The Rundown · Tech
Moss Live real-time multimodal voice
Streamers dissect Moss Live: real-time multimodal voice and long-context tradeoffs
Moss Live is a real-time multimodal conversational voice system with foreground conversation separation, background reasoning, voice distinction and meeting participation features; Maticus points viewers to the demo and API and offers a link. Maticus also explained how long attention and training phases affect recall for agent-style long sessions, saying "So losing the far views costs that average very little." Desk line: interesting tech, but long-context handling still shapes real-world behavior.
Takes
Maticus walks through how long-context training affects agents
Maticus explained how long attention and training phases change a model's ability to recall far-back tokens and how that matters for agent-style, long sessions. They described average token perplexity, the penalty from losing long-range context, and how longer training and phase choices can reduce that gap.
Maticus highlights Moss Live for real-time voice
Maticus describes Moss Live, a real-time multimodal conversational voice system that separates foreground conversation from background reasoning and can handle voice distinction and meeting participation; they note there's a demo and API but no model weights mentioned.