Summary
Sidney Primas, CTO and founder of LemonSlice, lays out his company’s technical approach to what he calls the “Avatar Turing test” — building a full-body photorealistic avatar that is indistinguishable from a human on a live video call. The talk opens with a concrete milestone: a partnership with Microsoft to deploy a real-time interactive Teddy Roosevelt avatar inside a replica Oval Office at the Roosevelt Presidential Library, which drew an unplanned 10-minute session from a high-profile visitor on launch day.
The core architectural bet is to use human-focused world models rather than the patchwork pipeline approaches most avatar companies use, trading higher upfront training cost for emergent properties like full-body motion, object interaction, and micro-expressions. Primas details two hard engineering problems: making the autoregressive video model causal (inference can only look backwards in time) and distilling 30-step denoising down to a single step for real-time throughput. He also describes a proprietary method for preventing error accumulation — the compounding artifact problem that degrades quality over long continuous runs — which allows LemonSlice avatars to generate for 8 to 16 hours without a reset.
Hardware cost reduction is framed as a prerequisite for commercial viability: continuous AI video generation at scale demands economics very different from a 5-second clip. Primas closes by arguing that the long-term default modality for human-AI interaction will include a visual layer, making photorealistic avatars a foundational infrastructure problem rather than a novelty.
📺 Source: AI Engineer · Published August 18, 2026
🏷️ Format: Keynote Launch
