The data required to advance AI has shifted from simple, low-skill labeling (e.g., images, basic text) to complex, expert-generated data. The new frontier is creating reinforcement learning (RL) environments that simulate real-world business workflows, teaching AI agents to perform multi-step, economically valuable tasks.
The speaker argues that the SaaS model is 'completely over' because foundation models will increasingly build agentic capabilities that replicate and replace the functions of standalone software applications. The future business model is not selling software subscriptions but providing services to build, fine-tune, and deploy custom AI agents for enterprises.
The development of AGI and superintelligence is framed as a high-stakes competition between a handful of frontier labs (e.g., OpenAI, Anthropic, DeepMind). This race is fueled by massive capital investments in compute, research, and data, with the winner expected to dominate multiple major markets from search to consumer devices.
The AI industry exhibits extreme power-law dynamics, with significant revenue concentration at key layers of the stack. NVIDIA's revenue, for example, is highly dependent on a few large customers, and Turing's business is focused on serving seven of the eight major frontier labs. The key competitive moat is shifting from software features to proprietary data feedback loops that continuously improve model performance.
Keep pulling the thread on Jonathan Siddharth.