Show HN: A real time AI video agent with under 1 second of latency
A new real-time AI video agent with sub‑second latency is impressing developers with how natural the conversations feel, even on lower‑end hardware, while still clearly revealing its artificial nature through glitches, odd mannerisms, and occasional hallucinations. Commenters explore potential applications—from sales, customer support, language tutoring, and education to companionship for the elderly and gaming—alongside skepticism about whether a video avatar actually improves UX over text chat. Many also raise privacy, fraud, and deepfake concerns around giving startups access to their face and voice, and question how such systems will affect work culture, call centers, and norms of human interaction.
Overall Reaction
- Many commenters found the real‑time AI video agent technically impressive, with sub‑second response often described as “future of interaction” or “call centers.”
- Others found it unsettling or “creepy,” especially the realism, constant eye contact, and difficulty hanging up.
UX & Use Cases
- Some users loved the conversational feel, found themselves being polite, and saw potential for:
- Sales agents, technical “co‑pilot” on calls, website reps.
- Language tutors and instructors.
- Mentors, edtech, interactive historical figures.
- Celebrity/influencer “digital twins,” ads, political outreach.
- A substantial group dislikes video calls in general and would prefer text chat or voice‑only.
- Multiple requests for:
- Text input option.
- Less intrusive commenting on personal appearance/background.
- Non‑human or more “cartoony” avatars to avoid uncanny valley.
Demo Experience & Limitations
- Reports of:
- Latency spikes, interruptions, and abrupt session endings (often attributed to “HN hug of death”).
- Avatars nodding or moving heads excessively; lip‑sync and expressions sometimes off.
- LLM unaware of its own visual appearance (e.g., denying wearing a hat) and surroundings in some cases.
- Trouble with name pronunciation and turn‑taking; sometimes cutting users off or misreading emotions.
- Vision features impressed many: reading book titles, noticing pets, decor, text on clothing, etc.
Technical & Infrastructure Discussion
- Stack involves fast LLMs, TTS/STT, and a custom video backbone using Gaussian Splatting rather than NeRF for speed on lower‑end hardware.
- Time‑to‑first‑token and end‑of‑turn detection highlighted as core challenges; discussion of VAD, speech confidence, and speculative decoding.
- Service pricing cited at ~$0.24/minute, billed in 6‑second increments.
- Comments on GPU multiplexing, pooling, and partnerships with GPU infra providers; recognition that costs are high but manageable.
Privacy, Security & Ethics
- Strong concern about giving face/voice to a startup; questions about data storage, cloning, hacking, future misuse, and acquisitions.
- Company reps state:
- No session audio/video stored by default.
- Users retain ownership of their clones; data used only for that clone and deletable.
- Many remain skeptical, referencing low trust in AI firms and analogies to genetic‑data companies.
- Broader worries about fraud, deepfakes, political manipulation, loneliness, and the morality of “owning” lifelike digital entities.