Yandex has detailed the architecture and training process behind its Alice AI Fast Answer, a large-scale generative product integrated into its search engine. The system is designed to provide concise answers within seconds, even under peak load. Key improvements in the latest release include faster response times through short infocontexts, a hybrid Encoder-Decoder and sparse MoE architecture, and learning from real user signals to enhance quality and usage. Yandex has also open-sourced the Alice AI-T5-35B-A0.6B Base model, offering inference via Hugging Face Transformers. AI
IMPACT This release provides insights into large-scale generative search product architecture and offers an open-source model for researchers.
RANK_REASON The item details the architecture and training of a generative AI model and releases a pre-trained model. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Mastodon — mastodon.social →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
