Labelbox's latest updates introduce several enhancements aimed at improving data exploration, labeling, and model error identification. Users can now upload custom embeddings, which represent data as numerical vectors, to better explore and find similar data, with support for up to 100 custom embeddings per organization. The updates also include SDK improvements, such as enhanced global keys for seamless annotation imports and a new Export v2 workflow that aligns with the import format, enhancing data export processes. A step interpolation feature has been added for video labeling, allowing annotations to remain fixed between keyframes rather than transitioning linearly, offering more precise control over video annotations. Additionally, architectural improvements promise faster labeling in the APAC region and a high-throughput data ingestion system to handle extremely large datasets. Model-assisted labeling workflows now support importing model predictions as pre-labels, facilitating human-in-the-loop reviews and enabling users to prioritize assets based on model confidence, ultimately aiming to address data discrepancies and improve model performance.