In the realm of conversational Artificial Intelligence (AI) research, the prevailing understanding is that augmenting model parameters and training data size significantly enhances the quality and capability of Large Language Models (LLMs). While the current trend involves scaling up models to staggering sizes, with state-of-the-art systems boasting hundreds of billions of parameters, this approach incurs substantial practical costs in terms of inference overheads. The quest for more compact and efficient chat AIs, capable of retaining user engagement and maintaining conversational quality akin to their larger counterparts, remains imperative.
While a single small model may struggle to rival massive state-of-the-art LLMs, an intriguing question emerges: Can a collective of moderately-sized LLMs collaboratively constitute a chat AI with equivalent or superior abilities? Motivated by this query, a recent paper titled “Blending Is All You Need: Cheaper, Better Alternative to Trillion-Parameters LLM” by a research team from the University of Cambridge and University College London introduces the Blended approach.
Blended proposes a simple yet potent method of amalgamating multiple chat AIs by stochastically selecting responses from diverse systems. This approach proves unexpectedly powerful, enabling a consortium of three 6-13B parameter models to outshine the engagement and retention capabilities of the 175B ChatGPT. At its core, Blended leverages the collaborative potential of existing small conversational LLMs to create a unified chat AI generating more captivating and diverse responses.
The Blended approach seeks to approximate samples from the genuine ensemble distribution. Achieving this approximation involves Blended randomly and uniformly selecting the chat AI responsible for generating the current response at each turn. The generated response is contingent on all preceding responses from previously selected chat AIs, allowing for an implicit influence on the output. This collaborative synergy results in a blended response, harnessing the strengths of individual chat AIs to craft an overall more engaging conversation.
The researchers validate the effectiveness of Blended through extensive large-scale A/B tests on real users on the CHAI platform. The findings underscore that a Blended ensemble, comprising three 6-13B parameter LLMs, surpasses the performance of OpenAI’s 175B+ parameter ChatGPT. Notably, blended ensembles exhibit significantly higher user retention, indicating users’ preference for the engaging, entertaining, and useful nature of Blended chat AIs. Remarkably, all these achievements come with a mere fraction of the inference cost and memory overhead typically associated with traditional approaches.
Blended emerges as a groundbreaking approach in the quest for efficient and engaging conversational AI. By strategically blending smaller models, it not only challenges the dominance of massive LLMs but also presents a cost-effective alternative that prioritizes user experience without compromising on performance. As the AI landscape evolves, Blended stands as a testament to the transformative power of collaboration among modestly-sized models, reshaping the future of conversational AI.
The paper Blending Is All You Need: Cheaper, Better Alternative to Trillion-Parameters LLM on arXiv.
Author: Hecate He | Editor: Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.
0 comments on “Beyond Behemoths: How Blended Chat AIs Outshine Trillion-Parameters ChatGPT with Elegance”