--- schema: 1 type: video title: >- SOTA Generative Media Panel — Dumitru Erhan, Shane Gu & Nicole Brichtova, Google DeepMind description: >- The team took captions from real videos, regenerated the same scenes with their own model, and ran a human eval. People largely preferred the generated version. Dumitru Erhan is quick to deflate that result. The output is not more realistic, it is sharper and more saturated with nicer skin tone, and human preference turns out to be an unreliable thing to optimize against. Nicole Brichtova has a name for the effect, the Instagram filter. The same blind spot surfaces in smaller ways. Their image model quietly began putting wedding rings on hands, and nobody internally caught it until an outside tester asked why, which is reward hacking arriving through the back door. Much of the hour circles one question, whether language is a good enough intermediate representation for any of this. Shane Gu argues it thins out exactly where humans are most sensitive. We carry poor vocabulary for audio, taste, smell and skin tone because those sit close to survival, and a professional wine taster he asked had resorted to borrowing the language people use to describe a date. The tell of AI video, swyx notes, is that everything sounds studio recorded, because studio recordings are the training data, and a model with no representation of standing further away cannot get that sound right either. Evaluation stays stubbornly manual underneath all of it. Ten people in a room, two videos side by side, pick one. Speaker info: Dumitru Erhan: - https://x.com/doomie - https://www.linkedin.com/in/dumitruerhan/ Shane Gu: - https://x.com/shaneguML - https://www.linkedin.com/in/shixiang-shane-gu/ Nicole Brichtova: - https://x.com/nbrichtova - https://www.linkedin.com/in/nicolebrichtova/ Timestamps: 0:00 - Introductions, and why this session exists 2:00 - Nano Banana 2 Lite and the Omni Flash APIs 4:30 - Beyond demos: storyboards, editing, education 7:51 - Video agents, or one model doing it all 10:26 - Video models as zero shot learners 13:01 - Will everything collapse into one model? 14:20 - Is captioning the right intermediate representation? 18:13 - What people actually mean by world models 19:44 - Vision researchers becoming generation researchers 24:51 - Audio, and why joint generation mattered 26:32 - The things language cannot describe 32:03 - Studio quality as the tell of AI video 34:12 - When humans prefer the generated video 36:07 - Prompt engineering and trained sensitivity 38:27 - Setting a default aesthetic for the world 41:09 - The wedding ring nobody noticed 42:10 - How you actually evaluate video 47:35 - What data they are looking for 53:04 - Forward deployed engineers as an eval channel --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/KLDdXOw6jIc-youtube-b01c3c14dd8f8e90af4c Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/sota-generative-media-panel-dumitru-erhan-shane-gu-59dcef3eb6c3 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/sota-generative-media-panel-dumitru-erhan-shane-gu-59dcef3eb6c3.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T04:17:00.669Z' language: null children: [] videoId: KLDdXOw6jIc durationSeconds: 3419 publishedAt: '2026-08-30T14:00:06.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 126 endSeconds: 137 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=126' heldReason: rights - ordinal: 1 text: null startSeconds: 137 endSeconds: 280 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=137' heldReason: rights - ordinal: 2 text: null startSeconds: 290 endSeconds: 471 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=290' heldReason: rights - ordinal: 3 text: null startSeconds: 290 endSeconds: 471 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=290' heldReason: rights - ordinal: 4 text: null startSeconds: 290 endSeconds: 471 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=290' heldReason: rights - ordinal: 5 text: null startSeconds: 290 endSeconds: 471 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=290' heldReason: rights - ordinal: 6 text: null startSeconds: 482 endSeconds: 630 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=482' heldReason: rights - ordinal: 7 text: null startSeconds: 482 endSeconds: 630 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=482' heldReason: rights - ordinal: 8 text: null startSeconds: 630 endSeconds: 641 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=630' heldReason: rights - ordinal: 9 text: null startSeconds: 641 endSeconds: 784 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=641' heldReason: rights - ordinal: 10 text: null startSeconds: 784 endSeconds: 796 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=784' heldReason: rights - ordinal: 11 text: null startSeconds: 796 endSeconds: 862 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=796' heldReason: rights - ordinal: 12 text: null startSeconds: 873 endSeconds: 1097 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=873' heldReason: rights - ordinal: 13 text: null startSeconds: 873 endSeconds: 1097 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=873' heldReason: rights - ordinal: 14 text: null startSeconds: 1107 endSeconds: 1186 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1107' heldReason: rights - ordinal: 15 text: null startSeconds: 1197 endSeconds: 1284 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1197' heldReason: rights - ordinal: 16 text: null startSeconds: 1284 endSeconds: 1295 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1284' heldReason: rights - ordinal: 17 text: null startSeconds: 1295 endSeconds: 1405 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1295' heldReason: rights - ordinal: 18 text: null startSeconds: 1416 endSeconds: 1494 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1416' heldReason: rights - ordinal: 19 text: null startSeconds: 1507 endSeconds: 1592 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1507' heldReason: rights - ordinal: 20 text: null startSeconds: 1592 endSeconds: 1602 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1592' heldReason: rights - ordinal: 21 text: null startSeconds: 1602 endSeconds: 1725 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1602' heldReason: rights - ordinal: 22 text: null startSeconds: 1725 endSeconds: 1735 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1725' heldReason: rights - ordinal: 23 text: null startSeconds: 1848 endSeconds: 1925 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1848' heldReason: rights - ordinal: 24 text: null startSeconds: 1936 endSeconds: 2057 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=1936' heldReason: rights - ordinal: 25 text: null startSeconds: 2057 endSeconds: 2068 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2057' heldReason: rights - ordinal: 26 text: null startSeconds: 2068 endSeconds: 2167 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2068' heldReason: rights - ordinal: 27 text: null startSeconds: 2178 endSeconds: 2315 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2178' heldReason: rights - ordinal: 28 text: null startSeconds: 2315 endSeconds: 2327 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2315' heldReason: rights - ordinal: 29 text: null startSeconds: 2327 endSeconds: 2473 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2327' heldReason: rights - ordinal: 30 text: null startSeconds: 2473 endSeconds: 2485 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2473' heldReason: rights - ordinal: 31 text: null startSeconds: 2543 endSeconds: 2633 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2543' heldReason: rights - ordinal: 32 text: null startSeconds: 2543 endSeconds: 2633 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2543' heldReason: rights - ordinal: 33 text: null startSeconds: 2644 endSeconds: 2765 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2644' heldReason: rights - ordinal: 34 text: null startSeconds: 2867 endSeconds: 2986 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=2867' heldReason: rights - ordinal: 35 text: null startSeconds: 3126 endSeconds: 3138 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=3126' heldReason: rights - ordinal: 36 text: null startSeconds: 3184 endSeconds: 3194 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=3184' heldReason: rights - ordinal: 37 text: null startSeconds: 3194 endSeconds: 3419 watch: 'https://liquidgold.cc/watch/KLDdXOw6jIc?t=3194' heldReason: rights --- # SOTA Generative Media Panel — Dumitru Erhan, Shane Gu & Nicole Brichtova, Google DeepMind The team took captions from real videos, regenerated the same scenes with their own model, and ran a human eval. People largely preferred the generated version. Dumitru Erhan is quick to deflate that result. The output is not more realistic, it is sharper and more saturated with nicer skin tone, and human preference turns out to be an unreliable thing to optimize against. Nicole Brichtova has a name for the effect, the Instagram filter. The same blind spot surfaces in smaller ways. Their image model quietly began putting wedding rings on hands, and nobody internally caught it until an outside tester asked why, which is reward hacking arriving through the back door. Much of the hour circles one question, whether language is a good enough intermediate representation for any of this. Shane Gu argues it thins out exactly where humans are most sensitive. We carry poor vocabulary for audio, taste, smell and skin tone because those sit close to survival, and a professional wine taster he asked had resorted to borrowing the language people use to describe a date. The tell of AI video, swyx notes, is that everything sounds studio recorded, because studio recordings are the training data, and a model with no representation of standing further away cannot get that sound right either. Evaluation stays stubbornly manual underneath all of it. Ten people in a room, two videos side by side, pick one. Speaker info: Dumitru Erhan: - https://x.com/doomie - https://www.linkedin.com/in/dumitruerhan/ Shane Gu: - https://x.com/shaneguML - https://www.linkedin.com/in/shixiang-shane-gu/ Nicole Brichtova: - https://x.com/nbrichtova - https://www.linkedin.com/in/nicolebrichtova/ Timestamps: 0:00 - Introductions, and why this session exists 2:00 - Nano Banana 2 Lite and the Omni Flash APIs 4:30 - Beyond demos: storyboards, editing, education 7:51 - Video agents, or one model doing it all 10:26 - Video models as zero shot learners 13:01 - Will everything collapse into one model? 14:20 - Is captioning the right intermediate representation? 18:13 - What people actually mean by world models 19:44 - Vision researchers becoming generation researchers 24:51 - Audio, and why joint generation mattered 26:32 - The things language cannot describe 32:03 - Studio quality as the tell of AI video 34:12 - When humans prefer the generated video 36:07 - Prompt engineering and trained sensitivity 38:27 - Setting a default aesthetic for the world 41:09 - The wedding ring nobody noticed 42:10 - How you actually evaluate video 47:35 - What data they are looking for 53:04 - Forward deployed engineers as an eval channel --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/KLDdXOw6jIc-youtube-b01c3c14dd8f8e90af4c Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/sota-generative-media-panel-dumitru-erhan-shane-gu-59dcef3eb6c3 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.