100% agree. Maximizing intimacy and scaling distribution pull in opposite directions. We’re experimenting with keeping the “character” consistent while letting personalization live in private memory and user-controlled settings. Still early, and this tension is real.
Appreciate it. If you try it and anything feels off (latency, turn-taking, uncanny moments), I’d love concrete feedback. That’s what we’re grinding on right now.
realtime api + elevenlabs but llms will be diversified based on persona moving forward. Using chatgpt/gemini as baseline model, we feel prompting has limitation
Totally fair reaction. We’re building this with clear boundaries: we don’t position it as therapy replacement, we add safety rails, and gives user a choice what mode they want and guardrails differ based on this. Plus, age restriction is there as safety boundary
For real-time: we use WebRTC for streaming. Input is streaming STT, then a low-latency LLM, then TTS, then we drive Live2D parameters on the client.
Lip sync: we currently do (simple phoneme / amplitude-based) and are testing viseme extraction. Rhubarb is on our list, but we’re cautious about added latency.
We’re not giving it unconstrained tool access. In-product, actions are either not available or gated behind explicit user intent and strict allowlists. The interesting part for us is the real-time conversational loop and memory personalization, not autonomous exploration.
I’m not trying to make any claims about consciousness. For us, the practical question is: does the interaction feel supportive and useful, while staying transparent that it’s a model. The rest is philosophy, and I’m happy to read more perspectives.
I think “parasocial” still captures part of it (one-to-many distribution, performer vibe), but there’s also a true interactive dyad here. It’s closer to “synthetic social interaction” or “responsive parasocial.” I don’t have a perfect word yet, but the asymmetry and the responsiveness both matter.