Editing Openai/691bd6f8-a5e8-8005-9752-1925a60e6a18 (section)

=== 2) What changed with Alexa+ (how training and design differ) ===
* Generative LLM backbone: Amazon says Alexa+ is “infused” with generative LLMs + Alexa’s knowledge base — that means instead of selecting mostly prewritten replies, Alexa+ can synthesize longer, more context-aware text responses on the fly. That requires pretraining/fine-tuning large language models and integrating retrieval (RAG) so answers remain factual and tied to Amazon services. About Amazon<ref>{{cite web|title=About Amazon|url=https://www.aboutamazon.com/news/devices/new-alexa-generative-artificial-intelligence|publisher=About Amazon|access-date=2025-11-20}}</ref>
* Fine-tuning + human preference alignment (RLHF-style workflows): Industry best practice for assistant behavior is supervised fine-tuning (SFT) on high-quality examples and then aligning outputs using human preference data (RLHF or similar). AWS/Amazon publishes tooling and research for RLHF and human feedback pipelines; Amazon’s public materials indicate they use these methods for aligning generative assistants. In short: Alexa+ training pipelines add more SFT + preference learning than earlier Alexa. Amazon Web Services, Inc.<ref>{{cite web|title=Amazon Web Services, Inc.|url=https://aws.amazon.com/blogs/machine-learning/fine-tune-large-language-models-with-reinforcement-learning-from-human-or-ai-feedback/|publisher=Amazon Web Services, Inc.|access-date=2025-11-20}}</ref>
* Richer context & longer memory: Alexa+ is built to hold longer multi-turn context (so replies can reflect longer conversations or user preferences). That requires training on multi-turn conversational data, context windows, and retrieval mechanisms from user-/service-specific data. Amazon has emphasized continuous learning and cloud-based models to support this. About Amazon<ref>{{cite web|title=About Amazon|url=https://www.aboutamazon.com/news/devices/new-alexa-generative-artificial-intelligence|publisher=About Amazon|access-date=2025-11-20}}</ref>
* Tighter integration with Amazon services & on-demand generation: Alexa+ is explicitly designed to synthesize outputs tied to Amazon Music, shopping, bookings, etc., so models are fine-tuned and instrumented to call APIs and produce formatted outputs (tickets, reservations). That means training/testing included many service-oriented examples and safety checks. The Verge<ref>{{cite web|title=The Verge|url=https://www.theverge.com/news/620005/amazon-alexa-plus-ai-smart-assistant-pricing-availability-devices|publisher=The Verge|access-date=2025-11-20}}</ref>
* More cloud dependence and privacy implications: Because Alexa+ uses larger generative models and retrieval from cloud knowledge, Amazon removed some local-only voice routing options; voice data must be sent to the cloud for the new features. Amazon documented that change as part of the Alexa+ rollout. That’s a training/deployment tradeoff (better model capacity vs. increased cloud processing). The Verge<ref>{{cite web|title=The Verge|url=https://www.theverge.com/news/630049/amazon-echo-discontinues-do-not-send-voice-recording-setting|publisher=The Verge|access-date=2025-11-20}}</ref>