「但……當我意識到自己只是做好本分工作時，我便放下了這種批判。」
Стало известно об отступлении ВСУ под Северском08:52
。电影对此有专业解读
Finally, there is the synthetic-data-driven, product closed-loop flywheel. Noin centers its approach on proprietary synthetic data, building a training system tailored to embodied manipulation: through scalable task generation, action/trajectory generation, and filtering mechanisms, it continuously produces high-quality training data that covers long-tail scenarios, which is then used to train embodied foundation models with stronger generalization. Compared with routes that rely heavily on demonstrations and real-world data collection, the company places greater emphasis on a “controllable, scalable, and iterative” synthetic-data pipeline, and feeds back product and real-hardware runtime signals—such as feedback, failure cases, and abstractions of critical scenarios—into its data generation and evaluation system, forming a closed-loop flywheel of “product feedback → synthetic enhancement → training iteration → experience improvement.” Backed by a high-quality synthetic-data pipeline, it continues to drive model capability gains, creating a hard-to-replicate self-evolving system and cementing long-term technical barriers. This route has a high engineering threshold; Noin has already validated the key links and established a sustainable gain-and-verification system for embodied manipulation and task generalization.
One more thing: pay attention to what you celebrate publicly. If every shout-out in your team channel is for the big, complex project, that’s what people will optimize for. Start recognizing the engineer who deleted code. The one who said “we don’t need this yet” and was right.