At a masterclass during the Lumière Festival, director Tsui Hark said AI is making it harder to judge the authenticity of content, as advertising, the internet, and AI blur the boundary between reality and illusion. He also linked AI to identity, saying its prevalence may affect how people understand themselves and the world. The festival is screening four of his classic films in a retrospective.
Shanghai AI Lab and SJTU LUMIA Lab released the NCP-ArchPreview Tech Report, proposing the Next Concept Prediction pre-training task. The first 8.9B-parameter discrete latent-space foundation model matched OLMo-3-7B's final pre-training loss using only 51.3% of the token budget, achieved a 1.95x equivalent convergence speedup and a nearly 6-point GSM8K improvement, with all assets open-sourced.
<body><div>On July 16, Xiaomi launched Xiaomi-Robotics-1, an embodied foundation model for real-world mobile manipulation. Pretrained on 100,000 hours of real-world data w</div></body>
<body><div>Global AI venture investment surged in Q1 this year with nearly 600 deals totaling over 110 billion yuan, up 185.4% year-on-year. Capital focuses on domestic la</div></body>
Google's AI Overviews feature, which powers search summaries, reportedly achieves an accuracy rate of approximately 90%. Given Google's annual search volume exceeding 5 trillion queries, this translates to a potential hourly generation of over 57 million incorrect answers, averaging nearly one million errors per minute. Independent evaluation by startup Oumi, analyzing 4,326 search results using the SimpleQA benchmark, indicates performance improvements. Accuracy for the Gemini 2 model was 85% in October last year, rising to 91% for the Gemini 3 model by February this year. However, Oumi's methodology, heavily reliant on AI tools for assessment, may introduce bias. Furthermore, Google can generate varying overviews for identical queries, adding to result inconsistency. A significant concern is the growing disconnect between AI Overviews and their source material. The rate of summaries not matching original source information increased from 37% with Gemini 2 to 56% with Gemini 3. This often leads users to encounter mismatched links or accurate summaries citing incorrect sources. Demonstrating vulnerability, AI Overviews have been manipulated, such as referencing content from a deliberately false blog post published just a day prior. Real-world examples highlight reliability issues. In one instance, a search for wrestler Hulk Hogan's death returned an AI Overview stating "no credible reports" of his passing, while directly below, a headline read "Mystery Deepens Over Hulk Hogan's Death," creating a contradictory user experience. Google has questioned Oumi's testing methodology, arguing it does not accurately reflect real-world search behavior. Key points underscore the tension between high accuracy rates and the substantial absolute volume of errors, alongside rising source discrepancies and demonstrable susceptibility to misinformation.