Search AI Products and News

Large Language Models (LLMs) have achieved significant progress in complex reasoning tasks by combining task prompts with large-scale reinforcement learning (RL), as demonstrated by models like Deepseek-R1-Zero, which directly apply reinforcement learning to base models, showcasing strong reasoning capabilities. However, this success is difficult to replicate across different base model families, especially within the Llama series. This raises a core question: what factors lead to inconsistent performance of different base models during reinforcement learning? How does reinforcement learning perform in

Recently, a research team from the Fraunhofer Institute for Process Engineering and Packaging in Germany developed an artificial intelligence molecular scent prediction algorithm called OWSum, which successfully distinguishes American whiskey from Scotch whiskey, surpassing human experts in accuracy. The team trained this AI tool using flavor descriptions and chemical data of whiskey to explore its potential in whiskey identification. Image source note: The image was generated by AI, and the image licensing provider Midjourney was involved in the study, where researchers selected 16 samples.

Recently, a group of tech giants and venture capitalists jointly issued a statement opposing California Senator Scott Wiener’s SB1047 bill. This bill aims to strengthen the regulation of artificial intelligence, prompting strong opposition from the tech industry. Marc Andreessen and Ben Horowitz, co-founders of Andreessen Horowitz, along with Microsoft CEO Satya Nadella.

OpenAI researchers recently discovered an interesting phenomenon: the usernames chosen by users when interacting with ChatGPT may subtly influence the AI's responses. However, this influence is generally very small and primarily limited to earlier, unoptimized versions of the model. This study delves into the differences in ChatGPT's reactions to identical questions when faced with usernames related to different cultural backgrounds, genders, and ethnicities. The choice of usernames as a focal point for the research was made because names often carry specific cultural, gender, and ethnic connotations.

The study reveals that there are perceptual differences between AI-generated fake news and human-made fake news. The survey shows that 12% of participants are willing to share fake news, with no significant difference in sharing willingness between the two sources. Socioeconomic factors influence users' trust levels towards fake news, with conservative participants being more likely to believe AI-generated fake news. The research suggests introducing new labels and enhancing education on the risks of AI-generated fake news, and possible regulatory measures may be needed to protect vulnerable groups.