Can smaller models compete with the big players in chat technology? – IT BOLTWISE® x Artificial Intelligence

CAMBRIDGE/MUNICH (IT BOLTWISE) – A new study examines whether combining smaller models in chat technology can be as effective or even better than one large model.

In the rapidly evolving field of chatbot technology, there is a trend towards massive AI models such as ChatGPT, Bard, and Gemini. These giants are known for their extensive training data and many model parameters aimed at significantly improving their performance. However, their large size also means that they require a significant amount of computing power, raising questions about their efficiency.

In one Research from the Universities of Cambridge and UCL It explores an interesting possibility: Could a group of smaller models be as effective or even better than a single large model?

The researchers present a new method called “scrambled,” which involves randomly selecting answers from a variety of small chat AI systems. This approach surprisingly outperforms larger models. Blended AI combines the best features of different systems and learns and adapts based on conversation history. This results in more diverse and engaging responses that improve the user experience. The effectiveness of Blended has been confirmed through extensive A/B testing on the CHAI platform with real users.

Designing chat AI involves defining the basic model and data needed for fine-tuning and type of feedback. Different combinations can lead to diverse systems with unique strengths. This study proposes mixing different chat AI systems, each with its own parameters, in a way consistent with Bayesian statistics.

The effectiveness of Blended AI is evaluated not only by the quality of its responses, but also by measuring user engagement and retention. These are reliable indicators of chat AI performance. In tests, Blended, which consists of different smaller AI models, showed even higher user engagement and retention rates than GPT-3.5, despite lower parameters and lower computational costs.

In summary, this study suggests that mixing smaller-scale and open source systems is an effective way to improve the chatbot experience without relying on very large models. This approach challenges the traditional view that bigger is always better in developing AI models and suggests a future where collaboration between models could be crucial.

Please send any additions and information to the editorial team via email to de-info[at]it-boltwise.de