Karina Nguyen and Jakub Zegzulka built the research demo around synthetic versions of Elon Musk and Joe Rogan, with generated dialogue labeled on screen.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [X](https://x.com/karinanguyen/status/2102806675174687176)
Why it matters #
ACTx486 makes the appeal and risk of interactive synthetic media visible in the same demo: viewers can steer familiar people through new scenes, while the team acknowledges that responses take minutes and its safeguards remain incomplete.
Karina Nguyen (@karinanguyen) and Jakub Zegzulka (@jakubzegzulka) introduced ACTx486 on September 23rd, a research demo that lets viewers interrupt a podcast, ask questions and direct changes to the scene. The exchanges shown so far were generated in advance, and a complete turn takes minutes.
https://x.com/karinanguyen/status/2102806675174687176 The pair built the demo around an episode of The Joe Rogan Experience featuring Elon Musk. ACTx486 combines clips from the original recording with generated segments showing the two men speaking and responding. The project labels generated sections on screen and says the simulated faces, voices and words are synthetic. Neither man said the generated dialogue, and the project is not affiliated with the show or its hosts.
The demo relies on viewers recognizing familiar people in a conversation that never happened. Its creators say they chose the episode because its subjects are recognizable and the risks are easy to see: a system that puts new words in a trusted person's mouth could make falsehoods more convincing.
Viewers can ask the simulated Musk for Los Angeles recommendations tailored to a viewer profile, request an answer in Italian, ask to visualize life on Mars or direct a change to the scene. The site also says the system can generate diagrams, carry changes across turns and send useful information to a phone. These are descriptions of the team's demo, not evidence of a deployed service or measured user demand.
ACTx486 routes a request through research, response writing and scene generation, according to the project. The published examples were made in advance. The team says it hopes to bring an ideal response under six seconds, a target it has not reached. Today, the demo runs more like a production pipeline than a live conversation. A real-time version would also need to address the computation, review and delay involved.
Nguyen's published biography describes work on post-training, model behavior and early products at Anthropic, followed by AI research and product work at OpenAI. Zegzulka's portfolio documents interaction-design work on Meta's Orion augmented-reality glasses and Apple's emerging-technology projects. The demo reflects those backgrounds in its focus on directing a scene through conversation rather than putting a chat window beside a video.
The creators say they started with existing media rather than a blank prompt. The original recording provides the scene's pacing and context; viewers can ask questions or alter moments without inventing the whole experience. ACTx486 calls the viewer a "microdirector" and leaves open how much control viewers should have over a work shaped by its original creator.
The safeguards described on the site cover a narrower set of cases than the demo's capabilities. ACTx486 says the simulation avoids using later parts of the episode, completes research before responding and declines personal and political questions on the real person's behalf. The team also says it needs a broader trust-and-safety layer for media that speaks as someone else. In the examples, authentic footage sits alongside invented speech in a real person's voice and invented lines based on public information. On-screen labels mark where the footage changes; the project does not claim those labels resolve the risks of synthetic likenesses.
ACTx486's website describes the project as a research demo and offers an email request form. Its creators say they are publishing it to develop safeguards in public. For now, the project has shown pre-generated scenes, not a responsive medium. Making it responsive will require much faster generation and safeguards that hold up when viewers can steer a recognizable person's likeness in unpredictable directions.