GLITCH introduces speech-to-film workflow in an 8-minute Michael Shamberg demo

• Starks ARQ says it used a system called GLITCH to make an AI short film with Michael Shamberg, whose credits in the thread include Pulp Fiction, Django Unchained, Erin Brockovich, and Contagion, in under eight minutes from a hotel room with one laptop, according to starks_arq's demo post and starks_arq's Shamberg follow-up.
• GLITCH is framed by starks_arq's product description as a "speech to film system" where a user directs out loud by casting characters, building the world, blocking the scene, and cutting the take.
• starks_arq's workflow post says the system follows a director's sequence, cast, world, scene, shot, while starks_arq's orchestration post says the underlying shift is multiple models and agents working in sequence rather than one model producing one clip.
• The strongest creative claim came from starks_arq's takeaway post, which pitched the tool as a way to turn ideas that died in budget meetings into films by the end of the day.
• Access still looks limited. starks_arq's contact post points people to ARQ's site for "serious projects," not a public product page or signup flow.

You can watch the original hotel-room demo and a second clip that places Shamberg inside the generated film. The thread also sketches the interface in unusually plain language, via GLITCH's speech-to-film description, the cast-to-shot sequence, and the orchestration claim. The only outward-facing link is ARQ's site, attached to a project inquiry post rather than a launch page.

Starks ARQ's main proof point was social, not technical: it brought in Michael Shamberg and ran the pitch as a live creative reveal.

The thread says Shamberg found the team on X, they flew out, opened one laptop in a hotel room, and had a film in under eight minutes. For AI filmmakers, that is the hook: not raw model quality, but the claim that a veteran producer could move from spoken direction to a watchable short almost immediately.

A second post extends the stunt by inserting Shamberg into the generated short itself.

The system description is concrete enough to map the workflow as a list:

1. Cast the characters.
2. Build the world.
3. Block the scene.
4. Cut the take.
5. Refine from cast to world to scene to shot.

That sequence matters because it mirrors film language instead of prompt-box language. The pitch is that you direct out loud the way you would on a soundstage, then the system turns those spoken instructions into the film.

Starks ARQ says the important jump is not one better video model. It says the leap is orchestration, multiple models and multiple agents running in sequence like a film crew.

That is the most useful line in the whole thread because it shifts the frame from clip generation to production workflow. The company does not break down the stack here, but the claim implies GLITCH is a coordinating layer that sequences specialized steps rather than a single end-to-end generator.

A background post adds that the company spent the last year recycling revenue into credits, generations, and workflows after starting an AI production company in El Salvador.

There is no public GLITCH product page in the evidence set. The call to action is a project inquiry.

ARQ's site is the only external link in the thread, and starks_arq's outreach post also offers to build opening scenes for selected directors, writers, or producers replying on X. That makes this look more like a studio capability being shown in public than a self-serve creative tool that creators can sign up for today.