如何让没有长时记忆的AI，完成持续数小时的复杂任务？Anthropic设计出一个更高效的长时智能体运行框架，让AI能够像人类工程师一样， 在跨越数小时的任务中渐进式推进。 假如你雇佣了一支24小时轮班的工程师团队，要求他们一起开发一款复杂应用。 但有一个奇怪规定：每位工程师一上班就完全忘记上一班做过什么，只能从零开始重新干。 无论他们技术多强，工作多努力，这个项目恐怕也做不成。 而这正是「长期运行智能体」在现实中遭遇的真实困境： 「上下文窗口一关，AI就失忆」。 模型没有真正的长期记忆，所有判断都依赖当下能看到的文本片段，上下文窗口一满或被关掉，就像白板被擦掉一样。 这种「记忆缺陷」，让智能体做不了长工程，一旦任务需要持续数小时、跨越多轮对话窗口时，这样的问题就会暴露出来。 由于上下文窗口有限，而大多数复杂项目无法在单一窗口完成，因此智能体必须找到一种能够跨越多轮编码会话的有效机制。 近日，Anthropic通过「偷师」人类工程师，形成了一套适用于长期运行智能体的有效框架。 https://www.anthropic.com/engineering/effective-harnesses-for-long-r ...

SemiAnalysis的这个报告已经出来两天，但周末一直比较忙，只能拖到今天才来写了。这篇TPU分析报告引起的 争议是不小的，下面我们结合报告内容，再加上自己的分析，来看下这个报告的内容，后面还会给出国内谷歌 的链的更新情况。 首先说一下国内外的一个通病，就是拉一踩一，最近谷歌链火了，在提谷歌的时候就非要踩一下英伟达。这是 完全没有必要的，英伟达现在依然是绝对的老大，CUDA依然是很强的护城河。 TPU的技术表现已明确引发竞争对手的高度关注，奥特曼公开表示，由于谷歌Gemini模型的强势表现抢占了行业 焦点，OpenAI正面临阶段性挑战。英伟达也发布公关声明以稳定市场预期，强调其在该领域仍保持领先优势。 近几个月来，谷歌DeepMind、谷歌云（GCP）与TPU形成的技术生态协同效应显著，实现了多项关键进展： TPU产能规划大幅上调，Anthropic宣布部署规模超1GW的TPU计算集群，Gemini 3、Opus 4.5等具备业界领先水 平的模型均基于TPU完成训练，且Meta、SSI、xAI、OpenAI等头部机构已陆续加入TPU采购队列，客户清单持 续扩容。 与此同时，以英伟达GPU为核心的供应链则面 ...

Core Insights - The launch of Gemini 3 and the Ironwood AI chip marks a significant comeback for Alphabet in the AI sector, generating excitement among experts and investors [1][2][3] Product Developments - Google introduced Ironwood, its seventh-generation tensor processing unit (TPU), which allows customers to run and scale large, data-intensive models [2] - Gemini 3, the latest AI model, requires less prompting and provides smarter answers compared to its predecessors [2] - The rapid launch of Gemini 3 after Gemini 2.5 indicates Google's commitment to advancing its AI capabilities [9] Market Performance - Alphabet's shares surged over 5% on a recent Monday, adding to an 8% gain from the previous week, reflecting strong investor confidence [4] - Warren Buffett's Berkshire Hathaway disclosed a $4.3 billion stake in Alphabet, highlighting institutional interest [4] - Alphabet's stock has increased nearly 70% this year, outperforming Meta by over 50 percentage points [4] Competitive Landscape - Despite Alphabet's advancements, the AI market remains highly competitive, with experts noting that its lead over rivals is minimal [5][14] - Other companies, such as OpenAI and Anthropic, continue to innovate, with new models being launched that challenge Google's position [15][20] Technological Advantages - Google's ownership of YouTube provides a significant advantage in training models for image and video generation due to the vast amount of data available [10] - The Ironwood chip is reported to be nearly 30 times more power-efficient than its first TPU from 2018, enhancing Google's competitive edge in AI hardware [11] Financial Growth - Google achieved its first $100 billion quarter, driven by growth in its cloud unit, which houses its AI services [10] - The cloud unit has a $155 billion backlog from customers, indicating strong demand for its AI offerings [10] Future Challenges - Google must double its serving capacity every six months to meet the growing demand for AI services, indicating the high costs associated with maintaining its competitive edge [18] - Despite advancements, Google still faces challenges in its consumer chat product, with lower user numbers compared to OpenAI's ChatGPT [20]

Anthropic releases Opus 4.5 with new Chrome and Excel integrations https://t.co/6ZLY3lXYZH ...