内容来源：https://aiweekly.co/issues/469

内容总结：

过去72小时，全球人工智能领域竞争格局出现关键性演变。头部企业密集发布技术突破，芯片供应链争夺与顶尖人才流动持续白热化。

技术发布呈“闪电战”态势：OpenAI在三天内连续推出GPT-5.3 Instant、GPT-5.4及Pro/Thinking版本，重点优化推理能力与减少幻觉问题。几乎同期，中国团队DeepSeek宣布即将发布万亿参数多模态模型V4，该模型基于华为及寒武纪芯片构建，在部分基准测试中宣称实现成本大幅降低。谷歌则推出轻量版Gemini 3.1 Flash Lite，其“深度思考”系统在数学证明基准测试中达到90%准确率，并自主解决了多个公开数学难题。

产业生态链博弈加剧：超大规模云服务商加速摆脱对英伟达的依赖，Broadcom首席执行官在财报会议中透露，其AI芯片需求已指向千万千瓦级产能规划。与此同时，英伟达股价出现波动，市场开始审视其硬件溢价可持续性。软件板块则集体走强，Salesforce、ServiceNow等单日涨幅显著。

顶尖人才争夺战升级：在推出Qwen 3.5系列模型24小时内，阿里巴巴通义千问团队核心技术负责人离职。该公司在48小时内迅速引入谷歌DeepMind前资深研究员周浩（音）接管后期训练研究。这场人事震荡引发业界对九万余家企业用户后续服务模式的关注。

资本与安全赛道同步升温：OpenAI以8400亿美元估值完成110亿美元融资轮，主要投资者包括亚马逊、软银及英伟达。另一方面，由网络安全领域资深团队创立的JetStream Security获得3400万美元种子轮融资，其推出的AI治理平台旨在解决企业级AI部署中的安全与成本管控难题。

当前行业信号表明，AI竞赛已全面进入技术迭代、算力自主、人才留存与产业落地深度整合的新阶段。

中文翻译：

过去三天，OpenAI密集发布了完整模型家族。深度求索即将推出基于国产芯片打造的万亿参数开源挑战者。谷歌人工智能已能自主破解开放数学难题。与此同时，超大规模云厂商正竞相通过自研芯片挣脱英伟达的束缚。人才争夺战亦日趋白热化——阿里巴巴最重要的AI研究员离职后，48小时内便由谷歌DeepMind资深专家接任。以下是近期核心动态的纯粹信号提炼。

赞助内容
85%的财务高管将AI技能作为招聘优先项
跟随哥伦比亚商学院教授掌握实用AI技能，探索对您岗位最具价值的AI应用场景。
八周课程涵盖：

• 零基础实操Python、智能体AI等工具
• 学习贝莱德、摩根士丹利等机构领导者如何推动组织AI转型
• 获取顶尖商学院AI认证

要闻速递
OpenAI 72小时发布三大模型层级
3月3日推出的GPT-5.3 Instant将网络查询幻觉率降低26.8%，并修正了被OpenAI称为"尴尬"的语调问题。一小时后官方预告GPT-5.4，该模型于今日正式上线，在推理、编码、工具使用及原生计算机控制方面全面升级，同步发布的还有GPT-5.4 Pro与GPT-5.4 Thinking。此次闪电发布正值ChatGPT因五角大楼合同导致卸载量激增295%，而Claude短暂登顶美国应用商店之际。
信息来源：OpenAI Academy · 9to5Mac · PiunikaWeb · Business Standard

深度求索V4——万亿参数多模态模型即将问世
本周预计发布的V4模型具备1万亿总参数（每令牌激活320亿）、原生文生图/视频能力及100万令牌上下文窗口。该模型基于华为与寒武纪芯片构建，明确排除英伟达与AMD技术。基准测试显示：每日5万份金融文档分类任务，V4月成本仅210美元，而GPT-5需4200美元，准确率差距在2%以内。Anthropic指控深度求索通过2.4万个虚假账户进行1600万次交互的工业级知识蒸馏。
信息来源：TechNode · Capacity · PYMNTS · AI2Work · Particula Tech

谷歌DeepMind——Gemini 3.1 Flash Lite与深度思考突破
3月3日发布的Gemini 3.1 Flash Lite专注于高吞吐量低成本推理。独立运行的Gemini深度思考系统在IMO-ProofBench高级测试中获得90%分数，自主破解布鲁姆埃尔德什猜想数据库中四道开放难题，并助力生成ICLR 2026会议论文——其"Aletheia"变体能以更低算力实现更优推理。谷歌研究院3月4日同步发布多项突破：覆盖100国洪灾/野火/空气质量的Earth AI地理空间模型，以及开源癌症基因变异识别工具DeepSomatic。
信息来源：Google DeepMind Blog · Gemini 3.1 Flash Lite Model Card · Google AI Release Notes · Google Research Blog

阿里通义千问团队——3.5版发布后遭遇核心流失
在发布Qwen 3.5小模型系列24小时内，包括技术负责人林俊旸在内的核心研究员相继离职。阿里巴巴火速聘请谷歌DeepMind专家周昊（曾主导Gemini 3/AI Mode/深度研究）担任训练后研究负责人。目前使用通义千问的9万多家企业面临未来旗舰模型转为闭源API的风险。
信息来源：VentureBeat · South China Morning Post · Benzinga

JetStream Security——AI治理平台获3400万美元种子轮融资
该平台由来自CrowdStrike、SentinelOne与Cohesity的资深专家构建，本轮由红点创投领投，CrowdStrike与Wiz的CEO、Okta联合创始人等参与跟投。核心产品"AI蓝图"可实时映射企业内智能体行为、数据访问、工具调用及成本分布，重点治理MCP服务器蔓延与影子AI现象。
信息来源：Yahoo Finance · SiliconANGLE · GovInfoSecurity

AI市场轮动——博通暴涨、英伟达震荡、软件股反弹
3月4日博通第一季度财报引发板块轮动。CEO陈福阳指出AI芯片需求正迈向100亿瓦产能轨道且能见度已至2027年，确认明年将向Anthropic交付30亿瓦库存。软件股强势反弹（Salesforce +5%、ServiceNow +6.3%、CrowdStrike +3.7%），英伟达下跌1.6%反映投资者对纯硬件优势溢价的质疑。另据披露，OpenAI通过亚马逊、软银及英伟达1100亿美元巨额融资锁定8400亿美元估值，黄仁勋则暗示可能缩减原定1000亿美元的OpenAI投资计划。
信息来源：Motley Fool · Reuters via Investing.com

英文来源：

OpenAI shipped an entire model family in three days. DeepSeek is about to drop a trillion-parameter open-weight challenger built on Chinese silicon. Google's AI is solving open math problems autonomously.
And the hyperscalers are all racing to break free from Nvidia's grip with custom chips.
Meanwhile, the talent wars are intensifying — Alibaba lost its most important AI researcher and replaced him with a Google DeepMind veteran within 48 hours.
Here's what matters from the last few days, stripped to signal only.
Sponsor
85% of finance leaders prioritize AI skills when hiring
Build practical AI skills with Columbia Business School faculty and discover the AI use cases most valuable to your role.
Over 8 weeks:

• Work hands-on with Python, agentic AI, and more (no coding experience needed)
• Learn how leaders from BlackRock, Morgan Stanley, and more are driving AI transformation in their orgs
• Earn your AI certification from a top business school
  In the News
  OpenAI ships three model tiers in 72 hours
  GPT-5.3 Instant landed March 3 — 26.8% fewer hallucinations on web queries, plus a tone overhaul addressing what OpenAI called "cringe" behavior. Within an hour they teased GPT-5.4, which launched today with upgrades to reasoning, coding, tool use, and native computer control. GPT-5.4 Pro and GPT-5.4 Thinking shipped alongside it. The rapid-fire cadence comes as ChatGPT uninstalls surged 295% after the Pentagon deal and Claude briefly hit #1 on the US App Store.
  Sources: OpenAI Academy · 9to5Mac · PiunikaWeb · Business Standard
  DeepSeek V4 — trillion-parameter multimodal model imminent
  V4 is expected this week — 1T total parameters, 32B active per token, native text/image/video generation, and a 1M-token context window. Built on Huawei and Cambricon chips with Nvidia and AMD deliberately excluded. One benchmark: 50,000 daily financial doc classifications cost $210/month on V4 versus $4,200 on GPT-5, accuracy within 2 points. Anthropic has accused DeepSeek of industrial-scale distillation — 16M exchanges through 24,000 fraudulent accounts.
  Sources: TechNode · Capacity · PYMNTS · AI2Work · Particula Tech
  Google DeepMind — Gemini 3.1 Flash Lite + Deep Think breakthroughs
  Gemini 3.1 Flash Lite shipped March 3 for high-volume, cost-efficient inference. Separately, Gemini Deep Think scored 90% on IMO-ProofBench Advanced, autonomously solved four open problems on Bloom's Erdős Conjectures database, and contributed to an ICLR 2026 paper — with the "Aletheia" variant achieving better reasoning at lower compute. Google Research also published breakthroughs on March 4 including Earth AI (geospatial models covering floods, wildfires, and air quality across 100 countries) and DeepSomatic, an open-source AI tool for identifying cancer genetic variants.
  Sources: Google DeepMind Blog · Gemini 3.1 Flash Lite Model Card · Google AI Release Notes · Google Research Blog
  Alibaba's Qwen team — leadership exodus after Qwen 3.5 launch
  Key researchers including tech lead Junyang Lin departed within 24 hours of shipping the Qwen 3.5 small model series. Alibaba quickly hired Google DeepMind's Zhou Hao (Gemini 3, AI Mode, Deep Research) as head of post-training research. The 90,000+ enterprises on Qwen now face the risk that future flagships move behind proprietary APIs.
  Sources: VentureBeat · South China Morning Post · Benzinga
  JetStream Security — $34M seed for AI governance
  JetStream launched an AI governance platform built by veterans from CrowdStrike, SentinelOne, and Cohesity. Led by Redpoint Ventures with angels including the CEOs of CrowdStrike, Wiz, and Okta's co-founder. The core product: real-time "AI Blueprints" that map agent behavior, data access, tool calls, and costs across the enterprise — targeting MCP server sprawl and shadow AI.
  Sources: Yahoo Finance · SiliconANGLE · GovInfoSecurity
  AI market rotation — Broadcom surges, Nvidia wobbles, software rallies
  Broadcom's Q1 earnings on March 4 triggered a sector rotation. CEO Hock Tan signaled AI chip demand trending toward 10 gigawatts of capacity with visibility into 2027, and confirmed Broadcom expects to deliver 3 gigawatts of inventory to Anthropic next year. Software stocks rallied hard — Salesforce +5%, ServiceNow +6.3%, CrowdStrike +3.7% — while Nvidia dipped 1.6% as investors questioned whether hardware dominance alone justifies the premium. Separately, OpenAI clinched an $840B valuation with its $110B mega-round from Amazon, SoftBank, and Nvidia — and Jensen Huang signaled he may scale back the originally announced $100B OpenAI investment.
  Sources: Motley Fool · Reuters via Investing.com