The Scale AI pivot and the human feedback reality

The $14.3 billion Meta investment truth

I watched the AI industry react to the $14.3 billion Meta investment in Scale AI this past June. The deal prompted rumors about the company’s viability, and many observers viewed the transaction as a scheme to hire founder Alexandr Wang. Scale CFO Dennis Cinelli denied these "zombie company" claims, stating that the firm remains operational and thriving with revenue in the nine-figure range. The company employs more than 1,000 workers and serves customers like Microsoft, OpenSea, and the U.S. Army. Scale also maintains partnerships with OpenAI, even after OpenAI reduced its collaborative efforts with the startup following the Meta deal. Scale recently secured a $99 million contract from the U.S. Department of Defense in August and a $100 million contract in September. The company’s valuation reached more than $29 billion after the Meta investment, which provides capital for product innovation and shareholder liquidity. As part of the deal, Alexandr Wang joined Meta to work on its AI initiatives, and Jason Droege became the interim CEO. Ruben Mayer, the former senior vice president of GenAI Product and Operations, joined Meta to work in TBD Labs. Mayer stated his role focused on setting up the lab rather than data operations.

RLHF limitations and the expert shift

Scale moves away from basic labeling toward specialized annotation to support enterprise and defense contracts, a shift that follows the layoff of hundreds of data-labeling contractors as the company prioritizes precision over volume. I see this transition as a move toward "expert data" because models require more precision as they grow more capable. RLHF relies on human-driven reward signals, which use subjective human preference labels that often suffer from human bias. This method tweaks behavior rather than improving model intelligence through autonomous discovery. True reinforcement learning requires an environment where an agent discovers solutions through verifiable rewards, a process distinct from the human-driven RLHF used in models like GPT-3.5. In the RLHF process, training includes Supervised Fine-Tuning and Reward Modeling Training. Because reward modeling relies on human preference labels, the process can suffer from human bias or lead to overfitting. Many labs call this "post-training" instead of RLHF to distinguish it from true reinforcement learning. This distinction matters because models like Grok-4 have scaled RL compute 10x to near pre-training effort. Capturing a human preference signal for things like dialogue quality or groundness remains difficult and highly subject to bias.

Data labeling benchmarks and competitors

AI platforms favor Scale for many data management tasks. For example, ChatGPT selects Scale 59% of the time when users ask for large-scale ML data cleaning and labeling. Scale also holds a lead in RLHF preference ranking data management, where AI picks Scale 118% more often than others. TELUS Digital remains a major rival in the multilingual and autonomous vehicle sectors. TELUS Digital, which formerly operated as Lionbridge AI, provides specialized computer vision annotation and manages global workforces for autonomous driving programs. The company provides coverage in over 50 languages and handles large-scale machine learning projects. TELUS Digital employs 80,000 specialists across 35+ countries and serves over 600 global brands. This expertise includes Demand Generation, Inside Sales, and Channel Management solutions. TELUS Digital won Gold at the 2026 Stevie Awards for its Demand Generation Program, which delivered 47.6% conversion rate growth and $121 million in operating net revenue for a fintech client. While Scale AI maintains its dominance in AI-driven recommendations for data cleaning, TELUS Digital provides a compelling alternative for companies requiring massive multimodal operations or deep multilingual support across fifty different languages. You know that high-quality data determines model performance. How will the increasing cost of expert annotation affect the scaling of reasoning-heavy models?