Agents that finish the ad, models that generate the world: two bets on AI video, 1.5 km apart

On the evening of 15 October two rooms in Venice, Los Angeles, 1.5 km apart, showed opposite ends of AI video. At Vizard × Alibaba Cloud an agent took one line of brief and one photo of a perfume bottle and returned a finished 30-second ad, and nobody opened an editor. At Visko × Google Cloud a model generated a snowy forest live, moved closer to a fox when asked out loud, and lit the fireworks the room had voted for on their phones. Both rooms said the same thing in different words: the hard part is now the idea and the story, not the footage. I was there with a camera as an observer. I build project-life2film, an AI video maker, so my takeaways from the drive home close the post.

The short version

• Agents now make the edit decisions. Vizard Agent, “Claude Code or Cursor, but for video” in the Vizard speaker’s words, wrote the script, made one anchor shot and checked it, generated four more in parallel, cut the timeline and reviewed its own render before handing over a 30-second ad.
• Live models generate faster than you watch. Visko’s Orbis generated two minutes of video in “one minute and 0.8 seconds” on four TPU v7x chips (Ivan Nardini, Google Cloud). Visko gave about $10 an hour as a near-term price, not today’s. The model is closed and does not run on-prem.
• Neither room thinks AI brings the story. “It’s not the wand. It’s the magician”, said director Jeffrey DeChausse; Buzz Hays of Google Cloud called storytelling a learned skill that AI does not give you.
• The agents evening ended on distribution, not production. A UGC talk claimed 1.4 billion views from 308,000 creator videos for under $3M, against $7-14 million as paid ads, and told founders that any app can be copied, so distribution is the moat.
• My takeaway: AI video makers are a crowded market, and story and editing are still the bottleneck. That is the part Life2Film works on, and creators are the channel I should use for it.

The whole evening in 22 minutes, with English and Russian subtitles. Jump to: Vizard × Alibaba · Google × Visko · my takeaways in the car.

Two bets, side by side

Vizard × Alibaba Cloud: an agent makes the edit decisions

“The Age of Video Agents”, at moss venice on Electric Ave, from 17:00. Four talks: Vizard on the agent, Alibaba Cloud on the model, a director on craft, a growth speaker on UGC. Times below are in the combined film.

The agent: five decisions a team used to make

The Vizard speaker opened with a chart: the video every channel wants (TikTok, Reels, Shorts, the Amazon listing) grows faster than what a team can produce, and most marketing teams live in the gap (1:21). AI video did not close it, because “what AI gave us is clips” (2:00): you still regenerate until the product looks right, assemble in Premiere or After Effects and resize for every channel. The hard part is “hundreds of decisions between an idea and a finished video”, and agents change who makes them (2:20).

The demo brief was one line, “create a 30-second product ad for this perfume for my Amazon listing”, plus one photo of the bottle (2:57). The agent broke it into five decisions (3:34):

1. What is the story? It reads the photo and writes the script.
2. Is this really our product? It makes one anchor shot and checks it, and every other shot builds on the same bottle.
3. Which shots? Four, generated at once, always with the same model.
4. How does it cut? It builds the timeline and renders.
5. Is it good enough? It reviews its own render for motion problems before delivering.

“These five decisions are what would take a team days” (4:10). A product photo is optional: a brand film came from a short brief with no footage, and a script plus one style image held a pixel-art look to the end (4:14). The one idea the speaker asked the room to keep: once an agent makes those decisions, “the bottleneck moves from production to ideas”. The question stops being whether you can afford the video and becomes which ideas to test, so the advantage is the brief, your taste and how fast you test (4:32). Three prompts to try, from the talk (not in the film): a screenshot of your product page into a TikTok ad, a website or reference video into a 30-second launch promo, a podcast or webinar into social clips.

The model: Wan 3.0 and OmniReference

Alibaba Cloud’s speaker, who does go-to-market for Qwen and Wan in the US and Europe, presented Wan 3.0 as built around five capabilities: native 30-second video, OmniReference, “reality-grade” rendering, pixel-level consistency and audio-visual impact. On stage the speaker said the model powers tools such as Vizard Agent. Most models give 5 to 15 seconds, “and for brand and marketing, that’s just not a story”; 30 seconds gets a start, a middle and an end (4:55). OmniReference takes a product shoot, a brand ambassador or a design and holds them across the video, “no drift, no hallucination” (5:22). That is the company’s claim. The examples come with audio, the speaker said, but the sound failed in the room. The models are on Alibaba Cloud Model Studio.

The director: five seconds out of ten

Jeffrey DeChausse, introduced as an Emmy Award-winning filmmaker who started in documentary and moved to commercials, gave the craft view. AI is a grocery store full of ingredients; someone still has to cook the meal (5:47). It has “no sensitivity or humanity”, so you work with what it gives you the way a jazz player improvises (6:00), and “it’s not the wand. It’s the magician” (6:27). He expects three production models side by side: traditional, hybrid and full AI (6:34).

His working rule: the model screws up, but maybe five seconds of a ten-second prompt work, or three, and pulling them out and slotting them in is the artist’s choice (6:50). Keep the final choice, and feed the machine original pictures and video from outside the digital space, or “it’s going to eat itself” (7:20). I asked him how he would approach editing real footage, which is what I build; his answer was that pre-production is everything, so organise your folders, your vision and your script first, or you will not understand the arc of your concept (7:45). Asked from the floor about editing, he said “I don’t trust editing yet to AI” (Q&A, not in the film). He closed with “you’re paid for your feelings” (8:23).

The growth talk: distribution is the moat

The last speakers came from the UGC side and said they work for Masterworks and other companies. Every number here is theirs, from their slides, and I have not checked any of it.

• Who does what. Founders who have not raised money should make founder-led content; once you have raised, UGC and paid ads (8:27).
• The headline case. For one client since March: 308,000 videos and 1.4 billion views, for less than $3 million, against $7-14 million had it been paid ads (8:46).
• Creators. Brutally consistent, posting at least twice a day (9:14). Hire them on retainer, $1,000 to $2,000 a month at minimum, because creators want fixed long-term work (11:29). In the talk, not the film: each creator opens new Instagram, TikTok and Facebook accounts that post only about your app, and the speaker had run over 12,000 creator interviews in seven years.
• Short form everywhere. Prime Video, HBO and Disney will each have a short-form feed, the speaker said, so “I’m more bullish on short form than anyone” (9:32).
• The moat. Most apps can be duplicated with Claude Code: Slack billed the speaker $10,000 for one month, so the team built its own. “You can copy any software”, and the hard part is distribution and users (9:52).
• Human or AI ads. Both perform; at Masterworks they produce 400,000 videos a month of both kinds (10:24).
• The hook is the first frame. Gen Z gives it 0.1 seconds, millennials maybe 0.5, boomers 1 to 2 (10:40).
• Series content builds a following: one series did over 25 million views and took an account from 6,000 to over 100,000 followers in a month (11:01).
• Tools. Vizard for paid ads, plus HeyGen, Higgsfield, Freepik and ElevenLabs, with the final cut still in Premiere and Final Cut Pro (11:13).

Visko × Google Cloud: a world generated while you watch

Live Models Night, at Google’s campus on Main St in Venice, 18:30 to 21:00. On stage: Ivan Nardini (Google Cloud, who ran the demo), Will (Qing) Yin (founder and CEO of Visko), Buzz Hays (global lead for media and entertainment at Google Cloud) and Dave Elliott (head of AI developer advocacy at Google Cloud, moderating). On 1 September SiliconANGLE reported that Visko, after raising $10M, debuted Orbis, its first live model. Times below are in the full panel.

The demo: directing a scene that does not exist yet

The idea: a director has a script, but a script is text and gives no picture. Live Director, an alpha Google built with Visko, loads a script and generates each scene while you watch, and you change it as you go (0:18). Act one, a fox in a snowy forest (2:04). “Go closer to the fox”: Gemini’s Live API turns the spoken instruction into something the video model understands, and the camera moves in (2:32). Then a wide view of the forest, then a sunset sky (3:28). For the last scene the audience scanned a QR code and voted; big fireworks won (3:54). “This was a real-time interaction. It wasn’t batch” (4:57).

Ivan’s point about speed: there was no loading bar because the model generates video faster than it plays, two minutes of video “in one minute and 0.8 seconds” (27:03). It runs on four chips (29:04), of a TPU v7x (43:48), and Visko got it running on TPUs in less than a month. The moderator called that part the Google TPU sales pitch.

What a live model is, according to Visko

• Physics as intuition. Will Yin spent about 11 years on 3D physical simulation before Visko. Equations always leave a sim-to-real gap, so Visko learns “low-fidelity physics”, human intuition, directly from video (11:06). Drop a stone and a rubber duck into a bowl of water: the stone sinks, the duck floats, and that comes from the data, not from rules, because “you can’t really tell the model every single one of them” (48:39).
• Low fidelity is a limit, not a trade-off. High-fidelity physics would need a language model’s help, not the video model alone (1:03:03). Surgery simulation is out; visualisation is fine (45:37).
• Content made for each viewer. Search moved to passive recommendation; next, content “will not be pre-generated”, but made on the fly for one person (18:01).
• Roadmap, as claimed. Memory is already built in (35:15); the goal for next year is running more than an hour without stopping (36:10); dialogue and voice come in the next version (37:35); a style reference is not supported yet (48:06).
• Controllability by design. Two people with the same prompt get a very similar world, and that is the goal: creators need “full control of the pixels in front of them” (57:27).
• What it cannot know. Asked for a flying car shown hours earlier, it will imagine one from the two words, not draw that car (1:04:25).
• Data and safety. Training data is licensed (42:00) and user input is not used for training at all (43:02). Gemini sits in the middle of the pipeline and refuses unsafe requests (55:33), and uploading a photo of a person requires verifying that it is you (56:47).
• Cost and access. The API goes through a third party and Google TPUs. One hour of play “probably costs you around $10 or even less”; asked whether that is now, the answer was “in the near term” (53:38). From the Google side: six months ago this was not possible, and unit costs are still high but falling fast (53:16). Not open source, no on-prem (54:04).

Google’s media view: previs, trailers, the craft

Buzz Hays, a Hollywood producer for over 30 years, said AI has not been an artist-friendly medium: “How many of you love APIs?” (14:18). Live action means shooting the same scene over and over from different setups, and live models keep the creative spark alive where a deadline would kill it (15:15). In the near term he sees previs: it is hard to sell a project now without something that looks like a trailer for a film you have not made, because studio executives have to see it to imagine it (22:31). Dave Elliott compared it with an all-day storytelling session at a Hollywood event a year earlier, built on Kling, Runway and Veo and spent fighting coherence; what the room had just seen “could have been done in 30 minutes” (24:08). Dave also repeated the usual line that the demo is the worst model you will ever see (16:52).

Then the line that became a short. Buzz tells the business side that text prompts should be “the last resort, not the first”, and that people wrongly believe AI makes you a good storyteller: “I’ve flown on airplanes hundreds of times, but I certainly couldn’t fly one myself” (25:15). Dave closed the panel with the story of Uncle Joe, a logger who left school in fourth grade and was the best storyteller in the family; when Joe died, the stories were gone, and a tool like this could have kept them (31:03). Buzz also told the story of The Wizard of Oz at Sphere, a 1939 film rebuilt for a screen about 500 feet wide and nearly 400 feet high, done when the team had no commercial models to use (6:59), and said the hype cycle is behind us and enterprises are now seriously trying these tools, much of it agentic (29:42).

Use cases named on stage

From Visko’s hackathon and customers: walking a street at night you have never been to, generated along a Google Maps route (1:06:10); rehabilitation, where a watch reads heart rate and the model turns anxiety into soothing content (1:06:50); a world view of a construction site; test environments for robotics (1:08:42); education, where Visko said it already has customers (50:58). Buzz added previs for anything expensive to build: a floor plan projected at full scale in an empty room, new cabins for cruise ships (1:07:40). One hackathon project described on stage took the last scene of a sports video and extended it into an ad, zooming in on the player’s shoes, so the viewer felt they were still watching the game (18:53). The panel ended with a call for an LA Visko × Google hackathon, and hands went up (1:09:15).

Where the two rooms split: one film for everyone, or one for each viewer

Vizard’s agent makes many variants of one ad for many channels. Visko’s pitch goes further, to content generated for each viewer. The Google panel was not sure people want that. Asked from the floor whether personal content isolates people, Buzz compared making things only for yourself to journaling and doubted there is much of an audience for it; the youngest filmgoers are going back to cinemas for a communal experience, “the complete opposite of doom scrolling” (59:56). Dave added that hyper-personalised entertainment was once everyone’s bet, yet people go to games more than ever (1:01:02). The exchange ended on “both formats will grow” (1:01:35).

What I said in the car

My words on the drive home, from the film’s English subtitles (I spoke Russian):

• The Alibaba evening showed an editor similar to SuperDuper AI, building a film from an idea on top of Wan 3.0. “What I was thinking of building two years ago, now just about everyone is building similar products. The market is overfilled” (16:28).
• The problem is the same one I talked about the day before, at the time of neu-wave-ai-film-festival-2026: consistency, and the author’s ideas come first; they said again that these systems do not replace the author (16:57). “But maybe that’s a bit disingenuous, because automated advertising doesn’t really need an author, it needs variety” (17:06). In the raw clip I add that it still needs control, an orchestrator.
• The agency case was the useful one: the main channel now is creators. Take a pool of a thousand small influencers, give them tasks, watch which videos take off. “I’ve wanted to do this for a long time, and now I’m more and more convinced that I need to take it on seriously”, including for Life2Film (18:02).
• Live models are the next level: inside the model you walk around the Statue of Liberty and down any street you like. Very hard and slow for now, but in 2 to 3 years reality may be generated on the fly and personalised (19:00). The catch is the dataset: famous places work, a place that was not in the training data is harder, so the first popular cases will be films the models were trained on (19:55). Most likely not 3D worlds like the ones Meta made, but generative worlds, seen through glasses, where we look for information and learn (20:37).
• Generated video or video shot on your iPhone, “the main problem right now is still building the story and editing”. What I did with generative AI and what I do with Life2Film both come down to auto-editing and building the story, which needs close work with an author who wants to tell their story; in my case, my own life turning into a film (21:07).

What a solo builder of video tools can act on

1. Copy the loop, not the generator. Vizard’s agent checks one anchor shot before it fans out, and it reviews its own render before anyone sees it. Both steps transfer to an editor of real footage: settle what must stay fixed, then cut, then run a critic over the result. My Chicago films already work this way, with a critic pass run at least twice (creative-video-agent-cut).
2. Leave the five seconds to the person, or learn how they pick them. Jeffrey keeps five seconds of ten, calls that choice the artist’s and does not trust editing to AI yet. A montage tool automates exactly that choice, so it has to learn a person’s taste rather than override it. That was the bet behind neon-personal-agents-life2film-director-2026.
3. Spend on distribution before features. The UGC numbers are one agency’s claims, but the playbook is concrete: creators on a $1,000-2,000 monthly retainer, posting twice a day, a hook that lands in the first frame. On my side of shipping, the hook and the cover are still picked by eye (creative-video-shipping).
4. Watch live models, do not build on them yet. My reading: a closed, cloud-only model at about $10 an hour, as a near-term target, does not fit a free app like project-life2film that runs its AI on the device. The claim to watch is the hour-long run with memory, because that is what would make a live model useful for long stories rather than single shots.

How this was put together

I filmed both evenings on a DJI camera, 30 clips over the day, and transcribed them on my laptop. The combined film and the full panel carry cleaned-up English and Russian subtitles, and every time in this post is a time in those videos. The film was cut automatically by Life2Film. The five shorts were cut by a script from three of the day’s clips: the director’s talk, the UGC talk and the Google panel. Quotes come from the corrected subtitles; the few points marked “not in the film” come from the raw transcripts of the full talks. People on stage are named when they were introduced by name and role or are listed in the video descriptions. I have left out the Alibaba Cloud and UGC speakers’ names, which reached me only through speech recognition, along with audience members and private conversations. Every number above is the speaker’s own claim; I did not check any of it.

Related: neu-wave-ai-film-festival-2026 (the evening before, 14 Oct, also LA Tech Week), neon-personal-agents-life2film-director-2026 (the SF Tech Week kickoff), creative-video-with-claude (how I cut films with an AI assistant), filmcraft-agent-built-editor (an AI-built editor, and what it means for Life2Film).