Quoting Luke Curley

Origin: A Complaint from an OpenAI Engineer The story begins with a pointed "wtf" from OpenAI engineer Luke Curley while discussing how the company delivers low-latency voice AI at scale. He highlighted that WebRTC, a protocol designed for real-time audio/video calls, has a core behavior—actively dropping audio packets during poor network conditions to maintain low latency—that runs counter to the needs of AI voice interaction. For users paying a premium to "boil the ocean" with complex prompts, they would rather wait an extra 200 milliseconds to ensure every word sent to the AI is accurate. However, the browser implementation of WebRTC is hard-coded, making it impossible to even retransmit a dropped packet. This isn't a technical capability issue but a fundamental conflict in underlying design philosophy. Breakdown: A Clash of Two 'Concepts of Time' To understand this矛盾, we need to break down two distinct "concepts of time." For traditional real-time calls (like Zoom, Tencent Meeting), the gold standard for user experience is "immediacy." The rhythm of conversation, interruptions, and emotional conveyance all rely on ultra-low latency. Occasional audio distortion or dropped words are often tolerable as long as the conversation continues. WebRTC was built for this; its "aggressive packet dropping" strategy is like a decisive traffic dispatcher willing to let a few vehicles (packets) disappear to keep the main artery (conversation flow) absolutely clear. But for AI voice interaction, the scenario changes. Users aren't engaging in rapid back-and-forth with another person; they are submitting a "query order" to a "super brain," which might contain complex instructions, long context, or critical information. The quality of this order (the prompt) directly determines the quality of the AI's response. Here, accuracy trumps immediacy. The user's mental model is "I make a request -> I wait -> I get a high-quality response." The waiting period in between, as long as it's not excessively long, is fully acceptable and expected. Luke's complaint精准地抓住了这一点: AI isn't particularly responsive anyway, so why is WebRTC so concerned? Trend Insight: AI-Native Applications Need 'Protocol-Level' Rethinking This incident is far more than a minor technical selection hiccup. It reveals a deeper trend: When we build applications with AI as the core engine, many tech stacks and protocols designed for the "pre-AI era" need to be re-examined. We are moving from an era of "connecting people" (real-time communication) to one of "connecting people with AI" (intelligent interaction). The communication pattern for the latter might be more akin to an "asynchronous, reliable, high-fidelity request-response" model rather than pure "real-time streaming." This could mean:

1. Need for new transport protocols or extensions: The industry may need to design or extend protocols specifically for AI voice/video input, prioritizing data integrity and accuracy over absolute low latency (e.g., within 500ms), with support for intelligent retransmission and error correction. 2. Edge computing and intelligent buffering: More intelligent buffering and preprocessing of user voice input at the client or edge node to ensure data packets are complete and ordered before being sent to the cloud-based large model. 3. Application-layer compensation: Within the constraints of existing protocols, developers may need to design cleverer solutions, such as proactively prompting users "ensuring your instruction is fully delivered" when network波动 is detected, or even allowing users to manually choose modes like "prioritize accuracy" or "prioritize speed." Practical Value and Counter-Intuition For teams developing AI voice features or agents, this is a crucial警示: Do not assume that the technology best suited for "real-time" is also best suited for "AI real-time." During technical selection, it's essential to deeply understand the core needs of AI interaction. Directly adopting WebRTC might plant user experience隐患 at the底层 level. The most counter-intuitive aspect of this story is that we often equate "fast" with "good," but in specific环节 of AI interaction, "accurate" is far more important than "fast." This forces us to redefine what constitutes a "good" communication experience in the AI era—it may no longer be a millisecond-level latency race, but achieving 100% intent-fidelity transmission within a user-acceptable delay range. Luke's complaint might just be the first trumpet blast heralding this protocol-level revolution.