Gergely Orosz Points Out Meta Software Engineers Assigned Manual Data Annotation Tasks
Gergely Orosz revealed that Scale AI, during Alexandr Wang's tenure as CEO, required software engineers to engage in manual data annotation, which the new leadership found shocking and immediately halted.
Now, Meta is reportedly forcing software engineers into full-time manual data annotation work, with about 5,000 engineers transitioning from roles in advertising and infrastructure to the AI team, causing dissatisfaction and a wave of resignations within the company.
This model reflects the enormous demand for high-quality annotated data in AI training. Although such tasks are typically performed by contractors, some AI leaders believe that having engineers annotate data personally helps build model intuition.
Source: Public Information
ABAB AI Insight
Alexandr Wang promoted engineer involvement in annotation during his time at Scale AI before leaving to join Meta to lead AI-related work. Previously, Scale AI had established a deep binding through a $14 billion investment from Meta, and Wang has long emphasized that data quality is a core bottleneck for AI model performance.
In terms of capital strategy, Meta prioritizes ensuring control over internal data pipelines and customized annotation through large-scale personnel reallocation and a significant stake in Scale AI, rather than relying entirely on the external labor market. This move aims to accelerate the iteration of its own models while locking in key resource investments in the intense AI competition.
Similar cases of engineer reallocation at Meta (such as the shift from infrastructure to AI) have also been seen in other major companies during peak model training periods; currently, Meta is in a phase of transitioning from reliance on outsourced AI infrastructure to deep involvement of internal engineers.
Essentially, this represents capital concentration: giants like Meta strengthen vertical integration of the training data supply chain by directly investing high-cost labor into data annotation, squeezing the space for small and medium annotation service providers, while leveraging engineers' professional backgrounds to enhance annotation accuracy and embed proprietary knowledge, creating difficult-to-replicate competitive barriers.
ABAB News · Cognitive Law
Data is the new oil, annotation is the refining: whoever controls the annotation chain, controls the sovereignty of the model.
Engineers sell time vs. sell structure: the former does annotation, the latter builds systems, with long-term outcomes clearly defined.