OpenAI’s Sora 2 is putting safety and censorship to the test

Reinforced with the new artificial intelligence model Sora 2, Openai’s Tiktok -style video application is the world’s most valuable private company, which is the world’s most valuable private company.

Despite the door -to -up version requiring an invitation code, the video creation tool has already been drawn to three number. AppleCEO Sam Altman’s viral clip, including a Deepfakes wave launched.

Internally, the presentation directed a long -standing debate on how security will be balanced with creative freedom in Openai.

A person who is familiar with the internal strategy in the company said that he considers the strict railings of leadership necessary, but that he was worried about being perceived as strangling creativity or being perceived as too much censorship.

This voltage was not solved.

Openai’s culture has been preferred for a long time, usually sending new vehicles in front of competitors and allowing the public to adapt in real time.

A former employee, who wanted to be named to discuss internal issues, said CNBC has a model of Openai leadership during his term of office. This strategy was fully exhibited after publishing a strong model of China’s construction of Deepseek’s deepseek at the end of last year than anything else than anything else than the Silicon Valley.

OpenAi answered within weeks and made two new outputs models What is seen as a defensive movement to maintain its leadership.

But Openai has an important advantage: the growing corporate muscle.

Once upon a time, a SCAPPY Research Laboratory in the Mission area of ​​San Francisco has become more configured since then, and has returned more quickly to functional teams and accelerated the development and distribution cycles for products such as Sora.

Openai said Sora contains multiple protection layers to prevent unalesi non -secure content using rapid filtering and exit moderation between video frames and sound transcripts. Open content, terrorist propaganda and self -harm material financial prohibitions. The application also uses the imitation of watermark and prohibition similarity.

However, some users have already found ways to skirt.

Sora 2, the AI ​​model that strengthens Openai’s application, is a sharp development on the first version. The new system produces longer, more consistent clips that look real in a striking way.

Multiple viral videos depict popular cartoon characters such as Pikachu and Spongabob Squarepants, while others are involved in Altman after allowing it to be used on the platform.

The content fueled Openai’s criticism once again that he had moved faster than his own railings. The use of materials protected with copyrights – unless rights holders disable – are consistent with the current policy of the company, but this approach is challenged in court.

Altman, in an article in X, saying that Sora is related to transparency – showing what technology can do – Openai’s concerns about creating commercial acceleration to finance the wider ambitions around the artificial general intelligence.

The launch comes in the midst of competition. Meta Opened Vibraphone Last week, a new short -shaped AI video supply in Meta AI application. Google BANDANCE AND Alibaba at the same time He revealed competitor systems.

Openai, in the meantime, decided on new expenses of $ 850 billion and turned to infrastructure and new generation models.

Experts say that entering the video is not only about drawing more users to the ecosystem with another adhesive consumer application.

Professor Hao LiA leading expert in video synthesis told CNBC that most AI system is still trained on linguistic data such as books and internet text. However, in order to move towards general intelligence, he said that the models should learn from visual and audio information and that he should discover the world just like a baby.

“We use AI to produce content to train another model to perform better.”

Li added that the laboratory uses a video created by AI to feed synthetic data back to the system to improve model performance.

It is part of a wider trend among researchers who see video production as a way of simulating reality and helping models as a way of helping people like people.

Former Openai ruler Zack Kass, who investigated the social effects of artificial intelligence of the book “The next Renaissance: the expansion of AI and human potential” reiterated this view.

Kass, on a wider question of how model producers should approach the deployment, argued that the changes in early publishing strong technology are worth it.

“The building has two alternatives in the open: it is not to build or specially buildings. And these alternatives are worse for me,” CNBC said. “If we have a groundbreaking technology, I think people should know it and use it to update it.”

WRISTWATCH: Openai cements as the most valuable private company in the world