AI digest: voice agents get faster, trust gets thinner
Salesforce cracks voice latency while Americans lose faith in AI outputs despite using the tools more.
The big tech releases keep coming, but the real action is in making AI actually work for users who don’t trust it.
Salesforce cuts voice agent latency by 316x
Salesforce’s VoiceAgentRAG solves the 200ms problem for voice AI. Standard vector database queries take too long for natural conversation, so they built a dual-agent memory router that dramatically speeds up retrieval. This matters because voice interfaces are where AI gets really useful, but only if they don’t pause awkwardly every few seconds.
Alibaba ships proper multimodal with Qwen3.5 Omni
The Qwen team released a native multimodal model that handles text, audio, and video without stitching separate encoders together. It’s positioned as a Gemini 3.1 Pro competitor, which suggests Alibaba thinks they can match Google’s flagship multimodal capabilities. We’re seeing the end of the “wrapper” era where vision and audio were bolted onto text models.
OpenAI’s Sora burns cash and users flee
Sora burned a million dollars daily while losing half its users, forcing OpenAI to shut it down. The company is redirecting those resources toward coding and enterprise tools. Turns out consumer video generation isn’t ready for the compute costs, even when you’re OpenAI.
Americans use AI more but trust it less
A Quinnipiac poll shows AI adoption is rising whilst trust drops. People are finding these tools useful enough to keep using them despite not believing the outputs. That’s actually a healthy dynamic, much better than blind faith in AI responses.