Continuous FunctionStart learning

Foundation Lab

AI Safety via Debate

Targets evaluation difficulty: we can judge arguments even when we can't judge answers

Concept 86 of 100Scaling & AlignmentPhase 12
Inspect equationAll Foundations
#86DebateScaling & Alignment
key equation
max⁡πAmin⁡πBE[J(τ)]\max_{\pi_A} \min_{\pi_B} \mathbb{E}[J(\tau)]
Reading map and next steps

Selected Foundation Object

Keep the equation fixed; move through the evidence.

Concept 86 of 100DebateScaling & Alignment / Phase 12: Advanced alignment & safety research
Current question

Think Socratic dialogue meets adversarial training

max⁡πAmin⁡πBE[J(τ)]\max_{\pi_A} \min_{\pi_B} \mathbb{E}[J(\tau)]
PredictionCommit before tracing the equation.

Ask what should change under a concrete input, then trace that expectation through the equation.

EvidenceCompare the equation and source.

Use the key equation and canonical papers as the available witnesses, without implying that a runnable panel exists.

InvariantName what survives notation changes.

The useful learning product is the reusable mechanism you can carry into another model, paper, or engineering tradeoff.

Next moveContinue through the atlas.

Use prerequisites, dependents, and semantic links to repair the next gap without leaving the object behind.

Inspect equationBrowse atlasNext foundation

Why It Matters for Modern Models

  • Targets evaluation difficulty: we can judge arguments even when we can't judge answers
  • Scalable oversight: judge weaker than debaters can still pick truth
  • Adversarial structure surfaces hidden flaws in reasoning

What Tutorials Skip

What is still poorly explained in textbooks and papers:

  • Think Socratic dialogue meets adversarial training
  • Claims + evidence + counterexample structure
  • Judge accuracy improves with debate length

Visualization Status

Core Math (Optional Deep Dive)

If you want intuition first, start with the key equation and cited sources. Come back here for the full walkthrough.

Key Equation
max⁡πAmin⁡πBE[J(τ)]\max_{\pi_A} \min_{\pi_B} \mathbb{E}[J(\tau)]

Two agents debate; judge picks winner. Zero-sum game:

max⁡πAmin⁡πBE[J(τ)]\max_{\pi_A} \min_{\pi_B} \mathbb{E}[J(\tau)]

where τ\tau is the debate transcript and J(τ)∈{0,1}J(\tau) \in \{0,1\} indicates A wins.

Key insight: self-play pushes agents toward truthful, checkable arguments.

Canonical Papers

AI safety via debate

Irving et al.2018arXiv
Read paper →

Connections

Prerequisites

⚖RLHF

Next Moves

Choose the next question to carry this object forward.