AI Video’s New Bottlenecks: Characters, Control, and Distribution

From weekend tutorials to shared character universes

Friends in creative fields who had dismissed AI images and videos as unusable were coming over for weekend tutorials. One guest describes that change early in 2025: people now wanted to know which tools to use and how to work with them. In Latent Space’s conversation with Olivia and Justine Moore, the hosts had a practical problem of their own: how to build an experimental secondary channel around AI video.

Italian Brain Rot shows how these tools can produce a shared entertainment universe. An initial creator introduced characters as images; other TikTok users added their own, and popular additions became informal canon. Animation enabled compilations, interactions, and continuing storylines. Remixing expanded both the cast and the narrative, without a single studio directing the universe.

A guest reports seeing toys, T-shirts, and plushies, and a child who recognized the characters as readily as television characters. Frequent encounters may help explain that attachment: a feed can deliver many character videos daily, while a television series might release an episode weekly. This is a proposed explanation, not a measured effect. The audience also extends beyond children: adults were making musicals, movies, and relationship dramas with the same characters.

The tools were also changing where trends began. A guest describes early AI video traveling from specialist Reddit communities to Twitter, only occasionally reaching TikTok or Instagram. By this conversation, more viral material was originating directly on consumer platforms, where nontechnical users could create and remix it. Repeated remixes and dedicated accounts became useful discovery signals: people were developing a format beyond a single striking generation.

One prompt can make a clip; control still takes work

Olivia Moore started an account after Veo 3 arrived to test how easy monetization might be. Her original sensory-video ideas, including opening dragon eggs, worked less well than following the existing fruit-slicing trend. She later tried gold-bar squishing and simulated Tide Pod consumption, cross-publishing to YouTube and Instagram. She found that different material worked on different platforms, without reporting comparative metrics.

Almost all her videos came directly from text prompts in Veo 3. The companion article highlights native audio generation—producing sound alongside the picture—as an advance that can remove separate lip-sync and sound-effects work. Olivia’s exception shows why those stages sometimes remain separate: for Star Wars characters performing Olympic dives, she used MiniMax for video and ElevenLabs for sound effects. Manually generating and aligning the effects gave her more control but took a long time.

At the time, starting from an image in Google’s interface switched generation to Veo 2. Creators could not combine a reference image with Veo 3’s audio in that workflow, making recurring characters harder to reproduce. Familiar figures such as Stormtroopers offered a workaround: the model could recreate their recognizable identities from text without a starting portrait.

Olivia reports generating cartoon characters such as Stitch while encountering restrictions on real people—an observation about tool behavior, not permission to use copyrighted characters. Ease also depended on the creator: one participant found simple sensory clips easy but narrative vlogs difficult; another, generating dozens of videos daily, found vlogs fairly straightforward.

Character design becomes a consistency workaround

The discussion proposes two ways characters attract attention: - Familiarity with a twist: a Stormtrooper supplies an immediate association, while unexpected behavior gives viewers a reason to continue. A participant reports testing formats, but supplies no test design or measured lift. - Strangeness that invites another look: Italian Brain Rot’s unfamiliar creatures and confusing language encourage viewers to work out what they are seeing. This is the guest’s explanation for how original characters can compete with established ones.

Kim the Gorilla adds a sustained story: Kim fights with a zookeeper named Becky and tries to escape the zoo. A participant reports roughly 300,000 followers, videos with hundreds of thousands of likes, and “Free Kim” merchandise, without specifying the growth period. The continuing conflict is the creative contribution. Access to Veo 3 alone does not supply an engaging premise or the judgment to sustain it.

Kim also illustrates recognizability without exact visual consistency. A participant suggests viewers notice differences between gorilla faces less readily, while a pink bow supplies an identity cue even when its placement varies. Keeping a character recognizable across several generations lets creators assemble stories beyond the eight-second clip limit described in the discussion. The workaround spreads through remixing: a Yeti vlog inspires a Stormtrooper version, then a gorilla, then a particular gorilla with a setting and an antagonist.

Virtual characters separate creative personality from on-screen appearance. One guest argues that funny, imaginative people can build an audience without personally fitting conventional influencer beauty standards. Another participant offers a deliberately cynical commercial interpretation: the creator controls the virtual influencer’s appearances, situations, and public messaging.

Retries determine the cost of a publishable video

The monetization routes discussed depend on different assets: - Platform payouts: earn from eligible views and engagement. - Traffic to a business: use videos as advertisements or bring viewers to an existing product. - Services and education: use creative work to attract consulting clients or course buyers; Nick St. Pierre is cited as an example. - Characters and merchandise: sell products built around audience attachment. For collectively developed characters, a hypothetical licensing deal raises an unresolved question: how should contributors share the revenue?

One participant needed about eight generations for a satisfactory glass-fruit cutting clip. Rejected attempts cut horizontally instead of vertically or revealed implausible internal layers. Once subscription credits ran out, retries required additional spending. The relevant production cost was therefore the cost of a publishable result, including failed attempts.

Revenue had two hurdles as well: qualify for a creator program, then earn enough from subsequent videos to justify production. A participant reports earning nothing from TikTok because they were not enrolled. No reliable payout rate emerges from the discussion; the estimates involve disagreement and confused denominators. A deliberate monetization route matters when every retry costs money.

Audience growth can still be rewarding without direct earnings. Another participant argues that people who previously struggled to grow an account now find substantial likes and followers motivating in themselves.

Applications compete on access and control

The model layer develops the generators; the application layer makes those models easier to use. A host proposes that proprietary model providers might capture most of the money, but the response only partly agrees. Even access to the same model can be valuable when an intermediary removes interface friction.

The examples expose different kinds of friction: - Kling combines common production steps: upload an image, generate video, and add sound effects in one interface. - Flow made model access difficult: participants describe subscription and Google-account confusion, Veo 2 as the default, hard-to-find model switching, and an unusable mobile experience at the time. - Krea and Replicate reduce commitment: creators reportedly used intermediaries for easier access and one-off, pay-as-you-go generation.

The categories overlap. HeyGen develops a talking-avatar model while also hosting other image and video models. Adobe prompts a disagreement: one participant speculates that weak adoption drove Firefly toward third-party models; another reports surprisingly strong Adobe usage in a State of AI Engineering survey, without supplying methodology or percentages. A participant’s own use offers a more precise distinction: editing utility can differ from generation appeal. They use Adobe’s generative fill for editing images rather than primarily making new ones.

ComfyUI is an open-source system of interconnected, dependent nodes for assembling complex workflows. Its value depends on the control a task requires: - Casual memes and short videos: a direct model interface may be sufficient. - Style transformation, consistency, and upscaling: configurable workflows provide controls missing from the model interfaces discussed, including transforming realistic footage into anime. - Character swapping: open-source tools offer advanced capabilities that remain difficult for ordinary users to access.

An educational avatar still needs an explanation

For Latent Space’s proposed educational channel, a guest suggests a format already appearing on Instagram: a synthetic celebrity delivers a script while instructional diagrams appear above the talking image. They report seeing such videos receive millions of views, though they do not know which tool produced the celebrity likeness. The presenter and the instructional visuals are separate production tasks.

The proposed workflow is: 1. Draft from existing material. Give YouTube links or podcast transcripts to Gemini 2.5 and request an entertaining script about a selected topic. 2. Edit the adaptation. Refine the script manually before production. 3. Animate a presenter. Latent Space already uses an ElevenLabs voice called Charlie; the suggestion is to give Hedra an audio clip and an image to make a talking character. 4. Assemble the instructional visuals. Create diagrams or reuse existing educational material separately from the avatar. This was a proposal, not a completed Latent Space experiment.

Presenter choice depends on the intended audience. A guest suggests existing listeners might prefer the actual hosts, while a celebrity likeness might attract unfamiliar viewers; no outcome comparison is reported. Accompanying footage can be another automated step: B-roll generators extract keywords from narration, then retrieve or generate images and video to accompany it.

Clipping agents need an editorial objective

Overlap offers a different route: monitor a YouTube channel, select excerpts from new uploads, and publish them across social platforms. The demonstrated agent watches the a16z channel and selects 30–140-second clips for a technology audience. Additional actions include subtitles, converting horizontal 16:9 footage to vertical 9:16, and smart zoom on a speaker’s face.

Editors can change clip boundaries, remove caption words or cut words from the video, remove filler and stutters, and restyle subtitles. Overlap can also write accompanying posts and publish automatically. Its predicted virality score is a product feature, not a validated performance measure in this discussion.

The hosts already have an internal tool that selects timestamps. Their harder problem is making those selections into videos people watch—and ensuring an automated selector values the passages a technical editor would choose. A guest proposes testing different selection instructions for different audiences: - YouTube: educational, informative, technical passages. - TikTok: funny or controversial passages aimed at younger viewers, reformatted for the platform. No results from that comparison are reported.

Whether excerpts can compete with videos conceived for short-form viewing remains contested. One host suspects repurposed interviews generally perform worse and says the team has considered changing how it records. A guest counters with observations of interview clips repeatedly going viral on X, especially when they feature recognizable people. These are competing judgments rather than measured comparisons.

Reach must preserve the speaker’s meaning. Participants describe resisting sensational framing that turns qualified remarks into exaggerated claims, including sweeping AI promotional language. An affiliated sub-brand is floated for experimenting with presentation styles, alongside the argument that wider reach requires adapting to how audiences receive information.

Attention and learning are separate outcomes. One participant worries that changing viewing habits could crowd out long-form intellectual writing. Another expects good writing and discussion to retain an audience, suggesting that some educational-character videos spread because their premise is funny rather than because viewers learn from them.

Prompt Theory turns generation into the plot

Prompt Theory makes the production mechanism part of the fiction: generated characters discover—or refuse to accept—that prompts control them. A participant compares it with a NotebookLM piece in which a generated host dramatizes discovering that he is not real and has no wife to call. These are fictional scenarios about self-awareness, not demonstrations of it.

The premise expands into speculation about humans being prompted characters in someone else’s universe. It also becomes ordinary comedy: generated young and old characters trade insults, including accusations that attractive hair was merely prompted. A guest finds realistic-looking characters especially effective at making that abstract premise striking.

That leads to an unresolved question about companionship. Using Reddit’s pseudonymous conversations as a starting point, a participant asks whether always-available AI interlocutors with interesting things to say would necessarily be disappointing. The question concerns what people value in interaction; it does not establish how many Reddit accounts are automated.

From audience attachment to physical products

The closing examples return to physical production. Viewers asked for sweatshirts featuring an AI possum character, and a participant made roughly 30: remove the background from a character design, place it on a sweatshirt, and have a screen-printing company produce it. The shirts went to friends, family, and the AI video community. A commercial brand remained a possibility.

AI furniture and home-design accounts offer a related path from imagined object to physical demand. A guest describes interest in designs such as a giant cat-shaped couch with glowing eyes, and reports seeing an AI gorilla-chair concept subsequently manufactured and made available to buy.

The practical discovery advice is to observe platforms separately: create accounts on YouTube, Instagram, TikTok, and X, follow AI creators, and let each feed surface its own community. Participants describe sensory AI videos spreading from TikTok and animal-diving videos gaining traction on Threads and Facebook products before moving elsewhere. One estimates a one- or two-day opportunity to carry a trend between platforms—an observed pattern, not a measured advantage.

For professional production, the recommendation is PJ Ace’s commercial work and shared workflows, including projects for musicians and brands. These provide a different view of AI video from casual meme creation.

The final startup proposal extends automated clipping into merchandise: use a creator’s latest videos, quotes, and images to generate products. The hypothesis is that good merchandise could become a monetization bottleneck as content production and audience growth become easier. That handoff—from generated media to something an audience wants to buy—remains open for someone to build.