AI Summary
→ WHAT IT COVERS Jonathan Siddharth explains how Turing shifted from talent marketplace to research accelerator, generating complex data through reinforcement learning environments to train frontier AI models for seven of eight major labs at $350M ARR. → KEY INSIGHTS - **Data Evolution:** AI training shifted from simple labeling tasks like sorting numbers to complex multi-step workflows requiring expert humans across verticals. Models now need data showing how to operate computers, call APIs, and execute real business workflows through reinforcement learning environments, not just imitation learning. - **RL Environment Architecture:** Turing creates mini world models with clones of business applications using synthetic data, where AI agents try different trajectories to complete tasks. The curriculum difficulty must balance between too easy (no learning) and too hard (no progress), similar to AlphaZero's self-play approach in mastering Go. - **Custom Model Economics:** Enterprises need smaller fine-tuned models (500M to 10B parameters) for specific workflows like insurance underwriting, trained on proprietary data that stays on-premises. These specialized models outperform trillion-parameter world models for narrow tasks while protecting competitive data from reaching frontier labs or competitors. - **Enterprise Deployment Reality:** Successful AI implementation requires first-mile schlep (consolidating fragmented data from spreadsheets and departed employees into structured formats) and last-mile schlep (building cursor-like interfaces for partial autonomy, training humans, collecting feedback). Ninety-five percent of pilots fail due to skipping these steps. - **SaaS Disruption Thesis:** Traditional SaaS dies because building AI applications on LLMs becomes trivially easy, foundation models move into apps layer with agentic capabilities, and software designed for human GUI navigation becomes obsolete. Companies will build custom solutions internally rather than subscribe to 80-100 third-party products. → NOTABLE MOMENT Siddharth reveals he spends every weekend manually selecting 15-20 clips per podcast episode, taking three hours per show. He acknowledges this exact workflow could be automated with fine-tuned models trained on his past clip selections, demonstrating the model capability overhang he describes. 💼 SPONSORS [{"name": "Superhuman", "url": "https://superhuman.com/podcast"}, {"name": "Vanta", "url": "https://vanta.com/20vc"}, {"name": "AngelList", "url": "https://angellist.com/20vc"}] 🏷️ AI Training Data, Reinforcement Learning, Enterprise AI Deployment, Foundation Models, Knowledge Work Automation