Core Insights - The event showcased China's AI model "AI Sun Wukong," which supports real-time interaction in Chinese, Japanese, and English, highlighting advancements in multilingual AI technology [1][3][6] - Experts from various countries expressed enthusiasm for the potential of multilingual AI models to facilitate cross-cultural dialogue and collaboration in the AI field [3][6] Group 1: Event Overview - Approximately 30 AI experts from 20 countries and regions in Asia and Europe gathered in Osaka, Japan, to experience the "AI Sun Wukong" exhibit [1] - The exhibit featured live demonstrations of the AI model's capabilities in multilingual interactions, creating an engaging atmosphere for attendees [1][3] Group 2: Expert Opinions - Gábor Prószéky, a linguistics expert from Hungary, emphasized the importance of multilingual AI models as bridges for civilizational dialogue and expressed hopes for collaboration between China and Hungary in AI [3] - Yu Yong Poh from Malaysia highlighted the cultural nuances reflected in the AI's responses and noted its support for over 130 languages, which is beneficial for researchers from smaller language countries [3][6] Group 3: Technological Capabilities - The "AI Sun Wukong" model, developed by iFLYTEK, integrates advanced technologies such as multilingual speech recognition, emotional voice synthesis, and multimodal interaction [3] - The model aims to democratize access to knowledge by allowing users to interact in their native languages, addressing the dominance of English in AI resources [6]

Core Viewpoint - The article discusses the introduction of MetaCLIP 2, a novel method for training the CLIP model on a global scale without relying on external resources, addressing the challenges of multilingual data processing and enhancing model performance across languages [2][4]. Group 1: MetaCLIP 2 Overview - MetaCLIP 2 is the first method to train CLIP from scratch on native global image-text pairs, overcoming the limitations of previous models that primarily focused on English data [2][5]. - The method includes three core innovations: metadata expansion to over 300 languages, a data filtering algorithm to balance concept distribution across languages, and a global training framework that proportionally increases the use of image-text pairs as non-English data is introduced [5][20]. Group 2: Performance Improvements - MetaCLIP 2 demonstrates that non-English data can enhance the capabilities of English models and vice versa, effectively breaking the "multilingual curse" [10][31]. - The model achieved state-of-the-art (SOTA) results in various multilingual benchmarks, including improvements of 3.8% on Babel-ImageNet and 1.1% on XM3600, among others [32][34]. Group 3: Training Methodology - The training framework of MetaCLIP 2 maintains consistency with OpenAI's CLIP architecture while introducing key components such as a multilingual text tokenizer and scaling of seen training pairs [26][30]. - The model's training data was expanded from 13 billion pairs to 29 billion pairs, resulting in significant performance enhancements across both English and multilingual tasks [38][39]. Group 4: Cultural and Linguistic Diversity - MetaCLIP 2 retains a comprehensive distribution of global images, enhancing geographical localization and regional recognition capabilities [13][15]. - The model directly learns from image descriptions written by native speakers, avoiding reliance on machine translation, which improves the authenticity and accuracy of the training data [12][16].

Core Viewpoint - The international academic seminar on high-level multilingual models highlighted the urgent need to address the digital divide affecting low-resource languages, as global mainstream models inadequately support these languages, posing a risk of marginalization for small language countries [1][5]. Group 1: Expert Opinions - Professor Vlado Delić from Serbia emphasized that Serbian language representation in general models is less than 0.1% in token usage, significantly lower than Slovenian, advocating for national-level models that reflect cultural identity to avoid translation errors in critical fields like healthcare and law [3]. - Avner Algom, founder of the Israeli Human Language Technology Association, stated that language services should not be designed solely for major languages, advocating for technological equity for small languages [5]. - Nipat Jongsawat from Thailand highlighted that language sovereignty is a strategic necessity for nations, not merely a choice [5]. Group 2: Technological Developments - Liu Cong, director of iFlytek Research Institute, reported that their multilingual model, Spark X1, expanded from supporting 81 languages in October 2022 to over 130 languages by July 2023, aiming to provide a comprehensive multilingual model and applications [5]. - The Spark X1 model reportedly outperforms GPT-4.1 in key languages such as Arabic, German, French, Korean, and Japanese, while the speech synthesis model supports 55 languages with leading industry performance [5]. - Zhang Xiao, deputy general manager of iFlytek's intelligent computing division, identified challenges in the rapid iteration of computing power and low efficiency in existing resources, proposing a five-element solution combining computing power, algorithms, data, applications, and ecosystem [7]. Group 3: Data Quality and Collaboration - Gábor Prószéky, director of the Hungarian Linguistics Research Center, stressed that data quality is more critical than quantity for building reliable large language models, noting the unique challenges posed by the Hungarian language's complex morphology [7]. - His team has developed the PULI model family, collaborating with Chinese AI counterparts to create a complete closed loop from training to application [7].

Core Viewpoint - The article highlights the risk of marginalization faced by low-resource language countries in the AI era due to insufficient support from general large models for these languages [1][2]. Group 1: Language Models and Their Importance - Serbian language usage is significantly lower than Slovenian, with its token representation in general models being less than 0.1% [2]. - The need for language models to reflect cultural identity is emphasized, particularly in critical fields like healthcare and law [2]. - Hungarian language presents unique challenges for language models due to its complex morphology, underscoring the importance of data quality over quantity [2]. Group 2: Initiatives and Collaborations - The Israel Association for Human Language Technology (IAHLT) has developed a bilingual model (Hebrew + English) based on open-source models, addressing the need for technology equity for small languages [4]. - iFLYTEK's latest model, X1, supports over 130 languages and aims to collaborate globally to create comprehensive multilingual models and applications [4].