OpenAI announced the termination of its partnership with Scale AI, a startup specializing in data annotation. This decision was made after Mark Zuckerberg invested $14.3 billion in the company and recruited its founder, Alexander Vang, to develop “superintelligence.” OpenAI noted that Scale AI was used only for a small part of their needs, and as models became more complex, there was a need to find more competent partners. They have now switched to a different provider—Mercor, which previously focused on recruitment and now helps in AI development.
This move is a blow to Scale AI’s reputation: within just one week, they lost one of their key clients and dropped from leading positions.
After a week of work, Midjourney finally introduced its first video model, which transforms images into short animated clips. The Animate feature allows turning any picture into dynamic video with automatic or manual control. There are Low Motion and High Motion modes to adjust movement, and the clip duration is limited to 16 seconds—with captions possible for each segment. External images are supported—they can be uploaded as starting frames and their movement controlled via text prompts.
Currently, videos are available in 480p resolution at 24 frames per second. The cost of generating a video is eight times higher than creating static images; however, the company claims it is roughly 25 times cheaper than competitors. Pro subscribers are testing the Video Relax Mode, which allows saving priority minutes of usage. Creators see this as an intermediate step toward integrating 3D content, video, and real-time capabilities on a single platform.
SandboxAQ, spun off from Google and supported by Nvidia, released an open dataset of three-dimensional structures of protein molecules and drug compounds. This database is designed to assist scientists in predicting interactions between substances and target proteins within the body—an essential task in early drug development stages. If a molecule does not “stick” to the target protein, the drug is unlikely to work.
Using Nvidia’s powerful chips, the company generated about 5.2 million theoretical molecules—artificial structures based on real data. This information trains AI models to quickly determine the likelihood of binding within minutes, significantly reducing manual calculation time.
A study by MIT revealed a concerning trend: using ChatGPT reduces brain activity in users by nearly a third—especially in areas responsible for creativity and memory. During the experiment, participants wrote essays either independently or with help from ChatGPT and Google. Those relying on AI showed noticeably lower neural activity; they began copying AI-generated answers and lost skills for independent idea formulation. The group without prompts used their brains more actively and could better generate ideas.
Researcher Natalia Kosmina warns that excessive reliance on AI may hinder critical thinking development in children and impair learning skills. During her work, she inserted special “traps” into the article text to test AI summarizers—finding that they started inventing versions of ChatGPT that were not present in the original data. The next phase of research will focus on AI’s impact on programming: initial results are even more alarming.
Baidu conducted a unique online marketing stream with digital avatars. The event featured two virtual characters—a digital twin of a popular Chinese streamer, Luo Yonghao, controlled by the ERNIE model. This was the first such case: both avatars hosted a live broadcast simultaneously.
Over six hours of streaming, the audience exceeded 13 million views; electronics products worth over 55 million yuan (approximately $7.6 million) were sold. The ERNIE technology provided realistic interaction between the avatars—they could communicate in real-time and generate detailed product descriptions exceeding 97,000 characters.
Such use of digital twins could fundamentally change approaches to online commerce and marketing in the future.