XNews

Jason Wei discusses AI frameworks at Stanford

Jason Wei discusses AI frameworks at Stanford

Jason Wei from Meta Superintelligence Labs 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 AI capabilities. He predicted timelines for various tasks from translation (done) to dating (impossible).

Key insights

Intelligence is becoming a commodity as costs drop

Wei showed that the cost to achieve a given level of intelligence on benchmarks like MMLU has been dropping every year. Adaptive computation (from o1 onwards) allows models to dynamically allocate compute based on task difficulty, pushing costs for simple tasks to near zero. The time to find public information has shrunk exponentially from pre-internet to agentic AI.

Verifier's Law: AI capability is proportional to task verifiability

Wei proposed that AI's ability on a task is proportional to how easily the task can be verified. Key factors: objective correctness, verification speed, parallelizability, noise level, and continuous vs binary reward. He cited AlphaEvolve as an example that exploits verification asymmetry by iterating on the same problem with automatic scoring.

AI capabilities form a jagged edge, not a smooth curve

Wei argued against rapid takeoff, noting AI's capability landscape is uneven: strong in some areas (competition math, programming) but weak in others (9.11 > 9.9, Tlingit language). He proposed three heuristics for progress speed: digital tasks iterate faster, human-easy tasks are often AI-easy (except where humans are physically limited), and data-rich tasks perform better.

Timeline predictions for various tasks

Wei provided a table: mainstream language translation (done), basic code debugging (2023), competition math (2024), AI research (~2027), chemistry research (later, non-digital), movie making (~2029), plumbing/haircuts (unlikely soon), traditional Uzbek carpet weaving (unlikely), and 'dating to girlfriend's satisfaction' (impossible, non-digital, lack of data).

Source posts