Low data, more intelligence: the power of SLMS – opinion

By Nuno Ruvo

The consequence of manufacturing artificial intelligence has brought many cases of utility, some with great success, and some other results have fallen to expectations. This expansion of AI-based systems is widely driven by linguistic models, and the race of the main manufacturers to develop the best language model in the world is proof of this. GPT (Openai/Microsoft), Gemini (Google), Gama Granite (IBM), Claude (Anthropic), Deepsek, Mistrol, Lama, along with others, are most common: they are mostly LLMs (large language models). But what does it mean? Generally, they are not only understanding the language, but also have huge information on multiple themes, in practice, “…There is everything and I don’t know what to do! ”

. In theory, it is very good, but in practice it raises a basic question: Do we need a very comprehensive model for all applications?

The answer is a round no. If I run the video AI system expert vehicle and claim analytics in claims, do I also know how to give cod recipes? Or, if my company works in the legal field, does it make sense to have knowledge of the model about Quantum Mechanics? Sharing the expression used by the client, the only speaking model of “dinners”, I don’t have to talk about telecommunications or other fields.

The reality is that only this additional data is occupied and requires high computational sources to support the implementation of these models. This is equivalent to the use of a ant to kill a ant: the waste of resources.

Alternatively, SLMs (small language models). These are trained for smaller designs, especially for a domain or knowledge area, and become real professionals in their field, without losing in Ramblings.

• There are many benefits of choosing SLMs instead of LLMs:Lightly infrastructure
• For those who seek on-enclosure solution, the hardware needed to drive SLM is significantly lower, receiving the AI ​​is more viable, especially for companies with security and privacy problems.Low cost
• .The lower likelihood of hallucinations
• : Since SLMs work with more limited and specific data sets, the possibility of providing irrelevant or incorrect answers is reduced.Sustainability

: This theme is also on the agenda and it can cause many companies. The carbon footprint associated with the training and implementation of SLM is much smaller than LLM, which allows the environment and effective approach to AI.

So why are we using LLMs?

In Portugal, the main reason for receiving SLMSs is the lack of special models that understand and properly process European Portuguese. Also, time and testing for every need to find the right model, which leads to the choice of the simplest path of many companies: resorting to an LLM and filtering irrelevant information.

Another alternative is to train specific models, but this approach is not yet practical for most organizations, because of costs and complexity and lack of human, physical and economic abilities.

Future: a more unique scenario

Studies have shown that betting on special SLMs will be a trend in the coming years, some companies have resorted to small and small models for certain tasks. In Portugal, this change is driven by the first major national SLM and Aamalia development. If successful, it represents a landmark in the way we receive the IA, which can be more efficient, accessible and will meet our needs.

At the end of the day, the question remains: do we want a little known patterns without specialization, or small, or centralized and effective models? The future of our projects can find this balance.