Luke Curley points out WebRTC's limitations

In response to OpenAI's "How OpenAI delivers low-latency voice AI at scale," Luke Curley pointed out that WebRTC is designed to drop prompts and audio when the network degrades.

He explained that this aggressive dropping of audio packets to reduce latency makes sense in environments like video conferencing, where immediate round trips matter. However, for LLM prompts, he believes it's better for users to wait a bit longer and get a more accurate response.

• He said that for a slow, expensive prompt, he'd rather wait an extra 200ms.
• However, retransmitting WebRTC audio packets within a browser is practically difficult, and he noted that Discord ran into the same limitation.
• Ultimately, the core issue is that the implementation itself is locked into prioritizing real-time latency, leaving almost no option to wait for the sake of accuracy.

Simon Willison quoted and posted this remark on May 9, 2026.