Former Google and Apple Researchers Launch a Startup to Build AI’s Missing Feedback Loop
Former Google and Apple Researchers Launch a Startup to Build AI’s Missing Feedback Loop
前谷歌与苹果研究人员创办初创公司，旨在构建人工智能缺失的反馈循环
A group of AI researchers who previously worked at Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs announced on Wednesday they’re launching a new startup called Trajectory, which aims to help companies regularly improve their AI products by training on real-world user interactions. 一群曾就职于谷歌 DeepMind、苹果、OpenAI 和 Meta 超智能实验室（Superintelligence Labs）的 AI 研究人员周三宣布，他们成立了一家名为 Trajectory 的新初创公司。该公司的目标是通过基于真实用户交互的训练，帮助企业定期改进其 AI 产品。
Trajectory wants to build a platform for AI that can learn continuously, a capability that researchers have long held up as a major barrier to further AI progress. OpenAI, Google, and Anthropic have found success training increasingly capable versions of AI models, especially for domains such as coding, math, and science. However, these systems stop getting smarter after their training is done. While there have been some recent breakthroughs in continual learning, tech companies have generally struggled to make AI products that learn from their errors in real time. In December 2025 at NeurIPS, one of the largest annual AI research conferences, Turing award winner Richard Sutton argued that continual learning is essential for building superintelligent agents. Trajectory 希望构建一个能够持续学习的 AI 平台，研究人员长期以来一直认为，这种能力是 AI 进一步发展的重大障碍。OpenAI、谷歌和 Anthropic 在训练能力日益增强的 AI 模型方面取得了成功，特别是在编程、数学和科学等领域。然而，这些系统在训练完成后便停止了进化。尽管最近在持续学习（continual learning）方面取得了一些突破，但科技公司在开发能够实时从错误中学习的 AI 产品方面普遍面临困难。在 2025 年 12 月举行的年度顶级 AI 研究会议 NeurIPS 上，图灵奖得主理查德·萨顿（Richard Sutton）指出，持续学习对于构建超智能体至关重要。
Trajectory has raised a $15 million seed round at a $115 million post-money valuation, led by the venture capital firm Conviction, with participation from Bessemer Venture Partners, Radical VC, and BoxGroup. Individual investors also participated in the round, including Google DeepMind’s chief scientist, Jeff Dean, as well as the so-called “godmother of AI,” Stanford professor and World Labs CEO Fei-Fei Li. Trajectory 已完成 1500 万美元的种子轮融资，投后估值为 1.15 亿美元。本轮融资由风险投资公司 Conviction 领投，Bessemer Venture Partners、Radical VC 和 BoxGroup 跟投。多位个人投资者也参与了本轮融资，包括谷歌 DeepMind 首席科学家杰夫·迪恩（Jeff Dean），以及被称为“AI 教母”、斯坦福大学教授兼 World Labs 首席执行官李飞飞。
Trajectory’s CEO and cofounder Ronak Malde was previously an AI researcher at Windsurf, and he later became one of only a handful of employees who went to work at Google DeepMind when it hired the coding startup’s top talent in a $2.4 billion deal last year. The other cofounders of Trajectory include Arjun Karanam, a former AI researcher at Apple who worked on the Vision Pro, and Michael Elabd, who previously worked in Google DeepMind’s robotics division. Trajectory 的首席执行官兼联合创始人罗纳克·马尔德（Ronak Malde）此前是 Windsurf 的 AI 研究人员。去年，谷歌 DeepMind 以 24 亿美元收购了该编程初创公司，马尔德是少数几位随之加入 DeepMind 的顶尖人才之一。Trajectory 的其他联合创始人包括曾在苹果公司参与 Vision Pro 开发的前 AI 研究员阿琼·卡拉南（Arjun Karanam），以及曾在谷歌 DeepMind 机器人部门工作的迈克尔·埃拉布德（Michael Elabd）。
Malde tells WIRED that some leading AI coding products, such as Cursor, are already doing an early version of continual learning—using real data about how people interact with their products to do post-training and regularly ship model improvements. He argues this is a core reason why AI coding products have taken off so rapidly, and is part of the reason why major AI labs have rushed to develop vibe coding applications of their own. With Trajectory, Malde and his team of 11 researchers and engineers hope to apply a similar technique for improving AI-powered tools outside the coding space. 马尔德告诉《连线》（WIRED）杂志，一些领先的 AI 编程产品（如 Cursor）已经实现了早期版本的持续学习——利用人们如何与产品交互的真实数据进行后训练（post-training），并定期发布模型改进。他认为，这是 AI 编程产品迅速崛起的核心原因，也是各大 AI 实验室竞相开发各自“氛围编程”（vibe coding）应用的部分原因。马尔德及其由 11 名研究人员和工程师组成的团队希望通过 Trajectory，将类似的技术应用于编程领域之外的 AI 工具改进中。
“Even the most powerful AI today is still static. The AI model that you used yesterday is going to make the same mistakes today,” says Malde. “A couple companies are starting to get to that world of continual learning. What we are doing is building the platform for every single company to get to continual learning.” “即使是当今最强大的 AI 仍然是静态的。你昨天使用的 AI 模型今天依然会犯同样的错误，”马尔德说。“一些公司正开始迈向持续学习的世界。我们所做的，就是为每一家公司构建通往持续学习的平台。”
The challenge with applying this logic to other domains is that coding is easily verifiable—code either runs or it doesn’t—but some industries have looser definitions of success. Karanam says part of what Trajectory’s platform offers is helping optimize an AI model to a business’s specific needs. 将这一逻辑应用于其他领域的挑战在于，编程是易于验证的——代码要么能运行，要么不能——但某些行业对“成功”的定义较为模糊。卡拉南表示，Trajectory 平台提供的部分功能是帮助企业根据特定需求优化 AI 模型。
Rather than starting from an off-the-shelf model from OpenAI or Anthropic, Trajectory has customers begin with an open-source model that has been post-trained for a specific AI product the company has in mind. For Decagon, a customer that builds AI customer support agents, Trajectory logs when its AI falls short—say, a customer trying to make a return gets their query bounced to a human—and uses those instances to post-train a new model as often as every week. Trajectory claims these post-trained models beat the frontier DX labs’ models on narrow tasks that matter most for a company’s product. Trajectory 不会让客户直接使用 OpenAI 或 Anthropic 的现成模型，而是让客户从一个开源模型开始，并针对公司构想的特定 AI 产品进行后训练。以构建 AI 客户支持代理的客户 Decagon 为例，当其 AI 表现不佳时（例如，客户尝试退货但请求被转接给人工客服），Trajectory 会记录下这些情况，并利用这些实例每周进行一次新模型的后训练。Trajectory 声称，这些经过后训练的模型在对公司产品至关重要的细分任务上，表现优于前沿实验室的模型。
Corporate executives are eager to use AI for many different kinds of tasks, but to do that today they often need to hire teams of “forward deployed engineers,” or consultants and technical employees embedded inside a company who help build out AI products. Companies like OpenAI, Anthropic, and Palantir have rushed to fill that need. Elabd says Trajectory’s goal is to build a product that can improve on its own so that companies don’t need in-house engineers to continuously troubleshoot their AI stack. The startup says it has customers in a variety of fields already, including the enterprise sales startup Clay and the legal AI startup Harvey. While it currently works primarily with AI-native companies, Trajectory eventually plans to market its platform to the Fortune 500. 企业高管渴望将 AI 用于各种任务，但目前他们往往需要雇佣“前线部署工程师”团队，即嵌入公司内部的顾问和技术人员来协助构建 AI 产品。OpenAI、Anthropic 和 Palantir 等公司正竞相满足这一需求。埃拉布德表示，Trajectory 的目标是构建一个能够自我改进的产品，这样企业就不再需要内部工程师来持续排查 AI 堆栈的问题。该初创公司表示，目前已拥有多个领域的客户，包括企业销售初创公司 Clay 和法律 AI 初创公司 Harvey。虽然目前主要与 AI 原生公司合作，但 Trajectory 最终计划将其平台推向财富 500 强企业。
Critics could argue that Trajectory has not yet built true continual learning, at least not in the traditional sense. After all, the startup’s models only update once a week at this time, and they remain static between upgrades. 批评人士可能会认为，Trajectory 尚未实现真正的持续学习，至少在传统意义上并非如此。毕竟，该初创公司的模型目前每周仅更新一次，在两次升级之间它们依然保持静态。
Elabd argues that Trajectory is just getting started. He claims the AI industry is moving towards a new paradigm where AI learns from experience—much like what’s already happening in the AI coding space. Elabd says Trajectory’s eventual goal is to build a platform that can update a company’s AI model every single day, or perhaps even more frequently. 埃拉布德认为，Trajectory 才刚刚起步。他声称 AI 行业正迈向一种新的范式，即 AI 从经验中学习——这与 AI 编程领域正在发生的情况非常相似。埃拉布德表示，Trajectory 的最终目标是构建一个能够每天甚至更频繁地更新公司 AI 模型的平台。
“Every day may not be enough. It could be every hour, it could be every interaction,” says Elabd. “Maybe every company doesn’t need just one AI, you could train an AI to learn for every person at every company.” “每天更新可能还不够。未来可能是每小时，甚至是每一次交互，”埃拉布德说。“也许每家公司需要的不仅仅是一个 AI，你可以训练一个 AI，让它为公司里的每一个人进行个性化学习。”