AI Briefing, March 26 — Issue #148

2026-03-26 16:00

Author: RadarAI Editorial Editor: RadarAI Editorial Last updated: 2026-05-11 Review status: Editorial review pending Brief 速报官方 AI动态开源

Anthropic launches Claude Coworker and Computer Use—its largest product release to date. Google unveils TurboQuant for 6x lossless KV cache compression. RISE and Itstone's AWE 3.0 advance embodied AI.

Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.

## 🔍 Key Insights Anthropic launched **Claude Cowork** and **Computer Use**, widely regarded as its **largest product release to date**, significantly boosting market engagement [12]. Meanwhile, Google unveiled the **TurboQuant** algorithm, achieving **6× lossless compression of KV cache**, setting a new benchmark for cost-efficient and high-throughput LLM inference [17]. In embodied AI, breakthroughs emerged from two distinct directions: the **RISE framework**, which trains robots via “world model imagination,” and **Ita Robotics’ AWE 3.0**, built on human-centric data and full-stack hardware—both tackling the field’s so-called “impossibility triangle” [10][23]. ## 🚀 Key Updates - **xAI’s founding team collapses: 11 co-founders down to just 1** [0]: Elon Musk is urgently restructuring xAI, bringing in key talent from Tesla and SpaceX while accelerating hiring to preserve core R&D capacity. - **Anthropic launches Claude Cowork and Computer Use** [12]: Officially labeled its most important update ever—greatly enhancing AI collaboration and system-level interaction capabilities. - **Google’s TurboQuant achieves 6× lossless KV cache compression** [17]: Leveraging innovations like polar-coordinate quantization, it slashes memory footprint and boosts inference throughput. - **RISE enables robots to self-evolve through “imagined” world models** [10]: Combines compositional world models with offline reinforcement learning—cutting real-world training costs dramatically. - **Ita Robotics’ AWE 3.0 breaks the embodied AI “impossibility triangle”** [23]: Validated via full-stack proprietary hardware and sub-millimeter flexible assembly—demonstrating an embodied AI scaling law. - **Shubham Saboo open-sources Gemini Embedding 2 for multimodal video search** [4]: Enables text- and image-based queries out of the box—delivering semantic-level video retrieval. - **Promptfoo launches: automated security testing for LLM applications** [24]: Supports red-teaming, penetration scanning, and vulnerability assessment—filling a critical gap in secure AI deployment. - **Moonshot AI unveils “Attention Residual” at GTC** [13]: Extends attention mechanisms across model depth—not just sequence length—improving long-range reasoning and inference stability. ## 🔗 Sources [0] Elon Musk’s xAI: 11 Co-Founders, 10 Gone — Only Tesla Insider Remains — https://www.bestblogs.dev/article/84c00739 [4] Multimodal Semantic Video Search Powered by Gemini Embedding 2 — https://www.bestblogs.dev/status/2037031910355198009 [10] RISE: Enabling Robots to Self-Evolve Through “Imagined” World Models — https://www.bestblogs.dev/article/7c4bcfc6 [12] [AI News] Claude’s Largest Launch Ever — https://www.bestblogs.dev/article/909bf5e0 [13] Moonshot AI Unveils “Attention Residual” at GTC — https://www.bestblogs.dev/status/2037010118957817988 [17] Google’s TurboQuant Paper: 6× Lossless KV Cache Compression — https://www.bestblogs.dev

Anthropic launched Claude Cowork and Computer Use, widely regarded as its largest product release to date, significantly boosting market engagement [12]. Meanwhile, Google unveiled the TurboQuant algorithm, achieving 6× lossless compression of KV cache, setting a new benchmark for cost-efficient and high-throughput LLM inference [17]. In embodied AI, breakthroughs emerged from two distinct directions: the RISE framework, which trains robots via “world model imagination,” and Ita Robotics’ AWE 3.0, built on human-centric data and full-stack hardware—both tackling the field’s so-called “impossibility triangle” [10][23].

🚀 Key Updates

xAI’s founding team collapses: 11 co-founders down to just 1 [0]: Elon Musk is urgently restructuring xAI, bringing in key talent from Tesla and SpaceX while accelerating hiring to preserve core R&D capacity.
Anthropic launches Claude Cowork and Computer Use [12]: Officially labeled its most important update ever—greatly enhancing AI collaboration and system-level interaction capabilities.
Google’s TurboQuant achieves 6× lossless KV cache compression [17]: Leveraging innovations like polar-coordinate quantization, it slashes memory footprint and boosts inference throughput.
RISE enables robots to self-evolve through “imagined” world models [10]: Combines compositional world models with offline reinforcement learning—cutting real-world training costs dramatically.
Ita Robotics’ AWE 3.0 breaks the embodied AI “impossibility triangle” [23]: Validated via full-stack proprietary hardware and sub-millimeter flexible assembly—demonstrating an embodied AI scaling law.
Shubham Saboo open-sources Gemini Embedding 2 for multimodal video search [4]: Enables text- and image-based queries out of the box—delivering semantic-level video retrieval.
Promptfoo launches: automated security testing for LLM applications [24]: Supports red-teaming, penetration scanning, and vulnerability assessment—filling a critical gap in secure AI deployment.
Moonshot AI unveils “Attention Residual” at GTC [13]: Extends attention mechanisms across model depth—not just sequence length—improving long-range reasoning and inference stability.

🔗 Sources

[0] Elon Musk’s xAI: 11 Co-Founders, 10 Gone — Only Tesla Insider Remains — https://www.bestblogs.dev/article/84c00739
[4] Multimodal Semantic Video Search Powered by Gemini Embedding 2 — https://www.bestblogs.dev/status/2037031910355198009
[10] RISE: Enabling Robots to Self-Evolve Through “Imagined” World Models — https://www.bestblogs.dev/article/7c4bcfc6
[12] [AI News] Claude’s Largest Launch Ever — https://www.bestblogs.dev/article/909bf5e0
[13] Moonshot AI Unveils “Attention Residual” at GTC — https://www.bestblogs.dev/status/2037010118957817988
[17] Google’s TurboQuant Paper: 6× Lossless KV Cache Compression — https://www.bestblogs.dev

← Back to Updates