Qualcomm and Xiomi Discuss RL-Based Training and Self-Improvement in AI

Jason Wei presented three frameworks for understanding AI in 2025: intelligence becoming a commodity, the Verifier's Law (AI capability correlates with task verifiability), and the jagged edge of intelligence. He predicts AI research automation by 2027 and notes that tasks with objective metrics are most susceptible to AI improvement.
Key insights
Intelligence as a Commodity
Jason Wei argues that AI intelligence is rapidly commoditizing, with the cost to achieve a given MMLU score dropping each year. This is driven by adaptive computation (test-time scaling) that allocates compute based on task difficulty. The time to access public information is also shrinking exponentially, from pre-internet to agentic AI. This democratizes knowledge but increases the relative value of private information.
Verifier's Law: AI Capability Tracks Verifiability
Wei proposes that AI's ability on a task is proportional to how easily the task can be verified. Key factors: objective correct answer, verification speed, parallelizability, noise level, and reward signal continuity. Tasks with high verifiability (like math and coding) are being solved first. He recommends inventing new measurement methods to make hard-to-verify tasks verifiable, enabling AI optimization.
The Jagged Edge of Intelligence and Self-Improvement
Wei rejects the 'fast takeoff' hypothesis, arguing AI self-improvement is gradual and task-specific. He provides three heuristics: AI excels at digital tasks (fast iteration), tasks humans find easy are often easy for AI (except those requiring superhuman perception), and data-rich tasks are easier. He predicts AI research automation by 2027, but physical tasks like plumbing and haircuts remain distant. A 75-page paper by 30 researchers classifies AI self-improvement into 5 levels, with 75% still at levels 1-2.