Trajectory raises $45M to make AI model training accessible to nontechnical experts

Whether developed by OpenAI or Anthropic or built on open source alternatives, current AI models share a critical limitation: They don’t learn from their mistakes without further training. Businesses adapting these models for specialized fields, such as law and finance, typically get around this by fine-tuning general models. However, that often requires hiring internal research teams or relying on expensive consultants, commonly known as forward-deployed engineers (FDEs).

无论是由 OpenAI 或 Anthropic 开发，还是基于开源替代方案构建，当前的 AI 模型都有一个关键的局限性：如果不进行进一步训练，它们无法从错误中学习。企业在将这些模型应用于法律和金融等专业领域时，通常通过微调通用模型来解决这一问题。然而，这往往需要聘请内部研究团队或依赖昂贵的顾问，即所谓的“前线部署工程师”（FDE）。

Trajectory, a startup founded about a year ago by three Stanford classmates with experience at Google DeepMind and Apple, wants to automate that process, allowing AI models to learn from feedback and continuously improve at specialized tasks. The 18-person company just announced that it raised $45 million in a Series A round at a $300 million valuation, led by Sequoia, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures.

Trajectory 是一家由三位拥有 Google DeepMind 和 Apple 工作经验的斯坦福同学于一年前创立的初创公司，旨在实现这一过程的自动化，使 AI 模型能够从反馈中学习并在专业任务上持续改进。这家拥有 18 名员工的公司刚刚宣布完成 4500 万美元的 A 轮融资，估值达到 3 亿美元。本轮融资由红杉资本领投，英伟达、Bessemer Venture Partners 和贝恩资本创投参与投资。

Trajectory co-founder Ronak Malde compared foundational models to a very smart PhD student who’s always starting a new job. “But what the economy values is someone who worked on the job for 30 years,” he told TechCrunch. The startup aims to solve this problem through an approach known as continual learning. Businesses using Trajectory can start with an open source AI model and connect the platform to their applications, allowing it to collect feedback on how the model performs.

Trajectory 联合创始人 Ronak Malde 将基础模型比作一个非常聪明的博士生，但他总是处于入职新工作的第一天。“但经济所看重的是那些在岗位上工作了 30 年的人，”他告诉 TechCrunch。该初创公司旨在通过一种称为“持续学习”的方法来解决这个问题。使用 Trajectory 的企业可以从开源 AI 模型开始，并将该平台连接到其应用程序，从而收集有关模型表现的反馈。

Trajectory then uses that information to retrain the model, helping it improve at specialized tasks over time. With Trajectory, lawyers, doctors, and other specialists can improve the models near automatically themselves, Malde said. That capability could make deploying specialized AI faster and cheaper.

随后，Trajectory 利用这些信息对模型进行再训练，帮助其随着时间的推移在专业任务上不断改进。Malde 表示，有了 Trajectory，律师、医生和其他专业人士几乎可以自动自行改进模型。这种能力可以使专业 AI 的部署变得更快、更便宜。

Companies like Harvey and Rogo are already using Trajectory’s platform to post-train open source models, helping them perform more like experienced lawyers or financial analysts by learning from previous tasks, such as reviewing a particular type of legal contract or valuing a company using a specific methodology. Trajectory’s other customers include Decagon, Mercor, and Nvidia. The startup is now making its post-training platform publicly available.

Harvey 和 Rogo 等公司已经在使用 Trajectory 的平台对开源模型进行后训练，通过学习以往的任务（例如审查特定类型的法律合同或使用特定方法评估公司价值），帮助它们表现得更像经验丰富的律师或金融分析师。Trajectory 的其他客户还包括 Decagon、Mercor 和英伟达。目前，该初创公司正将其后训练平台向公众开放。

The people using Trajectory decide “what good looks like for their organization and can turn that into improvements for their model,” Malde said. Trajectory’s approach to continual learning shares some similarities with recursive self-improvement (RSI), the idea that AI systems can improve their own capabilities. But co-founder Michael Elabd argues that keeping humans involved in the training process makes Trajectory’s approach safer, since autonomous self-improvement could lead to unintended or even dangerous outcomes.

Malde 说，使用 Trajectory 的用户可以决定“对于他们的组织来说，什么是好的表现，并将其转化为对其模型的改进”。Trajectory 的持续学习方法与递归自我改进（RSI）有一些相似之处，即 AI 系统可以提高自身能力的理念。但联合创始人 Michael Elabd 认为，让人员参与训练过程使 Trajectory 的方法更安全，因为自主的自我改进可能会导致意想不到甚至危险的结果。

Trajectory isn’t the only startup trying to automate post-training. It considers Thinking Machines Lab, founded by former OpenAI CTO Mira Murati and the company behind the fine-tuning platform Tinker, a direct competitor, along with Applied Compute, which was valued at $3.25 billion last month, and Prime Intellect, which helps companies build and train their own AI agents.

Trajectory 并不是唯一一家试图实现后训练自动化的初创公司。它将由前 OpenAI 首席技术官 Mira Murati 创立的 Thinking Machines Lab（微调平台 Tinker 背后的公司）视为直接竞争对手，此外还有上个月估值达 32.5 亿美元的 Applied Compute，以及帮助公司构建和训练自有 AI 代理的 Prime Intellect。

Malde believes the startup’s biggest advantage is its team. He was among the handful of researchers Google DeepMind acqui-hired from Windsurf. Elabd also worked at DeepMind, while Trajectory’s third co-founder, Arjun Karanam, previously worked on generative foundation models at Apple. The startup has also recruited researchers and product designers from OpenAI, Stripe, and Netflix. Why does design matter so much to a company focused on AI training? Karanam says the startup has developed an interface that makes post-training more accessible to its nontechnical customers.

Malde 认为该初创公司最大的优势在于其团队。他曾是 Google DeepMind 从 Windsurf 收购的少数研究人员之一。Elabd 也曾在 DeepMind 工作，而 Trajectory 的第三位联合创始人 Arjun Karanam 此前曾在 Apple 从事生成式基础模型的工作。该初创公司还从 OpenAI、Stripe 和 Netflix 招募了研究人员和产品设计师。为什么设计对于一家专注于 AI 训练的公司如此重要？Karanam 表示，该初创公司开发了一个界面，使非技术背景的客户能够更容易地进行后训练。