Faster, smarter, more powerful, but also more expensive. And frankly, it is almost impossible to keep track of which model is best for which task. In the Artific GenAI Research Lab we see many organisations struggling with that same challenge: which model should you choose for which task?
What choosing a language model affects
The language model you choose always affects three things:
- how quickly you get an answer (latency);
- how good that answer is (quality);
- and what it costs in computing power, but also in energy.
Yet most teams simply use the newest, heaviest model. That often means waiting longer for an answer and wasting energy on questions a simpler model could have answered perfectly well.
The right model for every task
It is like choosing how to travel: you don’t always need a racing car to reach your destination. Right now, though, customers usually grab the biggest car with the most options, when sometimes you would get there a lot faster by bike.
That is why at Artific we are building what I see as a traffic system for AI. A system that decides for itself: does this route need a racing car, or will another way of travelling do? In other words, this module can select the best language model for a task.
That way customers save time, money and energy, without the quality of the answers suffering.
Results of automatic model selection
ChatGPT has a similar feature. With GPT-5 you can switch between three modes yourself: auto, instant or thinking. So the idea is not new, but the execution is.
Around the world, a great deal of research is going into how to select a model for a task reliably without too much overhead in cost and response time. In the Artific GenAI Research Lab we have combined the ideas from several scientific publications and developed them further. The first results are promising:
- 50% lower costs on average
- 60% shorter loading times
- considerable energy savings
And all of that without any noticeable loss of quality or response time.
Clustering questions
We achieved this by grouping questions into clusters. Each cluster stands for a type of task, such as summarising, analysing or translating. For each cluster we know which AI model performs that kind of task best. When someone then asks the assistant a question that belongs to one of those clusters, we automatically route it to the model best suited to it.
Partners benefit from automatic model selection too
Our partners benefit from this development as well, because they can add their own clusters within our platform. That way the right model is chosen automatically for their very customer-specific tasks too.
This is how we make sure customers always use the right model: fast when it can be, powerful when it has to be.
More about Artific
Want to know more about Artific, our AI platform or the GenAI Research Lab? Get in touch with us at info@artific.nl or 053-203 0123.