DIJA团队 投稿 量子位 | 公众号 QbitAI 答案是：不能。 在最新研究《The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs》中，来自上海交通大学, 上海人工智能 实验室和中山大学的研究团队指出： 当前扩散语言模型存在根本性架构安全缺陷，在某些攻击场景下，几乎毫无防御能力。 他们提出的DIJA攻击框架，不需要训练，也不需要改写模型参数，就能令多个dLLMs大概率生成有害、非法、违规内容。这项研究不仅击穿 了扩散语言模型的"对齐护盾"，也正式拉开了dLLMs 安全研究时代的序幕。 扩散语言模型（Diffusion-based LLMs，简称 dLLMs）以其并行解码、双向上下文建模、灵活插入masked token进行解码的特性，成为一 个重要的发展方向。 相较传统的自回归语言模型，dLLMs 既可以一次性生成多个位置的词，也能更自然地完成文本插入、改写、补全等任务，被广泛应用于交互 式问答、代码生成、甚至多模态场景。 但在这股看似技术跃迁的浪潮背后，一场 潜藏的安全危机 正在酝酿：能否继续沿用自回归 ...

6/ New name, new modes, same vibes, Eternum is evolving into @realms_gg.They’re also launching a new game mode: Blitz (2h games), alongside the OG week-long Eternum sessions.Realms ☁️ (@realms_gg):Attention Eternum enjoyoors,The game's evolving, and with it, our name (& handle). Eternum is now 1 of a growing number of game modes:Realms: EternumRealms: BlitzDetails in the 🧵↓ https://t.co/bKsre69m0L ...

RT Realms ☁️ (@realms_gg)In Realms: Blitz, economic gameplay is simplified so players can focus on strategic action.Unlike Realms: Eternum, there's no Order Book/AMM or transfers between allies.Think 2hr onchain Battle Royale.Sharpen your swords & stack those $lords, anon.Blitz is coming. https://t.co/3MyBzXS5M6 ...

点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 今天自动驾驶之心为大家分享 上海科技大学&港中文ICCV'25中稿的最新工作— ReAL-AD！ 迈向类人推理的端到端自动驾驶。 如果您有相关工作需要分享，请在 文末联系我们！ 自动驾驶课程学习与技术交流群事宜，也欢迎添加小助理微信AIDriver004做进一 步咨询 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 论文作者 | Yuhang Lu等 编辑 | 自动驾驶之心 写在前面 & 笔者的个人理解 端到端自动驾驶作为一种统一感知、预测和规划的方法，能够在统一框架内减少信息损失并提高适应性， 近年来备受关注。然而，现有方法通常依赖于固定且稀疏的轨迹监督，这限制了它们捕捉人类驾驶员自然 采用的层次化推理过程的能力。为了弥补这一差距，上海科技大学和港中文的团队提出了ReAL-AD，这是 一种推理增强学习框架，基于三层人类认知模型（驾驶策略、驾驶决策和驾驶操作）来构建自动驾驶中的 决策过程，并引入视觉-语言模型（VLMs）以增强环境感知和结构化推理能力。具体而言，我们引入了： (1) 策略推理Injecto ...

RT Realms ☁️ (@realms_gg)In Realms: Blitz, settling takes place over a tense 30 minute period.Positions are randomised, with players spawning at equal distances between the coveted Hyper Structures.When the horn sounds, everything’s fair game.Prepare accordingly, anon. Blitz is coming. https://t.co/JreCoP1fTY ...

Solana DAO governance just got smoother 🤌- Advanced treasury management- Transparent voting heat maps- Mobile support and fully open sourceTry out Realms v2 now!Realms - The Home of Solana DAOs (@realmsdaos):Realms v2 is live.The next chapter for onchain governance on Solana.Built for the next wave of DAOs.Faster. Sharper. Yours. https://t.co/wWsxUoTt3O ...

Core Insights - The Second China (Shenyang) Comedy Film Week concluded on July 12, 2025, focusing on the theme "Joy in Liaoning, Laughter on Screen" [1] Event Overview - The event featured key activities such as the "Brilliant Chinese Cinema Tour", comedy film screenings, a roadshow, matchmaking sessions for comedy film scripts, and a seminar on the development of Chinese comedy films [4] - Sideline activities included "Savoring Cuisine through Films" and "On-Screen Journeys", emphasizing the connection between Shenyang and Chinese comedy films [4] Audience Engagement - The film week attracted over 100 movie fan clubs and nearly one million film lovers nationwide, highlighting fan-driven recommendations [5] - Award-winning comedy films released domestically from early 2023 to May 1, 2025, were unveiled during the closing ceremony, with the film "Successor" receiving the title of "Comedy Film Recommended by Movie Fan Clubs" [5] Organizational Goals - Co-hosted by various governmental bodies, the film week aims to enhance cultural originality, boost the film industry's high-quality development, and elevate Liaoning's cultural influence [6] - The initiative seeks to develop Shenyang's unique cultural identity and create a model zone integrating culture, sports, and tourism, contributing to Liaoning's status as a leading cultural province [6]

The future of AI is personal, where you can own your data, Run your AI, and Earn rewards.@PinLinkAi AI is a privacy-first platform for Personal AI, built to help you reclaim control over your data and run your own private AI, right on your device.Why It Matters?➡️ We all leave pieces of our data across Google, Apple, Meta, and Web3 platforms➡️ PIN AI lets you unlock that fragmented data and use it to train your own AI, without giving control to Big Tech.➡️ It’s powered by Trusted Execution Environments (TEE ...

Core Insights - Meta has proposed a compensation package exceeding $200 million to attract the head of Apple's Foundation Models team, Pang Ruoming [1][3] - Pang leads a team of approximately 100 people at Apple, responsible for developing large language models that are crucial for Apple Intelligence and other AI functionalities [1] - The proposed salary is second only to Apple's CEO Tim Cook and aligns with the compensation standards for other key members of Meta's newly established Superintelligence Team [3] Compensation Structure - The compensation package for the Superintelligence Team is among the highest in global corporate positions, surpassing that of CEOs at major banks [3] - A significant portion of the compensation is performance-based and requires long-term service to unlock fully, meaning early departure or poor stock performance could result in reduced payouts [3] - The structure includes base salary, signing bonuses, and Meta stock, with stock being the most critical component [3][4] Stock and Contract Terms - Meta typically ties stock payments to specific growth metrics of Meta's stock over designated years, often extending beyond the standard four-year vesting period [4]

Group 1 - Meta is offering over $200 million in compensation to recruit Ruoming Pang, head of Apple's Foundation Models team, which is approximately 1.436 billion RMB [1][3] - Ruoming Pang leads a team of about 100 people at Apple, responsible for developing large language models that are foundational to Apple Intelligence and other AI functionalities [1] - The compensation package offered by Meta is among the highest in the industry, second only to Apple's CEO Tim Cook [3] Group 2 - Meta's newly established Superintelligence Team (MSL) has compensation packages that align with Pang's offer, aimed at building AI systems that can perform tasks at or above human levels [3] - Key members of the MSL include former GitHub CEO Nat Friedman and AI startup founder Daniel Gross [3] - Meta's compensation structure includes base salary, signing bonuses, and stock options, with stock being the most significant component [3] Group 3 - The Superintelligence Team's compensation exceeds that of CEOs at major global banks, although much of it is performance-based and tied to long-term service [3] - Meta typically offers higher signing bonuses to compensate for equity forfeited from previous startups when hiring new employees [3] - Contracts often stipulate that stock payments are linked to specific growth metrics of Meta's stock over designated years, with many contracts extending beyond the typical four-year vesting period [3]