The small language models are reflex algorithms.

Researchers are interested in small language models. The reason for that is simple. Large language models include thousands of billions of parameters. Those parameters require lots of computer capacity. And those computers use lots of electricity. Training for the LLM costs billions of dollars in computer and electric bills. And the biggest problem is that the LLM requires a big data center. 

Normal companies have no money or other resources to run the LLM on their own servers  And in the ICT world those tools can cause data security problems. The LLM that runs on the Microsoft or Open AI servers runs on machines. That is under the potential competitor's control. The LLM is a good tool but it requires lots of power. 

If we want to operate robots independently using the LLM. We must be sure. The robot has an internet socket connected to the central computer. In that model, the robot sends orders first to the computer center. There the LLM transforms those orders into actions for the robot. That system is useless if something disturbs it. 

The robot cannot keep the connection in electromagnetic fields. Robots are planned to be used in high-risk environments like nuclear accidents and military work. The remote control is easy to jam, and that's why researchers in military and civil fields search for systems that can be compact and locally operated. 

One of the problems with LLM is this. Those systems search data from the entire internet. That makes them good tools for making things like doctoral theses. But if the AI-controlled robot uses that model it requires lots of energy and those things are slow. 

The robot requires the reflex algorithm. The RISC systems have a limited number of databases. That makes them compact and effective. The RISC system called the small language model, SLM is the tool that can make robots safer. And allows them to operate independently. 

Those reflex command bases can involve responses to things like: "Would you step away from the door". In the same way, the jet fighter cannot ask for advice from the computer centers if it sees an incoming missile. 

The robot must not make contact with the computer center. Every time, when it must react to some everyday things. When a robot hears something it must realize that it must not react to everything 

So, if we want to create an AI that can operate in a complicated environment we must modify the LLMs and create a lightweight version. The lightweight LLM or small language model SLM has only a couple of million or even less than a million algorithms. Sometimes those SLMs called RISC-language models. RISC (Reduced instruction set computer) systems are like pocket calculators. They might be more limited than large systems. But they are fast-reacting and they do their job very fast. 

Those lightweight language models can run on regular servers. They react very fast because they have only limited action libraries. The small language model can be installed on the aircraft's computer. And it can act as an assistant for the pilot. 

https://www.quantamagazine.org/why-do-researchers-care-about-small-language-models-20250310/

https://en.wikipedia.org/wiki/Reduced_instruction_set_computer