The increasingly impressive results of LLM models are truly captivating. However, some voices suggest that what we see now is merely a foundation for far more complex achievements yet to come. Visionaries capable of imagining what lies ahead often stir a sense of unease among audiences. So, what is the concern?
Daniel Kahneman, a psychologist and Nobel laureate in economics, described in his book Thinking, Fast and Slow a theory of two systems of thinking: System 1 – fast, intuitive, and reactive, and System 2 – slower, more analytical, and autonomous. Current language models largely function in line with System 1, relying on patterns and reflexes. Only by reaching the level of System 2 – with autonomous reasoning and the capacity for self-improvement – will the full potential of this technology be realized. This prospect may both amaze and alarm us.
Undeterred by these future challenges, we decided to create a "dream team" of language models that work together and complement each other on a single platform. Today, I'll show how we've managed to integrate multiple models from various providers to build a system with ever-improving efficiency and operational cohesion.
The Dream Team
The first and key element of the team is an algorithm responsible for delivering the primary service – in this case, providing IT advice, though this is only one example. Its effectiveness and quality are the results of the entire team’s collaborative efforts, constantly fine-tuning and optimizing its performance. The remaining team members play various essential roles:
- User Agent – Represents the end-user. It presents challenges, is demanding, and consistently strives to achieve its objectives.
- Arbiter – A strict judge of interaction quality. It evaluates responses according to eight criteria, including substance, format, response length, clarity, ethics, and response time.
- Teacher – Analyzes the Arbiter’s feedback and determines which changes could enhance the system. Based on these insights, it recommends improvements and oversees their implementation.
Each of these agents is an algorithm with varying levels of complexity, each of which has evolved over time. The models come from different "technology stables": included in the team are Facebook’s LLaMA, Google’s Gemma, and Poland’s Bielik, among others. All the models have undergone numerous tests and optimizations to best fulfill their assigned roles. Today, we can confidently say that this "dream team" is functioning more effectively than ever. 😊
Collaboration Mechanism
The internal mechanisms of the team resemble principles of reinforcement learning on a larger scale. This approach allows for systematic automation of testing and continuous improvement of the Personal Advisor. The result is a hybrid advisor that outperforms each of its individual components – much like the collective effort of a team surpasses the capabilities of any single member.
Author: Mariusz Jażdżyk
More info about our product: Personal Advisor
