Jason Wei's Three Ideas for Understanding AI in 2025

Intelligence as a Commodity

• 💡 The cost and accessibility of intelligence are being driven towards zero, making any public internet knowledge instantly available.
• 📈 AI progress involves two stages: frontier pushing with increasing compute and then commoditization, where the cost for the same intelligence dramatically decreases.
• 🧠 Adaptive compute is a key reason for continued cost reduction, allowing AI to spend more compute on harder problems and less on easier ones, optimizing task performance.
• ⏳ The concept of instant knowledge means the time to retrieve public information is converging to zero, moving from pre-internet searches to near-instant chatbot and agent responses.
• ✅ Implications include the democratization of fields like coding and personal health, higher value for private information, and extremely frictionless access to personalized information.

Verifier's Law and Asymmetry of Verification

• ⚖️ Asymmetry of verification describes tasks where it's much easier to verify a solution than to find one, such as Sudoku puzzles or building complex websites like Twitter.
• 🎯 Verifier's Law states that AI's ability to solve a task is proportional to how easily it can be verified, meaning any solvable and easily verifiable task will eventually be conquered by AI.
• 🔍 Key criteria for verifiability include objective truth, fast and scalable verification, low noise, and the presence of a continuous reward signal.
• 📊 This law explains why AI has quickly saturated many benchmarks, as good benchmarks are inherently easy to verify and meet most of these criteria.
• 🚀 Algorithms like AlphaFold exemplify this by using large language models to sample solutions, get continuous grades, and iteratively improve, especially for problems with clear evaluation metrics.

The Jagged Edge of Intelligence

• ⚠️ There will likely be **no