Group 1 - ByteDance is being questioned by Chinese regulators regarding the controversial smart assistant embedded in the upcoming Nubia M153 smartphone, raising concerns about cybersecurity, data security, and potential competition issues [4][5] - OpenAI and Microsoft are facing a lawsuit in the U.S. that links the ChatGPT AI chatbot to a murder case, marking the first instance of such a direct connection between AI chat tools and homicide [5] - Taobao has expanded its "no penalty for late delivery" policy nationwide, aiming to improve rider income and service efficiency through a positive incentive mechanism [5] Group 2 - JD.com is recruiting talent in the field of edge AI chips, focusing on integrated storage and computing chips for use in robots and smart home devices, offering salaries ranging from 40,000 to 100,000 [9][10] - Douyin has launched a "Douyin Pay" feature, allowing consumers to pay at merchant stores through the Douyin app, streamlining the payment process without needing to switch to third-party applications [10] - Honor's brand marketing president, Guo Rui, has left the company, potentially to pursue entrepreneurial ventures, after previously holding significant roles at Huawei and Honor [10] Group 3 - JD.com plans to invest 22 billion over the next five years to provide 150,000 housing units for delivery personnel, enhancing living conditions for its workers [13] - OpenAI has released GPT-5.2, which has shown improved benchmark scores compared to Google, indicating a competitive response to Google's advancements in AI [16] - SoftBank's Masayoshi Son has reduced the amount of shares pledged to lenders by 2.1 billion, reflecting a strategic move amid fluctuating tech wealth driven by AI investments [16] Group 4 - Wanda Film has completed a strategic investment in the interactive entertainment brand "Pailifang," aiming to explore new consumer scenarios in image socialization [17] - Several companies, including Nidejia and Bihui Biotechnology, have recently completed significant financing rounds, indicating active investment interest in various sectors [17][18] - Samsung has launched its first tri-fold smartphone, the Galaxy Z Trifold, in South Korea, priced at approximately 1.72 million KRW [18] Group 5 - Tesla's U.S. sales in November fell nearly 23% year-on-year, reaching the lowest level in four years, largely due to the removal of federal tax incentives for electric vehicles [23][27] - The global shipment of foldable smartphone panels is expected to grow by 46% by 2026, driven by demand from Apple's first foldable iPhone [23] - The Chinese market is seeing a significant increase in the box office for the 2025 New Year film season, surpassing 3.5 billion [25]

万万没想到，年初还是个小众方向的「扩散语言模型（dLLM）」，现在已经被扩展到千亿参数的规模了。 前段时间，我们在 HuggingFace 页面发现了两个新模型：LLaDA2.0-mini 和 LLaDA2.0-flash。它们来自蚂蚁集团与人大、浙大、西湖大学组成的联合团 队，都采用了 MoE 架构。前者总参数量为 16B，后者总参数量则高达 100B—— 在「扩散语言模型」这个领域，这是从未见过的规模。 更令人欣慰的是，模型变大了，也确实变强了：在涵盖知识、推理、编码、数学、智能体与对齐几大维度的 47 个基准测试中，LLaDA2.0-flash 平均得分 73.18，与强 AR（自回归）模型 Qwen3-30B-A3B-Instruct-2507（73.60）持平，在编码（如 HumanEval、MBPP）、智能体（BFCL）等复杂任务上优势显 著。 长期以来，自回归生成范式在大模型领域始终占据主导地位，这种从前到后依次生成下一个 token 的方法曾被寄予厚望。然而，其固有弊端也逐渐显现： 长文本生成的计算成本较高、推理速度较慢，且难以捕捉 token 之间的双向依赖关系。一旦前期生成的内容出现错误， ...

机器之心报道 编辑：杜伟、张倩 万万没想到，年初还是个小众方向的「扩散语言模型（dLLM）」，现在已经被扩展到千亿参数的规模了。 前段时间，我们在 HuggingFace 页面发现了两个新模型：LLaDA2.0-mini 和 LLaDA2.0-flash。它们 来自蚂蚁集团与人大、浙大、西湖大学组成的联合团队，都采用 了 MoE 架构。前者总参数量 为 16B，后者总参数量则高达 100B—— 在「扩散语言模型」这个领域，这是从未见过的规模。 更令人欣慰的是，模型变大了，也确实变强了：在涵盖知识、推理、编码、数学、智能体与对齐几大维度的 47 个基准测试中，LLaDA2.0-flash 平均得分 73.18， 与强 AR（自回归）模型 Qwen3-30B-A3B-Instruct-2507（73.60）持平 ，在编码（如 HumanEval、MBPP）、智能体（BFCL）等复杂任务上优势显著。 长期以来，自回归生成范式在大模型领域始终占据主导地位，这种从前到后依次生成下一个 token 的方法曾被寄予厚望。然而，其固有弊端也逐渐显现：长文本生 成的计算成本较高、推理速度较慢，且难以捕捉 token 之间的双向 ...