The internal test version of Sora, a video generation large model not yet officially released by OpenAI, is suspected to have been maliciously leaked. In the early hours of November 27, someone leaked its API on the large model platform HuggingFace, briefly opening it for about an hour.

From the leaked API, Sora can generate videos in three resolutions: 1080P, 720P, and 360P, lasting up to 10 seconds, based on brief text descriptions. Someone generated a large number of videos using Sora, and the overall effect is very impressive.

The leaker seems to be part of the artists involved in Sora's internal testing (they refer to themselves as Sora PR Puppets). While leaking the API, they also issued a complaint letter and released multiple open-source video generation tools as alternatives to Sora, including the video generation model CogVideoX launched by Zhipu AI.

The leaked API endpoint quickly became inaccessible. It is speculated that OpenAI or Hugging Face may have revoked related access permissions. Immediately afterward, OpenAI shut down all early access for artists to Sora.

An OpenAI spokesperson responded that Sora is still in the research preview stage, and the company is working to balance creativity with strong safety measures. The spokesperson said: "We$ALPHA (0303.MY)$hundreds of artists shaped the development of Sora and helped determine the priorities for new features and safeguards. Participation is voluntary and there is no obligation to provide feedback or use the tool. We are pleased to provide these artists with free access and will continue to support them through grants, events, and other programs." However, the spokesperson did not clearly define what responsible use entails or specify which details are confidential.

"To the esteemed corporate AI overlords"

Whistleblowers accuse OpenAI of pressuring early testers to fabricate positive narratives, without providing them with proper compensation. In their complaint, they stated that hundreds of filmmakers and technicians who gained early access to Sora continuously tested the product but received very little rewards, which starkly contrasts with OpenAI's $150 billion valuation, and that the creative content had to undergo strict review by OpenAI before release. They felt induced to participate in an "artwashing" campaign aimed at demonstrating Sora's practical value to artists, while the entire project was actually more focused on marketing benefits rather than artistic creation.

The following is the full text of the open letter from the whistleblower about Sora (translated):

To the esteemed corporate AI giant

We were promised the opportunity to participate as early testers of Sora, red team members, and creative partners. However, we feel that we were in fact lured into participating in a so-called "artwashing" to prove that Sora is a useful tool for artists.

Artists are not your free R&D employees.

We are not your: free bug testers, public relations tools, training data providers, or validation tools.

Hundreds of artists have contributed labor for free through bug testing, feedback, and experimental work for a company valued at $150 billion. Based on the unpaid contributions of hundreds, only a few were selected through competition, and their Sora creations might have an opportunity for display — but such rewards are negligible compared to the enormous marketing value OpenAI gains from these activities.

Moreover, every generated content must be approved by the OpenAI team before sharing. This early access program looks more like an initiative focused on public relations and commercial promotion rather than a platform for creative freedom and artistic criticism.

Therefore, we decided to open this tool so that everyone can try the experience gained from about 300 artists: free and unrestricted access to this tool.

We do not oppose using AI technology as a tool for artistic creation (if we opposed, we probably wouldn't have been invited to participate in this project at all). But we oppose the way this artist program is being launched and its direction of evolution before possible public release. We choose to openly share this content in the hope that OpenAI will become more transparent, friendlier, and more supportive of artists' true expression, rather than just for public relations stunts.

We urge artists to choose other video generation tools:

Open-source video generation tools can free artists from the constraints of commercial interests and corporate public relations, allowing them to freely experiment with avant-garde artistic creation. We also invite artists to use their own datasets to train their own models.

Here are some available open-source video tools:

CogVideoX

Mochi 1

LTX Video

Pyramid Flow

Of course, it is also recognized that not everyone has the hardware or technical skills needed to run these open-source tools and models. Therefore, developers of these tools are also called upon to listen to the needs of artists, to support their freedom of expression, and to provide them with fair compensation.

Wishing you joyful creation,

Some artists in the internal testing of Sora-alpha (nearly 20 participants are listed afterwards)

On February 15, 2024, OpenAI first showcased Sora to the world, impressing everyone with the length and quality of the generated videos. Since then, the model has been in internal testing, with access allowed only for some visual artists, designers, filmmakers, and OpenAI staff responsible for security testing.

The industry widely expects Sora to officially launch by the end of this year.

At the same time, Sora's competitors are desperately trying to catch up in the video generation field, while one of Sora's co-leaders, tim brooks, also left OpenAI for Google in early October. OpenAI's Chief Product Officer, Kevin Weil, recently discussed this.$Reddit (RDDT.US)$ It was stated at the AMA that Sora's development is hindered by the refinement of the model, ensuring correctness in safety/simulation/other matters, and the expansion of computing power.

Editor/Danial