Meta AI的研究者发表的论文《Remember When It Matters：长程Agent的主动记忆干预》，提出了一种将Agent记忆从被动检索转变为主动干预的全新范式。在长程任务中，决策相关信息虽然存在于对话历史或上下文窗口中，但不再可靠地影响后续决策——作者称之为行为状态衰减。该研究通过双Agent架构和选择性干预机制，在Terminal-Bench 2.0上将Claude...

In an era of rapid AI disruption, the winners may not be the companies that predict the future most accurately, but those that learn, reallocate resources, and adapt faster than their competitors.

Even before the Middle East conflict began, geopolitics surged as the foremost investor concern. AI disruption and capital allocation discipline are also top of mind.

Once you move from self-managed MySQL to a MySQL DB System with HeatWave on OCI, the tuning story changes in an important way. On a self-managed server, you worry about two layers: MySQL and the...

Five lessons can help leaders build two stronger companies.

Most organizations are getting AI adoption backwards. Instead of building the foundational scaffolding that makes AI agents work at scale, they chase flashy “big” use cases and set themselves up...

Fast Company has named Firefox to its 2025 “Brands That Matter” list, recognizing companies that go beyond acquiring customers to build meaningful relevance and cultural impact. For us, that honor...

When we talk about Artificial Intelligence in UX, we often hear: “How do we make users trust the system?” It sounds intuitive — after all, trust is central to how humans cooperate. But psychology...

Shifts in greenfield FDI are shaking up the industries of the future and rewiring global trade. Geographic distance is easy to understand: look at a map and countries that are eight centimeters...

Designers today are quick to praise how AI speeds up their workflows. And that makes sense — businesses now more than ever need fast drafts, rapid testing, and quick launches to keep users...

Are AI capabilities making learning to code obsolete? Not according to the leaders at Kode With Klossy, who say that coding can lead to many career skills that apply across industries.

A healthy culture can triple a corporate venture’s TSR by defining underlying beliefs, focusing on leadership, and modeling and celebrating desired behavior.

本研究针对混合专家（MoE）语言模型在资源受限环境中的应用挑战，提出了知识增强（KA）和学生感知路由器（SAR）两种知识蒸馏方法，有效提取专家知识，实验结果优于传统方法。

本研究提出了一种新的视角，针对多模态大语言模型中视觉标记的计算开销问题。通过DART修剪方法，能够在信息损失较少的情况下显著加快处理速度，实验表明DART可修剪88.9%的视觉标记并提升速度。

本研究提出了一种名为HILO的层次性方案，旨在优化大规模语言模型的微调过程。HILO通过动态调整适配器专家的数量和秩，以适应模型层的复杂性。实验结果表明，HILO在准确性和可训练参数方面优于现有方法，提供了高效的微调解决方案。

亚马逊云科技推出新系列Amazon Nova，强调多模型选择而非单一强大模型。新模型支持文本、图像和视频生成，成本显著降低，吸引开发者，推动市场多样化发展。

本研究提出了一种目标领域结构平滑（TDSS）方法，以解决无监督图领域适应中的知识传递问题。通过在目标图上实施结构平滑，TDSS显著提升了迁移效果，展现了广泛的应用潜力。

Apple turns to strategic investment post the iPhone 16 ban in Indonesia Apple plans to invest US$1 billion in Indonesia to strengthen its ties with Southeast Asia's largest economy. This move...