--- schema: 1 type: video title: 'Voice agents with Realtime Video — Sidney Primas, LemonSlice' description: >- An avatar of Teddy Roosevelt holds court in a replica Oval Office, generating video continuously for eight hours with no reset, and a second deployment is being built to run for sixteen. That duration is the hard part. Sidney Primas explains that a real time avatar can only look backward, because the future frames do not exist yet, so every block it generates inherits the errors of the blocks before it and compounds them. LemonSlice trains with an attention mask that enforces this during training rather than discovering it at inference, and collapses roughly 30 denoising steps down to a single step to hit real time. The less obvious bottleneck is audio. Emotion and facial expression turn out to depend on the audio embedding, and most audio encoders are trained on audiobooks, which are monotone by construction, so an expressive model needs its own. The wider bet is to take a world model and point it at humans, paying a harder training and deployment cost up front in exchange for full body movement, object interaction, and physics arriving closer to free. Two things surprised him. Serving this costs about what serving a voice model costs, despite the difference in pixels. And the model harness, meaning the orchestration of threads and queues across GPU and CPU so that video never stutters through an interrupt, is where he now thinks much of the durable value will sit. Speaker info: - https://www.linkedin.com/in/sidneyprimas/ Timestamps: 0:00 - Breaking the avatar Turing test 2:26 - Teddy Roosevelt in a replica Oval Office 4:40 - Why the visual layer matters 5:46 - Pointing a world model at humans 6:58 - One image in, any style out, and being the API layer 9:12 - Audio is what makes it expressive 10:15 - Making a video model interactive, then real time 12:22 - Error accumulation over hours of generation 14:34 - Cost parity with a voice model 15:36 - The model harness nobody talks about 16:38 - An emotion engine for the next model 19:55 - A single end to end EQ layer 22:09 - Questions: internal state, and a real Turing test canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/voice-agents-with-realtime-video-sidney-primas-lemo-c8a9ff15ea80 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/voice-agents-with-realtime-video-sidney-primas-lemo-c8a9ff15ea80.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T05:13:33.147Z' language: null children: [] videoId: z1dqv74SpUs durationSeconds: 1596 publishedAt: '2026-08-18T16:00:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 12 endSeconds: 23 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=12' heldReason: rights - ordinal: 1 text: null startSeconds: 23 endSeconds: 148 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=23' heldReason: rights - ordinal: 2 text: null startSeconds: 159 endSeconds: 290 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=159' heldReason: rights - ordinal: 3 text: null startSeconds: 302 endSeconds: 348 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=302' heldReason: rights - ordinal: 4 text: null startSeconds: 423 endSeconds: 434 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=423' heldReason: rights - ordinal: 5 text: null startSeconds: 434 endSeconds: 556 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=434' heldReason: rights - ordinal: 6 text: null startSeconds: 434 endSeconds: 556 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=434' heldReason: rights - ordinal: 7 text: null startSeconds: 556 endSeconds: 567 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=556' heldReason: rights - ordinal: 8 text: null startSeconds: 567 endSeconds: 623 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=567' heldReason: rights - ordinal: 9 text: null startSeconds: 635 endSeconds: 748 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=635' heldReason: rights - ordinal: 10 text: null startSeconds: 635 endSeconds: 748 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=635' heldReason: rights - ordinal: 11 text: null startSeconds: 748 endSeconds: 760 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=748' heldReason: rights - ordinal: 12 text: null startSeconds: 760 endSeconds: 878 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=760' heldReason: rights - ordinal: 13 text: null startSeconds: 760 endSeconds: 878 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=760' heldReason: rights - ordinal: 14 text: null startSeconds: 888 endSeconds: 941 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=888' heldReason: rights - ordinal: 15 text: null startSeconds: 1012 endSeconds: 1204 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=1012' heldReason: rights - ordinal: 16 text: null startSeconds: 1204 endSeconds: 1215 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=1204' heldReason: rights - ordinal: 17 text: null startSeconds: 1215 endSeconds: 1337 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=1215' heldReason: rights - ordinal: 18 text: null startSeconds: 1351 endSeconds: 1596 watch: 'https://liquidgold.cc/watch/z1dqv74SpUs?t=1351' heldReason: rights --- # Voice agents with Realtime Video — Sidney Primas, LemonSlice An avatar of Teddy Roosevelt holds court in a replica Oval Office, generating video continuously for eight hours with no reset, and a second deployment is being built to run for sixteen. That duration is the hard part. Sidney Primas explains that a real time avatar can only look backward, because the future frames do not exist yet, so every block it generates inherits the errors of the blocks before it and compounds them. LemonSlice trains with an attention mask that enforces this during training rather than discovering it at inference, and collapses roughly 30 denoising steps down to a single step to hit real time. The less obvious bottleneck is audio. Emotion and facial expression turn out to depend on the audio embedding, and most audio encoders are trained on audiobooks, which are monotone by construction, so an expressive model needs its own. The wider bet is to take a world model and point it at humans, paying a harder training and deployment cost up front in exchange for full body movement, object interaction, and physics arriving closer to free. Two things surprised him. Serving this costs about what serving a voice model costs, despite the difference in pixels. And the model harness, meaning the orchestration of threads and queues across GPU and CPU so that video never stutters through an interrupt, is where he now thinks much of the durable value will sit. Speaker info: - https://www.linkedin.com/in/sidneyprimas/ Timestamps: 0:00 - Breaking the avatar Turing test 2:26 - Teddy Roosevelt in a replica Oval Office 4:40 - Why the visual layer matters 5:46 - Pointing a world model at humans 6:58 - One image in, any style out, and being the API layer 9:12 - Audio is what makes it expressive 10:15 - Making a video model interactive, then real time 12:22 - Error accumulation over hours of generation 14:34 - Cost parity with a voice model 15:36 - The model harness nobody talks about 16:38 - An emotion engine for the next model 19:55 - A single end to end EQ layer 22:09 - Questions: internal state, and a real Turing test Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/voice-agents-with-realtime-video-sidney-primas-lemo-c8a9ff15ea80 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 2:39) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (2:39 to 5:02) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (5:02 to 7:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (7:03 to 9:16) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (9:16 to 10:23) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (10:23 to 12:28) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (12:28 to 14:38) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (14:38 to 15:41) > Held: transcript rights do not permit this chunk's text.