Large language models (LLM) are impressive. But can the small, compact language model be more effective?

Large language models (LLM) are impressive. But can the small, compact language model be more effective? 

Large language models can search and process any kind of data. That we might need. The thing that makes LLM a little bit complicated. They must analyze the data that they get. And that requires lots of calculation power. That power is not free. The LLM requires an entire group of supercomputers. 

The problem is that those LLMs require the most powerful hardware in the world. Large supercomputers require lots of electricity. 

Those systems use as much electric power as a small city. That means the LLM servers require their own power plants. That's why. Things like miniature nuclear reactors are under research. 

The only realistic way to use LLM is to use them over the net. The private actor doesn't have money for the electric bills that those servers require. And that brings one data vulnerability in the data handling process. When some other actor operates the server, where we share data. That another actor can get access to that data. The ability to bring the language model to its own servers gives the actor full access to the data. That means that the actor who administrates their servers owns and knows everything that the server does. 

The small language models (SLM) can act as multicore AI there each core can operate independently. 

The small language model (SLM) is the answer to that problem. The SLM can be like a droplet that is separated from the LLM. That lightweight language model is the tool that a normal actor can load to the power servers that are used in the CAD program's network use. The SLM is a limited language model used for special purposes. So the LLM can remain in the common use. But in secured use, the SML can be the tool, that is in the future. 

The SLMs can create a new model for the AI core. The SLMs can make it possible to create a multi-core AI system. Those cores can remain on their servers. They can operate independently without affecting other processes. 

The multicore models are tools that can turn the next page for the AI. That system can be more economical and environmentally friendly than the monolithic core. The system can switch unnecessary cores off when the system or users don't need them. 

https://www.wired.com/story/why-researchers-are-turning-to-small-language-models/