Gemini 3.8 Live with Live Avatar is now generally available for enterprise customers, delivering a production-ready, native speech-to-speech foundation that combines real-time audio, video avatars, and live visual understanding. Key capabilities include synchronized video avatars (custom avatars via allowlist), fluid interruption recovery without losing conversation context, background tool and API calling while the agent continues speaking, and support for 97 languages with automatic detection. The system can ingest camera feeds and screen shares alongside audio, enabling agents to see and act on what users show in real time. Enterprise features include US and EU endpoints, provisioned throughput, compliance and data governance, and imperceptible SynthID watermarks on generated audio/video to ensure provenance.
Demonstrations show building a custom avatar from a photo and audio sample, a claims-intake workflow that auto-fills forms while analyzing live video, and a real-time voice agent using the Agent Development Kit to stream audio to the Gemini Live API without routing through speech-to-text. Early customers report practical gains: Autotrader uses screen-highlighting and tool-calls for guided shopping, Equal AI cites better interruption handling across Indian languages, and partners note latency and VAD improvements. The offering is available in Gemini Enterprise with API docs, pricing, and guidance for provisioning and custom-avatar allowlisting; sales teams can assist with deployments.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Tesla's Full Self-Driving system exceeded speed limits in 55% of Brussels' 30 km/h zones, averaging 44 km/h. The Belgian road safety group Johanna.be raised concerns ahead of an EU vote on FSD approval, citing safety risks and misidentification of speed zones. (electrek.co)
GitHub Copilot is designed to enhance coding efficiency, but there is a gap between its potential and real-world productivity improvements. The article examines this efficiency-throughput gap in AI-assisted programming tools. (cacm.acm.org)
Large language models (LLMs) are shown to have significant limitations in performing accurate arithmetic. The article recommends using humans or specialized tools for calculations instead of relying on LLMs. (medium.com)
David Heinemeier Hansson announced his shift towards AI-driven development, emphasizing LLMs replacing traditional coding and native applications. He predicts that most programming will rely on AI and encourages embracing this technological transformation. (jardo.dev)
TypeSafe AI's Jev is a non-autoregressive model that provides structured answers with calibrated probabilities, prioritizing calibration over raw accuracy. It offers significant speed improvements and fits into production environments by answering multiple question types simultaneously. (kartikpansuriya.com)
The United States is investing heavily in artificial intelligence development, considering it the largest economic opportunity in history. This AI build-out aims to transform industries and boost economic growth on a national scale. (wsj.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.