- Yandex has announced the open-sourcing of its proprietary Alice AI Search Pretrain model, which underpins the instant answer functionality within its search engine.
- The Alice AI Search team has meticulously refined the entire instant answer pipeline to ensure users receive concise responses within seconds, even during peak usage periods.
- The Alice AI-T5-35B-A0.6B Base model is now publicly available, allowing external users to perform inference via Hugging Face Transformers.
Yandex Details Generative Answer Mechanisms
Yandex has announced the open-sourcing of its proprietary Alice AI Search Pretrain model, which underpins the instant answer functionality within its search engine. This model represents Yandex’s most widely adopted generative product, serving as the initial point of interaction for Search users with Alice.
The Alice AI Search team has meticulously refined the entire instant answer pipeline to ensure users receive concise responses within seconds, even during peak usage periods. The recent June release introduced several significant enhancements:
- Accelerated Responses: Achieved by leveraging shorter information contexts.
- Architectural Innovations: The Encoder-Decoder architecture has been combined with a sparse MoE (Mixture of Experts) architecture for improved efficiency.
- User Signal Training: Online RL (Reinforcement Learning) training, based on real user behavioral signals, has substantially enhanced product quality and utility.
The Alice AI-T5-35B-A0.6B Base model is now publicly available, allowing external users to perform inference via Hugging Face Transformers. However, optimized production-level inference remains exclusively within Yandex’s internal infrastructure for now.
It’s fascinating to see Yandex open-source their Alice AI Search Pretrain model. I’m particularly curious about how they balance the ‘shorter information contexts’ for accelerated responses with maintaining accuracy and comprehensiveness in the instant answers. Does this approach ever risk oversimplification for complex queries? And the combination of Encoder-Decoder with sparse MoE architecture sounds incredibly innovative for efficiency – I wonder if others in the community have seen similar performance gains with this hybrid approach in different NLP tasks?