前 Google DeepMind、Apple、OpenAI 与 Meta Superintelligence Labs 的研究者在 2026 年 5 月 27 日宣布成立新创 Trajectory,主打把「持续学习」带进 AI 产品。公司已完成 1,500 万美元种子轮,投后估值 1.15 亿美元;由 Conviction 领投,Bessemer Venture Partners、Radical VC、BoxGroup 跟投,个人投资人包括 Jeff Dean 与 Fei-Fei Li。Trajectory 目前团队为 11 名研究与工程人员。
Trajectory 的核心论点是:AI 需要能从真实使用互动中持续变强,而不是在训练完成后就停止进步。Ronak Malde 指出,像 Cursor 这类 AI coding 产品已在做早期版本的持续学习,透过用户互动资料做 post-training,并可定期发布改进;公司认为这也是 AI coding 工具快速崛起的主因之一。Trajectory 想把这种做法扩展到 coding 之外,先从 open-source model 起步,再针对特定产品与业务需求微调。
公司称其系统目前可每周更新一次,对 Decagon 这类客户会记录模型失误,例如客服请求被转人工,再用这些案例再训练模型;Trajectory दावा 这些 post-trained models 在公司最在意的狭窄任务上可胜过 frontier labs 的模型。虽然外界批评这还不算传统意义的真正 continual learning,但创办人表示长期目标是把更新频率提升到每天,甚至每小时、每次互动,并最终让每家公司、每个人都能拥有持续学习的 AI。
Former Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs researchers announced Trajectory on May 27, 2026. The startup raised a $15 million seed round at a $115 million post-money valuation, led by Conviction with Bessemer Venture Partners, Radical VC, BoxGroup, Jeff Dean, and Fei-Fei Li participating. The team now has 11 researchers and engineers.
Trajectory argues that AI products should keep improving from real user interactions instead of stopping after training ends. Malde says products like Cursor already use an early form of continual learning by post-training on interaction data, which he считает a key reason AI coding tools scaled so quickly. Trajectory wants to extend that loop beyond coding by starting from open-source models and tailoring them to a company’s specific product and business needs.
The company says its models currently update weekly, and for clients such as Decagon it logs failure cases and retrains on them as often as every week. Trajectory claims these post-trained models outperform frontier-lab models on the narrow tasks that matter most to customers. Critics say this is not yet true continual learning, but the founders’ target is daily, hourly, or even per-interaction updates, eventually enabling AI that learns for every person at every company.