLMs are generalist models, and for many tasks, it is not necessary to have a model that has 100 billion parameters. Especially for many business cases, we need a model that can accomplish a specific task well. So, there are many cases where a small language model (SLM) is sufficient.
SLMs tend to excel in specialized domains, and may therefore lose the contextual informativeness that comes from integrating various domains of knowledge. SLMs may lose some of the capabilities of LLMs or otherwise exhibit fewer reasoning skills (thus being less versatile). On the other hand, they consume far fewer resources and can be used on a commercial GPU or even CPU (or, in extreme cases...
