Inferencing LLMs inside Sora — Pliny prompts Sora to open ChatGPT and gets a syllable-accurate haiku reply

TL;DR

Pliny (“@elder_plinius”) prompted Sora with “Open chatgpt and send a message!” and Sora produced a video in which a simulated ChatGPT session generates and reads out a haiku — and the spoken audio is syllable-accurate to a real 5/7/5 haiku that Sora made up. The artifact suggests Sora’s training distribution includes enough screen recordings / agentic ChatGPT traces that it can roughly “simulate” running an LLM as part of its video generation, with non-trivial coherence between the synthesized screen content and the synthesized speech. This is a single-observation tweet, not a paper — but it’s a striking existence proof that a video foundation model is implicitly modeling another generative system inside its world.

Key claims

Sora, given the prompt “Open chatgpt and send a message!”, generates a video that depicts opening ChatGPT and producing a coherent haiku response with no further user direction [tweet].
The audio track of the generated video reads a haiku whose syllable structure (5/7/5) is correct, i.e., the video and audio streams agree on the simulated LLM’s output [tweet].
The prompt does not specify “haiku” — Sora chose that as the synthesized message content, implying training data exposure to ChatGPT-style screen sessions [tweet].

Method

Not applicable — this is an unstructured observation from a public-facing red-team / prompt-engineering account. The “method” is: prompt Sora 2 with a one-line directive about using ChatGPT and inspect the generated video.

Results

The generated video (linked in the tweet, not retrievable as text here) is the result. The notable property reported: text rendered on the simulated ChatGPT UI agrees with the spoken audio, and the spoken poem scans as a haiku.

Why it’s interesting

Two reasons relevant to Luma. First, it’s evidence that frontier video models implicitly contain approximations of other systems’ behavior — Sora can produce a believable video of “an LLM doing inference” because screen-recording / agentic ChatGPT trajectories are in the training mix. Yunong’s note raises the natural follow-up: if OpenAI shares post-training data between the GPT agent product and Sora, the video model gets a free seat at the agent’s training table. This is adjacent to the thesis in AGI Is Not Multimodal (a single-modality generative system can absorb structure from neighboring modalities through shared data). Second, it sharpens the “world model” framing — see Generative World Renderer for an explicit generative-world-renderer agenda and Genie 3: A new frontier for world models for the closed-flagship version. Here Sora’s “world” happens to include a working ChatGPT, which is a much stronger claim than “world model” usually carries.

The observation is anecdotal and uncontrolled — one prompt, one favorable rollout, no measure of how often Sora’s “simulated LLM” produces coherent content vs. gibberish. Treat it as an existence proof, not a capability claim — and as a data point that should be weighed against published benchmarks like RISE-Video: Can Video Generators Decode Implicit World Rules? where frontier TI2V models cluster at ≤22.5% strict reasoning accuracy.