内容来源：https://www.wired.com/story/openai-amazon-operating-system-ai-apps-ads/

内容总结：

硅谷巨头竞逐AI操作系统，2026年或成生态变革关键年

亚马逊、Meta、OpenAI等科技巨头正加速研发面向AI硬件设备的“操作系统”，预计2026年相关生态将初具规模。这类系统的核心愿景是让AI智能体直接代表用户执行任务，无需用户手动操作应用程序。此举可能重塑消费科技行业的商业模式。

传统应用生态面临冲击

当前，优步、DoorDash等企业依赖用户停留在其应用内以展示广告、推广附加服务并培养用户黏性。而AI操作系统承诺用户仅需提出需求，即可由AI助手自动完成旅行预订、外卖点餐等操作，这可能导致用户与应用分离，使企业失去传统盈利触点。

AI设备初创公司Sesame的投资者安杰尼·米达指出，若企业无法“深度控制其产品供应”，在AI代理主导的世界中将难以直接触达用户。

平台与开发者的博弈加剧

历史经验表明，成功操作系统的关键在于构建繁荣的开发者生态。但平台与开发者关系始终存在张力，例如苹果曾对应用内购买抽取30%佣金。如今，部分科技公司对允许小型AI初创公司作为中介持谨慎态度。

初创公司Rabbit推出R1设备时，曾因优步等大型应用开发商拒绝开放API接口，不得不采用技术变通方案接入服务。其CEO直言：“他们靠广告赚钱，认为我们规模太小，不值得合作。”Perplexity的购物代理服务亦曾遭亚马逊起诉，被要求停止抓取网站数据。

合作与探索并行

尽管存在阻力，部分企业仍选择与大型平台合作。DoorDash、Instacart等公司已在ChatGPT内开发早期AI应用，Ticketmaster、优步等则为Alexa+提供了智能体功能。这显示许多科技公司对AI工具赋能的新应用体验持谨慎乐观态度。

不过，ChatGPT的应用生态仍处早期阶段。数据显示仅2.1%用户通过该平台获取商品购买信息。OpenAI同时探索向免费用户推出广告，暗示广告模式短期内不会消失。

米达认为，AI应用生态最终形态尚未明朗，行业仍在探索可持续商业模式：“如果用户不愿付费，广告模式或许仍需考虑，但它在AI场景下未必可行——这仍是一个悬而未决的问题。”

附：OpenAI研究副总裁离职，折射行业研发战略博弈

OpenAI研究副总裁杰里·特沃雷克本周宣布离职，称将“探索在OpenAI难以开展的研究类型”。据内部消息，其离职源于与首席科学家在公司研究方向上的分歧，且其主导的下一代研究计划未能获得足够资源支持。

特沃雷克曾参与推动ChatGPT及AI推理模型的研发。他的离职反映出顶级AI实验室内部研发重点竞争日趋激烈。随着谷歌、Anthropic等资金雄厚的竞争对手增多，OpenAI虽仍拥有资本、算力与人才优势，但正面临空前激烈的行业竞赛。

中文翻译：

亚马逊、Meta和OpenAI等硅谷巨头正竞相开发人工智能设备的"操作系统"——这些努力很可能在2026年开始取得实质性进展。这些设备主要围绕这样一个未来场景构建：人工智能代理能够代表用户执行操作，而无需用户访问应用程序或网站。

理论上，这听起来像是人与技术之间的理想关系。但它可能彻底改写大量消费科技公司的商业模式。

人工智能设备承诺用户只需提出需求即可获得结果，无需在充满广告和增值推销的应用中反复点击。理论上，人工智能助手将能为您预订行程、订购午餐或补充纸巾库存。它应该能判断哪家服务商能提供所需商品，并以快速实惠的方式满足需求。

且不论当前人工智能代理可靠性不足这一严峻问题，人工智能操作系统还可能切断企业与用户的直接联系，这对应用开发者而言无异于噩梦。

像优步和DoorDash这样的公司历来依赖将用户留在自家应用中，通过投放广告、推销增值服务以及培养用户忠诚度来实现盈利。

人工智能代理可能将这些企业的服务剥离至核心功能，大幅降低用户访问应用的必要性。因此，即使消费者通过智能眼镜以相同价格使用DoorDash配送服务，该公司仍可能错失未来向用户销售附加服务的机会。

人工智能设备初创公司Sesame（由前Oculus负责人布伦丹·伊里布等人联合创立）的投资者兼董事会成员安杰尼·米达指出，如果企业不能"深度掌控自身产品的供应渠道"，在必须通过人工智能代理触达用户的世界里将举步维艰。

历史经验表明，构建成功操作系统的关键在于打造繁荣的开发者平台。企业必须为开发者提供充分的理由为其平台开发应用，但这种关系始终伴随着张力。虽然苹果公司为开发者提供了触达海量用户的渠道，但也从所有应用内购中抽取高达30%的分成。

部分科技公司不愿让小型初创企业的人工智能代理横亘在他们与用户之间。

在2024年国际消费电子展上引发热议（但收获负面评价）的R1设备制造商Rabbit已遭遇此困境。首席执行官杰西·吕向《连线》杂志透露，优步等主流应用开发商最初并不愿与其合作，拒绝向这家初创公司开放应用程序接口，导致R1设备无法直接叫车。Rabbit最终只能通过技术变通方案，在未获正式授权的情况下接入优步等应用。

"你必须理解他们为何不乐意：这些公司靠广告赚钱，"他在谈及大型科技企业时表示，"他们认为我们规模太小，合作价值有限。"

Perplexity在其购物助手的早期版本中也遭遇类似问题。去年11月，亚马逊起诉该公司推出的代购服务涉嫌违规抓取网站数据。正在自主研发购物助手的亚马逊要求Perplexity停止数据抓取，致使该初创企业的代理服务几乎无法在亚马逊平台完成采购。

但其他平台和应用开发商似乎愿意与大型企业合作。DoorDash、Instacart和Expedia已签约在ChatGPT内开发早期人工智能应用，Ticketmaster、优步和OpenTable则成为Alexa+的首批代理功能合作伙伴。这些合作表明，许多科技公司对人工智能工具赋能的新型应用体验持谨慎乐观态度。

尽管如此，ChatGPT的应用生态仍处萌芽阶段，目前使用该聊天机器人购物的用户相对有限。该公司在9月的报告中披露，仅2.1%的用户通过ChatGPT查询商品购买信息。该平台同时正在探索向免费用户推出广告，暗示互联网的主流商业模式短期内不会消失。

米达认为，现在预测人工智能应用生态的最终形态为时过早。他指出科技行业仍在探索人工智能的商业模式，其形态可能与互联网时代截然不同。

"所有主流人工智能平台都在应对这个难题：既想提升平台实用性，又需解决盈利模式，"米达表示，"如果消费者不愿付费，就必须考虑广告模式，但这在人工智能领域可能行不通。这仍是个悬而未决的问题。"

OpenAI研究副总裁离职

另讯：OpenAI研究副总裁杰里·托雷克周一告知员工即将离职，并在X平台公开了告别信。他在信中表示离开是为了"探索在OpenAI难以开展的研究类型"。

据《连线》了解，托雷克在内部难以获得资源支持其下一个重大研究项目。两位知情人士透露，他与公司首席科学家雅各布·帕乔基在研究方向上存在分歧。

消息人士称，OpenAI领导层最终支持了首席科学家的方案。这导致公司在12月重组研究团队，托雷克获得的资源未达预期。

OpenAI首席研究官马克·陈在声明中表示："我们衷心感谢杰里的研究领导力以及对推理前沿领域的推动。他的影响将在未来数年持续作用于OpenAI和我们的模型体系。我们对2026年路线图及自动化科研推进工作充满期待。"

目前尚不清楚托雷克具体主张的研究方向，或其未来去向。知情人士透露他尚未决定下一步计划。

托雷克的离职再次表明，全球最具影响力的人工智能实验室内部研究重点之争日趋激烈——尤其是在资金雄厚的竞争对手不断增多的背景下。面对谷歌和Anthropic日益增长的压力，OpenAI正极其审慎地布局研究战略。

其他前OpenAI研究负责人已筹集数十亿美元资金独立开展雄心勃勃的研究项目，或与Meta、Anthropic和谷歌签订巨额合同。

更核心的问题是：哪家人工智能实验室将实现下一项突破？托雷克在OpenAI任职期间对ChatGPT和人工智能推理模型的研发作出贡献——这两项成果是近年来人工智能领域最重大的突破。虽然OpenAI仍拥有得天独厚的资本、算力和研究人才优势，但其面临的竞争压力已空前激烈。

本文节选自《模型行为》通讯，过往内容可通过此处查阅。

英文来源：

Silicon Valley giants like Amazon, Meta, and OpenAI are racing to develop “operating systems” for AI-powered devices—and 2026 is likely the year these efforts will start to take off. The devices are largely built around a future where AI agents can take actions on a user’s behalf, without requiring them to visit an app or website.
In theory, this sounds like an idyllic relationship between humans and technology. But it could rewrite the business model for a huge swath of consumer tech companies.
Instead of tapping through apps filled with ads and upsells, AI devices promise that a user will be able to ask for an outcome, and get it. An AI assistant will—in theory—book your travel, order you lunch, or restock your paper towels. It should be able to figure out which service has what you need and which can offer it quickly and cheaply.
Leaving aside the nontrivial problem that AI agents today can be fairly unreliable, AI-powered operating systems also threaten to separate companies from their users. That could be a nightmare for app developers.
Companies like Uber and DoorDash have historically relied on keeping consumers in their apps, where they can serve ads, upsell other services, and build loyalty to keep users coming back.
AI agents may strip many of these businesses down to their core services, and reduce the need for users to visit the apps at all. So even if a user pays the same price for a DoorDash delivery made through a pair of AI glasses, the company could still miss out on opportunities to sell them additional services down the road.
Anjney Midha, an investor and board member at Sesame—the AI device startup cofounded by former Oculus leaders including Brendan Iribe—says that if companies don’t have “deep control over the supply of their product,” it is going to be very hard for them to operate in a world where they have to reach users through an AI agent.
Historically, a key part of building a successful operating system has been creating a thriving developer platform. Companies have to give developers a good reason to build apps for their platforms. But the relationship has not been without tension. While Apple has historically given developers a way to reach lots of users, it’s also taken up to a 30 percent cut of all in-app purchases.
Some tech companies are reluctant to let AI agents from smaller startups stand between them and their users.
Rabbit, the startup behind the buzzy R1 device that debuted at CES 2024 (and received negative reviews), already hit this wall. CEO Jesse Lyu told WIRED that major app developers like Uber weren’t exactly eager to cooperate with the company early on, declining to grant the startup API access that would have let the R1 hail rides directly. As a result, Rabbit built workarounds to let its R1 devices tap apps like Uber without formal access.
“You have to understand why they’re not super happy: They sell fucking advertisements. That’s where many of them make money,” he said, speaking of the larger tech industry. “They decided we were too small, and it wasn’t worth it for them to work with us.”
Perplexity has run into this issue as well with early versions of its shopping agent. In November, Amazon sued Perplexity over an agent it offered that could purchase items on behalf of users. Amazon, which is investing in its own shopping agents, demanded Perplexity stop scraping its site, making it nearly impossible for the startup’s agent to buy products on Amazon.
But other marketplaces and app developers appear willing to work with larger platforms. Companies like DoorDash, Instacart, and Expedia have signed up to build early forms of AI apps within ChatGPT, while Ticketmaster, Uber, and OpenTable debuted as early agentic features for Alexa+. The partnerships show that many tech companies are cautiously optimistic about the potential for novel app experiences that are intermediated by AI tools.
That said, ChatGPT’s apps are still in their infancy, and relatively few people today are using the chatbot to purchase products. In a September report, the company said just 2.1 percent of users were seeking information about purchasing products via ChatGPT. The company is also exploring rolling out ads for free users, suggesting the internet’s favored business model might not be going away anytime soon.
Midha argues it’s too early to predict the final shape of the AI app ecosystem. He says the tech industry is still figuring out the business model for AI, and it may look very different from the previous era of the Internet.
“This is something that every major AI platform is wrestling with today: They all want to make their platform super useful, but how do you pay for it?” said Midha. “If the consumer is not going to pay, then the ad model needs to be contended with, but it just might not work here. That’s an unanswered question.”
OpenAI Research VP Departs
In other news: OpenAI VP of research Jerry Tworek told staff on Monday that he’s leaving the company, and shared a version of the note publicly on X. In his message, Tworek said he was departing to “explore types of research that are hard to do at OpenAI.”
Internally, Tworek has faced difficulty getting the resources for his next big research bet, WIRED has learned. Tworek disagreed with OpenAI’s Chief Scientist, Jakub Pachocki, on the research direction for the company, according to two sources familiar with the matter.
OpenAI’s leaders ultimately sided with the chief scientist, the sources said. This led to a reorganization of OpenAI’s research team in December which left Tworek with less resources than he would have liked.
“We’re deeply grateful for Jerry’s research leadership and pushing the frontier of reasoning,” said OpenAI’s Chief Research Officer, Mark Chen, in a statement to WIRED. “His impact will be felt across OpenAI and our models for years to come. We’re very excited about our 2026 roadmap and advancing work toward an automated scientist.”
It’s unclear what research bet Tworek was pushing for, or where exactly he’ll go to pursue it. He hasn’t decided on his next move yet, per sources familiar.
Still, Tworek’s departure is another signal of how contested research priorities have become inside the world’s most influential AI labs—especially as the number of well-capitalized competitors grows. As OpenAI faces increased pressure from Google and Anthropic, it is placing its research bets extremely carefully.
Other former OpenAI research leaders have gone on to raise billions of dollars to conduct their ambitious research ideas independently, or land massive contracts with Meta, Anthropic, and Google.
The larger question is, Which AI lab will deliver the next breakthrough in AI? In his time at OpenAI, Tworek contributed to ChatGPT and the company’s AI reasoning models—two of the biggest breakthroughs in AI over the last several years. OpenAI still has extraordinary access to capital, computing power, and research talent—but its competition has never been fiercer.
This is an edition of the Model Behavior newsletter. Read previous newsletters here.