# Anyscale News: Post-training and RL across industries Canonical: https://brew.new/templates/anyscale/anyscale-news-post-training-and-rl-across-industries Brand: anyscale.com Category: general ![Preview of Anyscale News: Post-training and RL across industries](https://cdn.brew.new/email-preview-5bfec72f0dd436b1-tracking_r57rw3t4s8gt3b4vav7tey6c998gax3y-1792084883754.png) ## Email content Hi there, Post-training, whether its fine-tuning or running Reinforcement Learning (RL), is quickly becoming the way organizations are building moats where proprietary data is used to customize models. Whether detecting complex financial fraud, accelerating autonomous vehicle training, or discovering life-saving drugs, enterprises are leveraging post-training to deliver more value to their customers using AI. The concept is straightforward, combine large-scale training, online inference, and environments in a single tightly coupled loop, if you have the necessary AI infrastructure foundations that is. RL on Anyscale → What Makes RL the Hardest Workload in AI What makes RL the hardest_Robert Nishihara_Day 1 Reinforcement learning represents one of the most demanding distributed systems challenges in modern AI. In his Ray Summit 2026 keynote, Robert Nishihara (creator of Ray and co-founder of Anyscale) breaks down the anatomy of RL workloads. He explores why combining large-scale training with large-scale inference in a single loop stresses infrastructure in unprecedented ways, and how Ray was specifically architected to handle this exact shape of problem. Watch session recording → Coinbase: Fraud Agent Evals & Post-Training 791522da-ad81-44b0-b2b5-b48065a8ad52 The specialized model outperformed Claude Opus 4.5 across all four fraud-detection metrics—achieving a 9.6% higher F1 score and a 35.4% increase in dollar-weighted recall—while reducing end-to-end online serving latency by 55%. This milestone went viral when Brian Armstrong, CEO of Coinbase, shared the post-training achievements on X, highlighting the power of owning and specializing domain-specific AI models on Anyscale. Read the Coinbase Blog → Lila Sciences: Building an AI Science Factory Building an AI science factory_Andrew Beam_Day 1 Only ~5% of scientific hypotheses can be verified cheaply; the rest require physical laboratory experiments. Andrew Beam, CTO of Lila Sciences, demonstrates how Lila's AI science factory closes the loop between models and physical experiments. Trained on Anyscale, Lila's platform measures scientific scaling laws across thousands of experimental tasks—from designing quantum dots to reaching in vivo CAR-T candidates in months instead of years. Watch the Lila Sciences Presentation → Torc Robotics: Unifying Simulation and RL Unifying Sim_Neil Wadhvana_Torc Autonomous trucking requires complex task planning and massive scale for petabytes worth of data captured by each truck. Neil Wadhvana and Yashovardhan Chaturvedi from Torc Robotics share how they unified object simulation and end-to-end RL training with neural rendering on Ray. Discover how Torc adapted existing LLM post-training frameworks to scale RL loops for autonomous heavy-duty trucks. Watch the Torc Robotics Session → SkyRL: Scaling RL with an open-source stack 1200x675_2x (3) Want to dive deeper into building and scaling post-training and RL workloads? Join us for our upcoming live event where we'll discuss practical architectures, cost optimization strategies, and real-world implementation insights from top engineering teams. Register for the Live Event → Anyscale Try Free Anyscale_logo_grey_blue Advance your AI Platform with Anyscale. LinkedIn X Facebook Anyscale, Inc., 600 Harrison St, San Francisco, CA 94107, USA Unsubscribe Manage preferences [Open and remix this design](https://brew.new/templates/anyscale/anyscale-news-post-training-and-rl-across-industries) [Browse email designs](https://brew.new/browse/templates)