![[20b7c83f-3374-4295-b0f1-014803862980.mp3]]

Anthropic发布开源“电路追踪”工具，提升AI模型透明度与可控性

简报：

• 5月29日，Anthropic宣布开源“电路追踪”(Circuit Tracing)工具，为大语言模型的内部决策路径提供可视化归因图，帮助研究人员深入探索AI模型如何从输入推理到输出结果的过程。
• 该工具结合Neuronpedia交互式前端，让用户能够直观分析神经元活动并测试不同假设，显著降低了AI决策机制研究门槛。
• Anthropic此举被认为是推动AI可解释性和透明化发展的重要里程碑，为理解和优化大模型性能及应对模型偏见或幻觉问题提供了有力技术支持。

相关链接：

• Anthropic开源“电路追踪”工具，解锁AI大脑决策机制

Anthropic Releases Open-Source "Circuit Tracing" Tool to Enhance AI Model Transparency and Control

Newsletter:

• On May 29, Anthropic announced the open-sourcing of its "Circuit Tracing" tool, which provides visual attribution maps for the internal decision-making paths of large language models. This helps researchers deeply explore how AI models reason from input to output. 🧠
• The tool integrates with the Neuronpedia interactive frontend, allowing users to intuitively analyze neuron activity and test various hypotheses, significantly lowering the barrier to studying AI decision-making mechanisms. 📊
• Anthropic's move is seen as a significant milestone in advancing AI interpretability and transparency, offering robust technical support for understanding and optimizing the performance of large models, as well as addressing issues like model bias or hallucinations. 🚀

Related Links:

• Anthropic Open-Sources "Circuit Tracing" Tool, Unlocking AI Brain Decision Mechanisms

EnCharge AI发布高能效AI加速器EN100，能耗仅为传统芯片的1/20

简报：

• AI芯片初创公司EnCharge AI正式推出基于模拟内存计算技术的AI加速器EN100，宣称能效比传统芯片高出20倍；
• EN100涵盖笔记本M.2和工作站PCIe两种规格，分别针对便携与高阶AI应用，计算能力高达1PetaOPS；
• 芯片采用独特架构，使每瓦AI计算性能及计算密度大幅提升，并可在本地运行先进AI模型，降低对数据中心和云的依赖；
• 国际能源署预计数据中心能耗将激增，EnCharge AI新产品被视为缓解能源危机的潜在解决方案；
• 公司由普林斯顿大学团队创立，已获1.44亿美元融资，并受DARPA等机构支持，正加速布局AI PC和边缘市场。

相关链接：

• EnCharge AI发布革命性AI加速器EN100，能效提升20倍
• AI芯片EnCharge EN100 能效高20倍，专为边缘及AI PC而生
• EnCharge AI raises over $100 million in funding to bring AI inference chips to market

EnCharge AI Unveils High-Efficiency AI Accelerator EN100, Consuming Only 1/20th the Power of Traditional Chips

Brief:

• AI chip startup EnCharge AI officially launches the AI accelerator EN100, based on analog in-memory computing technology, claiming 20x greater energy efficiency compared to traditional chips;
• EN100 is available in two formats—M.2 for laptops and PCIe for workstations—targeting portable and high-end AI applications, with computing power reaching up to 1 PetaOPS;
• The chip features a unique architecture that significantly boosts AI computing performance per watt and computational density, enabling advanced AI models to run locally and reducing reliance on data centers and the cloud;
• The International Energy Agency predicts a surge in data center energy consumption, and EnCharge AI's new product is seen as a potential solution to mitigate the energy crisis 🌍💡;
• Founded by a Princeton University team, the company has secured $144 million in funding and is supported by institutions like DARPA, accelerating its expansion into AI PC and edge markets 🚀.

Related Links:

• EnCharge AI Releases Revolutionary AI Accelerator EN100, Boosting Efficiency by 20x
• AI Chip EnCharge EN100 Offers 20x Higher Efficiency, Designed for Edge and AI PC
• EnCharge AI Raises Over $100 Million in Funding to Bring AI Inference Chips to Market

华为昇腾大模型实现无GPU高效解答数学难题

简报：

• 华为最新发布的“昇腾 + Pangu Ultra MoE”系统展示了近万亿参数的MoE大模型，在无GPU条件下，仅用2秒即可解答高等数学题。
• 该成果依赖于华为在训练系统性能优化、智能并行策略、通信机制改进以及负载均衡上的多项技术创新，实现了集群高效训练和单节点算力的显著提升。
• 这些技术突破大幅提升了大规模MoE模型的训练效率，并为AI模型的未来创新和应用带来了更多可能性。

相关链接：

• 华为发布昇腾大模型：无GPU训练高数大题仅需2秒

Huawei Ascend Large Model Achieves Efficient Math Problem Solving Without GPU

Newsletter:

• Huawei's newly released "Ascend + Pangu Ultra MoE" system showcases a MoE large model with nearly a trillion parameters, capable of solving advanced mathematics problems in just 2 seconds without a GPU. 🧮
• This achievement relies on Huawei's multiple technological innovations in training system performance optimization, intelligent parallel strategies, improved communication mechanisms, and load balancing, resulting in highly efficient cluster training and a significant boost in single-node computing power. 💻
• These technological breakthroughs greatly enhance the training efficiency of large-scale MoE models and open up more possibilities for future innovation and application of AI models. 🚀

Related Links:

• Huawei Releases Ascend Large Model: Solving Advanced Math Problems Without GPU in Just 2 Seconds

云从科技“从容 V2.0”多模态模型荣登OpenCompass全球榜首

简报：

• 近日，云从科技自主研发的多模态大模型“从容 V2.0”以80.7分的成绩，登顶全球知名OpenCompass多模态榜单，展现出在视觉感知、认知理解和跨领域应用等方面的突出表现。
• “从容 V2.0”具备处理图像、文本和音频等多种数据的能力，并在自动驾驶、智能安防、医疗影像等实际应用场景中表现优异。
• 此次成绩彰显了中国企业在多模态人工智能领域的创新突破，为相关行业的技术进步和智能化转型带来新动力。

相关链接：

• 云从科技多模态大模型登顶OpenCompass榜单
• 云从科技“从容 V2.0”模型登顶OpenCompass多模态榜单
• OpenCompass多模态榜单最新动态

CloudWalk Technology's "CongRong V2.0" Multimodal Model Tops OpenCompass Global Rankings

Newsletter:

• Recently, CloudWalk Technology's independently developed multimodal large model "CongRong V2.0" achieved a remarkable score of 80.7, securing the top position on the globally recognized OpenCompass multimodal leaderboard, demonstrating outstanding performance in visual perception, cognitive understanding, and cross-domain applications. 🌟
• "CongRong V2.0" is capable of processing various data types, including images, text, and audio, and has shown excellent results in real-world application scenarios such as autonomous driving, intelligent security, and medical imaging. 🚗
• This achievement highlights the innovative breakthroughs of Chinese enterprises in the field of multimodal artificial intelligence, bringing new momentum to technological advancements and intelligent transformation in related industries. 💡

Related Links:

• CloudWalk Technology's Multimodal Large Model Tops OpenCompass Leaderboard
• CloudWalk Technology's "CongRong V2.0" Model Secures Top Spot on OpenCompass Multimodal Leaderboard
• Latest Updates on OpenCompass Multimodal Leaderboard

“龙跃”生成式人形机器人运动大模型发布，上海加快智能数据生态建设

简报：

• 在2025张江具身智能开发者大会上，全球首款生成式人形机器人运动大模型“龙跃”（MindLoongGPT）正式发布，实现了以自然语言驱动高保真运动动作的多模态生成，推动人形机器人运动控制与数据采集能力的突破。
• 会上，国家地方共建人形机器人创新中心牵头，联合多地创新中心共同启动了产业标准互认计划，目标为人形机器人领域建立统一标准。
• 创新中心计划构建全国最大高质量数据集，目标包含500万条实体数据与2500万条合作数据，以促进高智能模型训练和应用。
• 同时，国际人形机器人技能大赛启动，吸引全国60余支顶尖队伍，设有五大赛道、覆盖28个高难度场景，推动技术创新落地。
• 浦东新区官方表示将加大对具身智能产业的政策投入，助力建设世界级人工智能产业集群，加速制造业与服务业等领域的智能转型。

相关链接：

• AIBASE新闻报道

"DragonLeap" Generative Humanoid Robot Motion Model Released, Shanghai Accelerates Smart Data Ecosystem Development

Newsletter:

• At the 2025 Zhangjiang Embodied Intelligence Developer Conference, the world's first generative humanoid robot motion model, "DragonLeap" (MindLoongGPT), was officially unveiled. It achieves multimodal generation of high-fidelity motion actions driven by natural language, marking a breakthrough in humanoid robot motion control and data collection capabilities. 🚀
• During the event, the National-Local Joint Humanoid Robot Innovation Center took the lead, collaborating with innovation centers across multiple regions to launch an industry standard mutual recognition program, aiming to establish unified standards for the humanoid robot sector.
• The Innovation Center plans to build the largest high-quality dataset in China, targeting 5 million entity data entries and 25 million collaborative data entries to facilitate the training and application of highly intelligent models.
• Meanwhile, the International Humanoid Robot Skills Competition was launched, attracting over 60 top teams from across the country. The competition features five major tracks covering 28 challenging scenarios, driving the practical implementation of technological innovation. 🤖
• Officials from Pudong New Area announced increased policy support for the embodied intelligence industry, aiming to help build a world-class artificial intelligence industry cluster and accelerate the smart transformation of sectors such as manufacturing and services. 🌟

Related Links:

• AIBASE News Report