DeepSeek’s AI Impact, Legal Challenges, and Industry Innovations

AI : DeepSeek's AI Impact, Legal Challenges, and Industry Innovations

A Chinese AI company, DeepSeek, has released a new model called Janus-Pro. This model is an upgrade from its predecessor, Janus, and is designed to be larger and more powerful. Janus-Pro can create and understand images, with improvements in following prompts and analyzing pictures. However, it still has a limitation with image resolution at 384 × 384 pixels, affecting fine details like facial recognition. DeepSeek plans to address this with higher resolutions in the future. Janus-Pro, like other DeepSeek models, is available on Hugging Face.

DeepSeek’s recent activities have caused significant reactions in the stock market. Although their models R1 and V3 were released earlier, DeepSeek gained attention when investor Marc Andreessen praised it as a major breakthrough. Over the weekend, DeepSeek’s app surpassed competitors like OpenAI’s ChatGPT in the US Apple App Store, leading to market panic on Monday. DeepSeek claims their models are trained at lower costs than competitors, causing Nvidia’s stock to drop by over 17%, with Oracle and Microsoft also seeing losses. Even companies like Siemens Energy, which could benefit from AI’s energy demands, lost nearly 20% of their stock value. Experts are divided on whether this panic is justified, with some questioning DeepSeek’s cost claims.

In India, major news outlets have joined a lawsuit against OpenAI, accusing it of using their content without permission for training ChatGPT. The Indian Express and Hindustan Times are among those involved. Publishers demand OpenAI stop accessing copyrighted content and delete datasets used for training. OpenAI has not responded but generally denies similar claims, arguing its systems use publicly available data. The company claims Indian courts cannot handle copyright cases since servers are abroad. However, publishers argue OpenAI operates in India, thus subject to Indian law. Similar lawsuits exist in the US and Canada.

A new AI benchmark called RealCritic has been developed by researchers from the Chinese University of Hong Kong, Shenzhen, Alibaba’s Qwen Team, and the Shenzhen Research Institute of Big Data. It tests AI models’ ability to self-critique and correct errors. OpenAI’s o1-mini model performed best in these tests, surpassing competitors like GPT-4o, Qwen2.5, and Llama 3.1. The benchmark evaluates self-critique, critique of other models, and iterative critique over multiple rounds. OpenAI’s model excelled in all areas, notably improving its performance through self-critique.

Meta, led by Mark Zuckerberg, plans to invest $60-65 billion in data centers for AI algorithm training. By year-end, Meta aims to operate over 1.3 million GPU accelerators, including new additions. The company is building a new data center in Rishland Parish, Louisiana, designed for two gigawatts, roughly half the size of Manhattan. This facility will accommodate new hardware, with no details on the energy source provided.

The AI startup Perplexity AI has proposed a merger with TikTok US. The plan involves the US government holding up to 50% of a new holding company after a public offering valued at $300 billion. ByteDance would contribute TikTok’s US operations without the recommendation algorithm. Perplexity would join the holding, gaining access to TikTok’s video content, potentially integrating its AI search technology. This merger presents an opportunity for Perplexity to compete with Google and expand its user base.

Three Swedish researchers have raised concerns about AI-generated “research” appearing in Google Scholar, databases, and journals. These false entries could overwhelm quality control in science, threatening the integrity of scientific records. Generative AI could create misleading documents that appear scientific, prioritized by search engines like Google Scholar. This undermines trust in science and poses societal risks. Researchers suggest addressing this issue with technology, education, and regulation, emphasizing the need to understand how fraudulent works reach audiences and why they persist.

UBTech, a Chinese robotics company, is preparing for mass production of its Walker-S humanoid robots, aiming to complete this by 2025. Between 500 and 1000 units of Walker S, Walker S1, and the upcoming Walker S2 will be delivered to industries such as automotive, logistics, and Apple supplier Foxconn. The Walker S1 has 41 degrees of freedom and various sensors, equipped with AI for task planning and autonomous operation. UBTech plans to introduce the Walker S2, a lighter, stronger, and more precise model with advanced image processing and AI algorithms for faster task planning.