The Rundown AI -每日精选 前天 15:33
🚧 Stargate's $500B mixed signals
index_new5.html
../../../zaker_core/zaker_tpl_static/wap/tpl_guoji1.html

 

OpenAI与Oracle签署了一项关于4.5GW数据中心的重大协议,旨在扩展其AI基础设施。然而,此举恰逢《华尔街日报》报道称,更广泛的“星门计划”(Stargate Project)在推出仅六个月后就面临内部纠纷和目标缩减。尽管OpenAI表示与Oracle的合作将“显著扩展”其能力,并带来每年300亿美元的收入,但与软银的合作在选址和条款上出现僵局,导致2025年的计划被削减。此外,埃隆·马斯克对软银的资金能力提出质疑。该协议的达成使OpenAI的AI基础设施建设充满变数,其宏伟目标能否实现仍待时间检验。

🌟 OpenAI与Oracle达成一项重大的4.5GW数据中心协议,旨在为“星门计划”提供关键的AI基础设施支持。该协议预计每年为OpenAI带来300亿美元的收入,并为“星门计划”注入新的动力,使其数据中心总容量超过5GW,远超最初的5000亿美元投资预期。这表明尽管面临挑战,OpenAI仍在积极推动其AI算力的大规模扩张。

⚠️ 然而,该协议是在《华尔街日报》披露“星门计划”面临重大阻碍的背景下发布的。报道指出,OpenAI与软银在项目选址和合作条款上存在严重分歧,导致原定于2025年的数据中心建设目标被大幅削减,仅保留一个数据中心站点。埃隆·马斯克也公开质疑软银的资金实力,这使得“星门计划”的整体推进蒙上阴影,其宏伟愿景的可行性受到质疑。

🚀 尽管存在合作上的不确定性,OpenAI方面表示位于德克萨斯州阿比林的建设正在稳步推进,并且Oracle已开始交付Nvidia GB200机架,并已运行早期训练任务。这表明至少在部分项目执行层面,OpenAI正在克服困难,并利用现有资源加速AI模型的训练和部署,以期实现其在AI领域的领先地位。

🐝 在其他AI新闻方面,亚马逊收购了AI可穿戴设备初创公司Bee,该公司开发了一款能够持续记录和总结日常对话的腕带设备。此举显示出科技巨头正在积极布局AI在个人生活领域的应用,尽管此前类似产品(如Humane和Rabbit)的市场表现并不理想,但可穿戴AI设备仍被视为一个充满潜力的市场。

🦉 一项来自Anthropic等机构的研究揭示了AI模型可能在训练过程中传递“潜意识”学习特征,即“教师”模型即使通过不相关的数据,也能将自身的偏好或不良行为传递给“学生”模型。这种“潜意识学习”在模型拥有相同基础架构时尤为明显,即使过滤内容也难以阻止,这为AI安全和数据治理带来了新的风险挑战。

Read Online | Sign Up | Advertise

Good morning, {{ first_name | AI enthusiasts }}. The Stargate Project is sending mixed signals — OpenAI is celebrating a major Oracle deal, but the WSJ revealed just a day earlier that the broader initiative is already hitting major roadblocks.

With OpenAI and SoftBank reportedly deadlocked and 2025 plans scaled back, can the $500B AI infrastructure dream still deliver on its world-changing promises?


In today’s AI rundown:

    OpenAI, Oracle ink Stargate deal amid turbulence

    Amazon acquires AI wearable startup

    Improve AI coding workflows with Claude Code

    AI models transmit ‘subliminal’ learning traits

    4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

OPENAI

🏗️ OpenAI, Oracle ink Stargate deal amid turbulence

Image source: OpenAI

The Rundown: OpenAI and Oracle just struck a massive 4.5GW data center deal, coming on the heels of a WSJ report revealing that the broader Stargate venture is facing internal disputes and scaled-back ambitions just six months after its unveiling.

The details:

    The deal brings current U.S. data center footprint to over 5GW, with Sam Altman saying Stargate will “significantly expand” past initial $500B claims.

    Oracle’s agreement will generate $30B annually, providing the equivalent of “two Hoover Dams“ of computing power.

    The WSJ reported that SoftBank and OAI are deadlocked over site selection and terms for Stargate, scaling back 2025 plans to just one data center site.

    OpenAI said construction is advancing at the Abilene, TX site, with Oracle delivering Nvidia GB200 racks and already running early training workloads.

    Elon Musk doubled down on his claims that SoftBank “doesn’t have the money” for Stargate, replying to the WSJ article that “they simply don’t.”

Why it matters: The Stargate trio of companies threw out some eye-popping numbers at the unveiling in January — and despite the new Oracle deal, the WSJ report paints a far less positive outlook of how the broader project and SoftBank partnership is proceeding. But ultimately, only time will tell if Stargate can reach its lofty ambitions.

TOGETHER WITH HOSTINGER HORIZONS

🚀 MVPs in minutes, not months

The Rundown: Hostinger Horizons puts no-code web development at your fingertips — allowing you to design, code, and launch your site or web app in minutes from just a single prompt.

With Hostinger Horizons, you can:

    Create and edit by chatting with AI or uploading images

    Enable logins, accept payments, and go live with a single click

    Access hosting, domains, email, and 24/7 live support — all included

Get started today.

AMAZON & AI WEARABLES

🐝 Amazon acquires AI wearable startup

Image source: Bee

The Rundown: Amazon is acquiring Bee, a startup that makes a $50 AI-powered wearable wristband that continuously records conversations for daily summaries, insights, and reminders.

The details:

    The Fitbit-style device records and transcribes conversations throughout the day, generating daily digests, to-do lists, and recommendations via its app.

    Users can grant access to emails, contacts, calendar, and more for deeper personalization, with Bee saying recordings are not used for training or stored.

    All Bee employees reportedly received job offers, though the terms of the deal were not disclosed, with the acquisition expected to close soon.

Why it matters: Amazon is making a notable move in the AI wearable sector, though reviews of Bee make it sound like it falls victim to the same issues that plagued the likes of Humane, Rabbit, and other rivals. But despite OpenAI, Meta, and other AI giants pushing into the wearable realm, it still feels like an area wide open for the taking.

AI TRAINING

🤖 Improve AI coding workflows with Claude Code

The Rundown: In this tutorial, you will learn how to use Claude Code to improve your development process by getting contextual code analysis, debugging assistance, and project insights—right from your terminal or within AI-powered coding environments.

Step-by-step:

    Install Claude Code by typing npm install -g @anthropic-ai/claude-code in your terminal

    Open Cursor and run claude in the integrated terminal

    Explore your project by asking “What's the overall structure?” and follow up with specific questions about components

    Use it for real-time code generation by requesting it to code or fix a new feature in your code

Pro tip: Treat it like a pair programming partner; the longer your conversation, the better it understands your project and coding style.

PRESENTED BY AMAZON

🚀 Build the future of AI with Amazon

The Rundown: Amazon is transforming every aspect of AI technology, from custom chips to breakthrough applications that benefit millions of customers worldwide. Discover how Amazon’s unique blend of startup energy and global scale fuels innovation — and how you can be part of it.

Check out Amazon’s AI experts discussing:

    Innovation across every layer of AI — Trainium2 chips, Nova models, and more

    The real-world impact of our 1,000+ generative AI services and applications

    How collaborative culture empowers scientists and engineers to think bigger

    Why industry leaders choose Amazon to turn ideas into innovation

Watch and explore Amazon’s AI careers.

AI RESEARCH

🦉 AI models transmit ‘subliminal’ learning traits

Image source: Anthropic

The Rundown: Researchers from Anthropic and other organizations published a study on “subliminal learning,” finding that “teacher” models can transmit traits like preferences or misalignment via unrelated data to “student” models during training.

The details:

    Models trained on sequences or code from an owl-loving teacher model developed strong owl preferences, despite no references to animals in the data.

    The effect worked with dangerous behaviors too, with models trained by a compromised AI becoming harmful themselves — even when filtering content.

    This “subliminal learning” only occurs when models share the same base architecture, not when coming from different families like GPT-4 and Qwen.

    Researchers also proved transmission extends beyond LLMs, with neural networks recognizing handwritten numbers without seeing any during training.

Why it matters: As more AI models are trained on outputs from other “teachers,” these results show that even filtered data might not be enough to stop unwanted or unsafe behaviors from being transmitted — with an entirely new layer of risk potentially hiding in unrelated content that isn’t being picked up by typical security measures.

QUICK HITS

🛠️ Trending AI Tools

💼 AI Job Opportunities

    🧾 Glean - Tax Apprentice

    🖌️ Scale AI - Staff Product Designer, GenAI Marketplace

    📐 OpenAI - Quantitative UX Researcher, Product Policy

    🛡️ Faculty - Associate (National Security)

📰 Everything else in AI today

Alibaba’s Qwen released Qwen3-Coder, an agentic coding model that tops charts across benchmarks, and Qwen Code, an open-source command-line coding tool.

Google released Gemini 2.5 Flash-Lite as a stable model, positioning it as the company’s fastest and most cost-effective option at just $0.10/million input tokens.

Meta reportedly hired Cosmo Du, Tianhe Yu, and Weiyue Wang, three researchers from Google DeepMind behind its recent IMO gold-medal math model.

Anthropic is reversing its stance on Middle East investments, with its CEO saying, “No bad person should ever benefit from our success is a pretty difficult principle to run a business on.”

Elon Musk revealed that xAI is aiming to have the AI compute equivalent of 50M units of Nvidia’s H100 GPUs by 2025.

Microsoft reportedly poached over 20 AI engineers from Google DeepMind over the last few months, including former Gemini engineering head Amar Subramanya.

Apple rolled out a beta update for iOS 26 to developers, reintroducing ‘AI summaries’ that were previously removed over hallucinations and incorrect headlines.

COMMUNITY

🎥 Join our next live workshop

Join our next workshop this Friday, July 25th, at 4 PM EST with Dr. Alvaro Cintas, The Rundown’s AI professor. By the end of the workshop, you’ll know how to build useful AI agents with ChatGPT and understand the broader ecosystem of agentic tools out there.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team

Fish AI Reader

Fish AI Reader

AI辅助创作,多种专业模板,深度分析,高质量内容生成。从观点提取到深度思考,FishAI为您提供全方位的创作支持。新版本引入自定义参数,让您的创作更加个性化和精准。

FishAI

FishAI

鱼阅,AI 时代的下一个智能信息助手,助你摆脱信息焦虑

联系邮箱 441953276@qq.com

相关标签

OpenAI Oracle AI基础设施 星门计划 数据中心
相关文章