Lessons from Trillion Token Deployments at Fortune 500s
summary
Claims that reinforcement learning is the production mechanism not a nicety backed by a concrete stack: 95% of GenAI pilots fail because feedback integration is missing; the deployed pattern uses synthetic data as a byproduct of environment training plus mock environments for safe failure and LLM judges replace expensive annotation to codify rewards. An RL Ops platform (PPO-style orchestration of multiple LLMs) runs at enterprise tokenomics, serving Fortune 500 use cases like AT&T Manulife CCS and cutting inference cost and latency while continuously improving.
AI-written summary · about
chapters (13)
- 0:00RL Ops and Adaptive ML
- 1:20Myth of last mile to production
- 3:23MVP to production marathon
- 4:04Reinforcement learning outperforms others
- 5:09Cost savings with smaller models
- 6:36Ownership over AI solution
- 7:19Agents make production harder
- 8:21Training agents with workflows
- 9:47Synthetic data pipelines
- 11:38Human feedback in RL
- 13:24Industrializing RL Ops
- 15:10Overcoming RL complexity
- 15:53Q&A
related talks
19mTraining Agentic Reasoners
32mBeyond LLMs: Building an AI-Powered Decision System at Scale
26mCan Reinforcement Learning Lead to AGI?
12mKeynote: How LLMs Have Changed the Way Organizations Do AI
43mLLMs in the Wild - Streaming, RAG, and Real Time GenAI at Scale
19mRL for Autonomous Coding
36mSkyRL: Democratizing RL Training for Researchers
28mNavigating LLM deployment: Tips, tricks, and techniques