Scale AI

Today marks a huge step forward in our mission to advance AI with human values at its core. The future of AI is bright and we’re just getting started! 🚀

At Scale AI, we're working on using RL to build agents that work and learn like humans. It is very exciting to see how using rubrics, we can encode expert knowledge into real world AI agents.

The next 6 months in AI won’t look like the last. On Human in the Loop, we unpack what’s changing and why it matters. 🎧 Listen in: bit.ly/45hbgij

🔍 How do you scale quality control for data that challenges even PhDs and today’s most advanced LLMs? Demand for PhD-level, multimodal reasoning data is rising fast — critical for training models through reinforcement learning to handle complex, extended reasoning. To meet the challenge, we built autoraters: LLM agents that use multi-agent debate to evaluate expert-level data. 💡 The impact: → 4x more errors detected (23% → 82%) → 87% improvement in answer quality By combining AI capabilities with human expertise, we’re pushing the boundaries of data quality and model performance.

The U.S. AI Action Plan is a major step towards securing American leadership in AI. We’re proud to see many of the priorities we outlined in our “Promote, Unleash, Innovate” framework reflected in the Plan—advancing innovation, building critical infrastructure, and establishing global standards rooted in democratic values. The Plan outlines a strong path forward: deploying AI across the government, expanding access to U.S. technology for allies, and empowering the American workforce. We’re committed to supporting this vision by building safe, high-impact AI systems that strengthen America’s position in the race for AI dominance. Full response here: https://lnkd.in/gtxRXZ4k

🎙️ In today’s episode of Human in the Loop, we unpack the latest in AI and why Reinforcement Learning with Verifiable Rewards (RLVR) is standing out as a game changer for enterprises. Catch the full episode: https://scl.ai/hitl-e10

Meet MultiNRC, our latest SEAL Leaderboard that’s raising the bar for multilingual reasoning. Most benchmarks rely on translated English questions, missing important cultural and linguistic nuances. MultiNRC takes a different approach: tasks are natively written in French, Spanish, and Chinese across four reasoning categories, plus a translated subset to reveal gaps between languages. The results are clear: significant gaps in how models reason across languages and domains highlight the critical need for AI trained on richer, more diverse data that reflects the full spectrum of global language and culture – so we can build AI that works for everyone.

AI investments are no longer experimental, they’re delivering real results. Tune in 🎧 http://bit.ly/40aKYMO

🎙️ Live from SFHQ, it’s Human in the Loop! In today’s episode, our Enterprise team votes on the internet’s hottest AI takes. Watch the full video: http://bit.ly/40aKYMO