YouTuber recreates Google's faked Gemini demo with GPT-4V

Greg Technology's open-source ChatGPT Video interface connects your webcam and audio directly to OpenAI's GPT-4V, delivering real-time multimodal video interaction straight to your personal laptop.

The project was built directly in response to Google's controversial Gemini Ultra launch video. While Google's promotional clip showed an assistant seamlessly watching and responding to live video, it turned out to be staged using still photos, manual text prompts, and post-production voiceover. This open-source tool proves that the same core concept can already run genuinely in real time.

What it does

The tool continuously grabs video frames from a camera, sends them to GPT-4V alongside audio prompts, and reads out responses in a natural tone similar to ChatGPT Voice.

In demonstrations, the setup cleanly handles the same interactions Google faked. It observes hand gestures to play rock-paper-scissors, watches drawings unfold stroke by stroke to identify a duck floating on water, and solves visual puzzles. Tech reporters testing it on an M2 MacBook Air had it running within minutes, noting it accurately identified hand gestures, a glass coffee mug, and read both the title and author off a book held up to the camera.

How to get it running

The entire codebase is freely available on GitHub.

Running it does not require an expensive GPU cluster or heavy local machine learning pipelines. All you need is a basic computer, the repository code, and an active OpenAI API key with access to GPT-4V.

It is certainly less polished than a multi-million-dollar marketing production. Responses take a couple of seconds to process over the API rather than hitting zero latency. But unlike corporate teasers, every reaction happens live without post-production tricks or pre-fed still images.

Why it matters

For developers exploring camera-based AI agents, this project is a practical starting blueprint you can inspect, run, and modify today without waiting on enterprise waitlists.

We think this highlights an amusing contrast in modern AI development. While Google relied on clever video editing to make its model appear revolutionary, an independent builder recreated the actual live workflow using existing OpenAI infrastructure. It sets a healthy bar for the industry: real-time multimodal capability is already here, and consumers will demand the real thing over staged demos.