The Second Half
Source: https://ysymyth.github.io/The-Second-Half/ Author: Shunyu Yao Date: 2025-08-06
Summary
Shunyu Yao (ReAct, Tree of Thoughts author) argues that AI’s “first half” was about solution methods (how to answer questions) and the “second half” is about problem definition (which questions to ask). The shift from answering to asking is the next frontier.
Key Claims
- First half (done): AI has learned to solve well-defined problems — coding, math, games with clear rules.
- Second half (current challenge): defining the problem is often harder than solving it. AI systems need to figure out what to optimize.
- ReAct as first half: ReAct showed agents could solve problems step-by-step. The second half is agents that define the task.
- Application to research: the value isn’t in running experiments — it’s in designing experiments. That’s still mostly human.
- Evaluation problem: you can’t evaluate problem-definition with standard benchmarks — they assume the problem is given.
- Prediction: the next major capability jump will be agents that can reformulate problems, not just solve them.
Entities
- Shunyu Yao — Princeton AI researcher; author of ReAct and Tree of Thoughts
Concepts
- Autonomous Research — problem-definition as the next frontier for research agents
- Coding Agents — application to coding task discovery
- Test-Time Compute — problem-reformulation as test-time compute strategy