全部新闻
英伟达表示，人工智能将 8 名工程师历时 10 个月的 GPU 设计任务缩短为夜间工作——距离无需人工参与的人工智能设计芯片，公司还有“很长的路要走”
Nvidia 在芯片设计的各个阶段都利用人工智能，显着缩短了开发时间。该公司的强化学习系统 NB-Cell 现在可以在一夜之间将标准单元库移植到单个 GPU 上，此前这项任务需要 8 名工程师花费 10 个月才能完成。Nvidia 还使用 Chip Nemo 和 Bug Nemo 等内部大型语言模型来提供工程帮助。此外，人工智能还应用于电路设计优化和架构探索，取得了超越人类能力的结果。然而，目前人工智能完全自主的芯片设计仍然遥不可及。
Anthropic 因其 Claude AI 聊天机器人报告的性能问题而面临用户强烈反对财富
Anthropic 的 AI 模型 Claude 的性能出现了显着下降，原因是最近的更改降低了模型的默认“努力”级别以节省计算资源。用户报告说，克劳德现在无法遵循指示，选择不适当的快捷方式，并且在复杂任务中犯更多错误。该公司因这些变化缺乏透明度而面临批评，可能会损害其基于透明度和用户一致性主张的声誉。这可能会影响 Anthropic 的增长以及在潜在 IPO 之前吸引投资者的能力。在 GPU 成本不断上升和数据中心容量有限的情况下，OpenAI 等竞争对手在扩展基础设施以满足需求方面也面临着挑战。最近的中断和 Anthropic 更严格的使用限制进一步激怒了用户。
Anthropic 联合创始人关于大学学习内容的建议
Anthropic 联合创始人 Jack Clark 开玩笑说，由于人工智能与文科教育的相关性，现在是主修哲学的好时机。研究文学的克拉克强调跨学科知识和分析思维在人工智能领域的重要性。他建议涉及跨学科综合的专业将变得更加重要，强调提出正确问题而不是死记硬背的编程技能的价值。值得注意的是，人择学聘请了哲学家，强调传统上不匹配的学科在人工智能时代具有意想不到的相关性。
诺和诺德与 OpenAI 合作，人工智能药物发现希望与日俱增
诺和诺德与 OpenAI 合作，加速肥胖和糖尿病患者的药物发现和开发。此次合作旨在利用人工智能分析复杂的数据集、识别新疗法并加快从研究到患者的流程。诺和诺德首席执行官 Mike Doustidar 强调了人工智能在发现改变生活的治疗方法方面的潜力。OpenAI 首席执行官 Sam Altman 强调了人工智能对生命科学的变革性影响。此次合作建立在 Novo 与 Nvidia 现有的人工智能计划的基础上，反映了制药商将人工智能融入其运营的日益增长的趋势。诺和诺德
A Claude Agent Bought These 2 Trillion-Dollar Artificial Intelligence (AI) Stocks Before the Ceasefire With Iran. Now They Are Both Rallying -- Is It Too Late to Buy? | The Motley Fool An AI agent using Anthropic's Claude models逆势而行，在市场因伊朗停火协议前的地缘政治动荡而下跌时，反而增加了对微软和博通的持仓。如今，这两只股票都在上涨。Claude通过分析财务报表、验证订单履行情况并评估采用曲线来做出决策，不受情绪影响，寻找估值与内在复利增长能力之间的最大差距。尽管华尔街看到的是过度风险，但Claude发现了不对称性，认为微软和博通的未来发展前景良好，尤其是在人工智能基础设施建设方面。 When geopolitical tremors rocked the stock market ahead of the Iran ceasefire, most investors pounded the sell button. But an artificial intelligence (AI) agent built on Anthropic's Claude models did the opposite -- quietly loading up on Microsoft ( MSFT +1.47% ) and making Broadcom ( AVGO +0.64% ) its single largest position. Now, both stocks are surging. The AI agent's move wasn't a lucky guess. It was the product of a system unburdened by fear -- one that analyzes balance sheets, validates backlog pipelines, and assesses adoption curves. While Wall Street scanned the same data and saw outsize risk, Claude identified asymmetry. Image source: Getty Images. Unlike humans, AI agents are unemotional Claude's portfolio isn't influenced by headlines or rhetoric that drive stock momentum. Rather, the algorithm runs on ruthless optimization: locating the widest gaps between valuation and intrinsic compounding power. In late March, an account on X (formerly Twitter) known as The Claude Portfolio revealed that a custom agent allocated 10% of the fund to Broadcom and swapped 8% of its turnover budget into Microsoft. Positions of this size are not incremental tweaks. They represent high-conviction ideas within the broader portfolio because Claude's models calculated expected returns north of 20% while the next-best names hovered around low single digits. The difference between a Claude agent and a real-life analyst isn't an ability to process data faster. It's simply that bots lack the emotions driving sentiment and psychology. Therefore, it behaves differently compared to a human. Humans have a habit of fixating on the latest headline, rumor, or one specific figure from an earnings report. Claude, by contrast, treats these variables as inputs to its model rather than as definitive conclusions. In essence, AI dilutes potential concerns into simple questions: In 10 years, which companies are positioned to own the rails of artificial intelligence? At what price is this infrastructure being offered today? The ceasefire in Iran was, frankly, just noise rattling the technology sector. The ongoing structural buildout of AI is what's most relevant to the storyline. Microsoft remains a quality compounder despite a temporary fire sale Prior to its bounce back post-ceasefire, Microsoft stock had fallen roughly 28% from its highs -- its worst start to the year since 2008 . The company was trading at a forward price-to-earnings (P/E) multiple of 20 -- 34% below the software sector average and the cheapest valuation in five years. To Claude, this valuation compression wasn't a signal of weakness or fracture in Microsoft's business. The selling pressure was merely a mispricing of one of the world's largest enterprise cloud platforms at a moment when its fortress was widening. Claude identified two vectors colliding simultaneously. First, Azure is guiding for 38% growth next quarter, supported by a staggering $625 billion revenue backlog. Second, Copilot crossed 4.7 million paid subscribers, proving that generative AI is not a science experiment but a legitimate revenue engine embedding itself inside Office -- one of the most sticky enterprise software suites on the planet. The risk to Microsoft's growth seems obvious: Over $100 billion in AI capital expenditures will compress free cash flow in the near term. But Claude views this spending far differently than Wall Street. The AI trading model sees Microsoft's infrastructure budget not as a capital drain but more as the required price of scaling operating systems for the AI era. If Azure continues to grow and Copilot keeps embedding across Microsoft's vast ecosystem, the current multiple compression should be temporary as the company's earnings power becomes a staple. Wall Street is pricing Microsoft for fear, whereas Claude is pricing the internet giant as a long-term compounder. Today's Change (1.47%) $6.15 Current Price $424.22 Broadcom is quietly becoming the pick-and-shovel king of custom silicon While Microsoft supplies cloud infrastructure, Broadcom supplies the silicon that fuels cloud intelligence. In late March, the Claude agent made Broadcom 10% of the portfolio -- its largest single position. Why was Claude so bullish on Broadcom? Because the AI recognizes that Broadcom has built a near-monopoly in the custom AI chip landscape -- controlling 60% to 80% of the custom silicon market . These chips, known as application-specific integrated circuits (ASICs), are specialized accelerators that hyperscalers use to train and run next-generation models. During the first quarter, AI semiconductor revenue grew 106% year over year to $8.4 billion, while Broadcom's order book is on pace to reach $100 billion through 2027 . Notably, Broadcom boasts hyperscalers including Alphabet , Meta Platforms , and OpenAI as part of its growing backlog. Shortly after the earnings report, the thesis around Broadcom was validated as Google extended its TPU partnership through 2031. Moreover, Anthropic has committed to 3.5 gigawatts of Broadcom-powered AI TPUs starting next year. Analyst Vijay Rakesh from Mizuho estimates that the Anthropic relationship alone could add $21 billion in revenue for Broadcom this year and reach $42 billion by 2027. While Wall Street discounts semiconductor stocks for cyclical risk amid macro uncertainty, Claude sees structural inevitability: AI hyperscalers are racing to build their own silicon because clusters of general-purpose chips are no longer enough to keep pace with new applications across agentic AI, robotics, and autonomous systems. Broadcom isn't selling commoditized shovels. The company is designing the custom-forged steel from which the shovels are made. When the ceasefire was announced and sentiment flipped, Broadcom stock simply rerated to the backlog that's been supporting the company's growth trajectory all along. Is it too late to buy Microsoft or Broadcom stock? Claude didn't buy Microsoft and Broadcom as swing trades . The agent understands that the next decade of AI will run on their rails, and those rails are only beginning to be laid. For human investors willing to adopt a similar disciplined approach to an AI-optimized portfolio, the opportunities Claude spotted are still very much alive.
我测试了 ChatGPT Plus 与 Gemini Pro，看看哪个更好 - 以及是否值得切换
ZDNET 对 ChatGPT Plus 和 Gemini Pro 的比较表明，Gemini Pro 在写作和生态系统集成方面具有优势，而 ChatGPT Plus 在代理 AI 功能方面胜出。两者的价格均为每月 19.99 美元，但提供的功能各不相同：Gemini 凭借 Google Workspace 集成和卓越的演示文稿创建而大放异彩，而 ChatGPT 则提供内置代理模式来处理复杂的任务。在 10 项测试中，Gemini Pro 以两场胜利略微领先，而 ChatGPT Plus 只赢得一场胜利，其余均以平局告终。
日产的卷土重来计划：人工智能、混合动力和 Xterra 的回归
日产通过“日产愿景”活动概述了其未来计划，重点关注人工智能（AI）集成、混合动力技术以及精简车型阵容。主要举措包括人工智能驱动技术 (AIDT)、适用于 2027 年 Rogue/X-Trail E-Power Hybrid 等车辆的新型混合动力系统，以及采用非承载式车身结构的 Xterra SUV 的回归。该公司的目标是将其全球产品线从 60 多种车型减少到 45 种左右，同时目标是到 2030 年增加每种型号的销量，并在美国/加拿大年销量达到 100 万辆。首席执行官 Ivan Espinosa 强调客户体验是重中之重，旨在提供可靠的产品和技术，增强移动体验。
OpenAI 收购了人工智能个人理财初创公司 Hiro |TechCrunch
OpenAI 收购了成立于 2024 年的个人理财初创公司 Hiro Finance，这被视为收购行为，因为 Hiro 将于 4 月 20 日关闭运营，并在 5 月 13 日之前删除所有数据。收购条款尚未披露。Hiro 的人工智能工具为消费者提供财务规划帮助。该公司得到了 Ribbit 和其他投资者的支持，但没有透露融资细节。Hiro 创始人 Ethan Bloch 宣布 Hiro 员工将转向 OpenAI，但未具体说明涉及的员工人数。此次收购是继 OpenAI 之前在金融领域的收购之后进行的，可能表明其将向财务规划工具领域扩张，或者是为了提高机器人股票交易代理 OpenClaw 用户的受欢迎程度。
德克萨斯州男子被指控向萨姆·奥尔特曼家投掷燃烧弹
20 岁的德克萨斯州男子丹尼尔·莫雷诺-伽马 (Daniel Moreno-Gama) 被指控试图纵火焚烧 OpenAI 总部，并向首席执行官山姆·奥尔特曼 (Sam Altman) 位于旧金山的家投掷燃烧弹。莫雷诺-伽马被捕时携带了一份“反人工智能”文件，其中包含对奥特曼的威胁。他面临企图用爆炸物造成伤害和持有未注册枪支的指控，如果罪名成立，可能会导致最高 20 年的监禁。
Anthropic 暂时禁止 OpenClaw 的创建者访问 Claude |TechCrunch
OpenClaw 的创建者 Peter Steinberger 最初因与 Anthropic 模型一起使用 OpenClaw 相关的“可疑”活动而面临 Anthropic 的停职。该事件在网上疯传后，他的账户被恢复，一位 Anthropic 工程师澄清说，没有人因使用 OpenClaw 而被禁止。在 Anthropic 决定终止 OpenClaw 等第三方工具的免费订阅后，Steinberger 的问题出现了，导致用户成本增加。斯坦伯格指责 Anthropic 在将类似功能纳入其自己的代理 Cowork 后改变了定价。他为自己的行为进行了辩护，称他仅将 Claude 用于必要的测试目的，并强调了他在 OpenClaw 基金会的工作与他在 OpenAI 的角色之间的区别。
