Aggregator

Google TurboQuant AI 压缩算法大幅减少大模型内存使用

Solidot
1 week 2 days ago
Google 研究院发布了压缩算法 TurboQuant，能在大幅减少大模型内存占用的同时提高速度和维持精度。TurboQuant 旨在减小键值缓存的大小，被称为是储存重要信息减少再计算的“数字查找表（digital cheat sheet）”。大模型并不理解任何东西，它通过映射词元文本语义的向量去模拟对事物的理解。大模型的向量通常使用 XYZ 坐标进行编码，而实现 TurboQuant 压缩的系统将向量转换为笛卡尔坐标系的极坐标，向量被简化为两类信息：半径（核心数据强度）和方向（数据含义）。如果使用 XYZ 坐标编码向量，那么特定位置可以编码为“向东走 3 个街区，向北走 4 个街区”，采用笛卡尔坐标编码向量，那么同样的信息编码为“沿 37 度方向走 5 个街区” ，简化了空间节省了计算。Google 的早期测试显示，TurboQuant 在部分测试中实现了 8 倍的性能提升，内存占用减少到原来的六分之一，同时质量没有损失。实现 TurboQuant 算法将有助于降低 AI 模型的运行成本和内存占用，但也可能推动更复杂模型的出现，因此对降低内存价格可能没有什么效果。

AI Threat Landscape Digest January-February 2026

Check Point Research
1 week 2 days ago

KEY FINDINGS AI-assisted malware development has reached operational maturity.VoidLink framework, which is modular, professionally engineered, and fully functional, was built by a single developer using a commercial AI-powered IDE within a compressed timeframe. AI-assisted development is no longer experimental but produces deployment ready output. AI-assisted development is not always obvious from the final product.VoidLink was […]

The post AI Threat Landscape Digest January-February 2026 appeared first on Check Point Research.

matthewsu

AI Agents Redefine Enterprise Cybersecurity Risk

BankInfoSecurity.com
1 week 2 days ago
Menlo Ventures' Rama Sekhar on Securing AI Agents and Non-Human Identities
As AI evolves from assistants to autonomous agents, enterprises face a new attack surface driven by non-human identities. Rama Sekhar, partner at Menlo Ventures, explains why visibility, governance and AI-driven remediation are critical to securing this evolution.

AI Versus AI: The Future of Cyber Defense

BankInfoSecurity.com
1 week 2 days ago
Segura's Joe Carson on Agentic AI, Cyber Resilience and Estonia's Lessons
AI is accelerating both attackers and defenders, transforming cybersecurity into an AI-versus-AI battle. Segura's Joe Carson discusses why organizations must treat agentic AI as a force multiplier, not a replacement, and how to harness it responsibly in a future driven by autonomous agents.

Where AI Labs Will and Won't Disrupt Cybersecurity

BankInfoSecurity.com
1 week 2 days ago
Foundation Capital's Sid Trivedi on the Three Markets AI Labs Can't Easily Enter
AI labs are moving into application security, but three structural barriers define where they won't go, and that's where the next generation of durable security companies will be built, said Sid Trivedi, partner at Foundation Capital.

AI Agents Redefine Enterprise Cybersecurity Risk

DataBreachToday.com
1 week 2 days ago
Menlo Ventures' Rama Sekhar on Securing AI Agents and Non-Human Identities
As AI evolves from assistants to autonomous agents, enterprises face a new attack surface driven by non-human identities. Rama Sekhar, partner at Menlo Ventures, explains why visibility, governance and AI-driven remediation are critical to securing this evolution.

AI Versus AI: The Future of Cyber Defense

DataBreachToday.com
1 week 2 days ago
Segura's Joe Carson on Agentic AI, Cyber Resilience and Estonia's Lessons
AI is accelerating both attackers and defenders, transforming cybersecurity into an AI-versus-AI battle. Segura's Joe Carson discusses why organizations must treat agentic AI as a force multiplier, not a replacement, and how to harness it responsibly in a future driven by autonomous agents.

Managed ad