IDA achieves superhuman performance at coding

What AI 2027 Predicted

The scenario describes Iterated Distillation and Amplification (IDA) being deployed at scale around early 2027. IDA, originally proposed by Paul Christiano, involves a cycle: amplify a model’s capability through careful decomposition and verification, then distill the amplified capability back into the model. AI 2027 predicts that by 2027, models become good enough at verifying “subjective” outputs — not just math proofs or game outcomes — enabling IDA to improve performance across many real-world tasks. This is framed as one of several algorithmic breakthroughs that accelerate the path to superhuman AI.

How We Track This

We monitor:

• Self-play and self-improvement methods in frontier model training
• RL-based training with model-generated verification (constitutional AI, RLAIF)
• AlphaProof-style methods applied beyond formal mathematics
• Lab announcements about recursive self-improvement in training pipelines
• Academic work on IDA, debate, and amplification approaches

Current Evidence

Elements of IDA-adjacent techniques are increasingly visible in frontier AI development, though not yet at the scale or generality the scenario envisions:

AlphaProof (Google DeepMind): Published in Nature in November 2025, AlphaProof uses AlphaZero-inspired reinforcement learning to discover formal mathematical proofs. It trains on millions of auto-formalized problems — a clear example of amplification-distillation in a domain with formal verification. This represents IDA working in a domain with clear verifiers.

Constitutional AI and RLAIF: Anthropic’s Constitutional AI and the broader RLAIF paradigm (using AI feedback to train AI) represent a form of distillation and amplification, where model outputs are judged by models and the judgments are used for training. This is now standard practice at most frontier labs.

RL post-training scaling: The o1/o3 family and similar reasoning models use extensive RL to improve model capability, with reward models providing verification. This is a distillation step where reasoning strategies are baked into the model.

The gap: These techniques work where verification is relatively easy (math, code, clear-cut factual questions). AI 2027’s prediction specifically calls out IDA working for “subjective” tasks — creative writing, research taste, strategic planning — where verification is much harder. This remains undemonstrated.

Sources:

• Olympiad-level mathematical reasoning with RL — Nature
• Grading AI 2027’s 2025 Predictions — AI Futures Project
• Checking in on AI 2027 — LessWrong

Counterevidence & Limitations

• Current IDA-like techniques only work reliably in domains with formal or near-formal verification (math, code, games)
• The “subjective verification” breakthrough the scenario requires has no clear path — it’s the hard part of the prediction
• Self-play / RL-from-model-feedback can lead to reward hacking and mode collapse, limiting its generality
• It’s unclear whether constitutional AI / RLAIF truly counts as IDA or is a more limited form of the concept
• The gap between “IDA for math” and “IDA for everything” may be larger than the scenario implies

What Would Change Our Assessment

• Upgrade to “on-track”: A frontier lab demonstrates IDA-style self-improvement on open-ended tasks (research, writing, strategy) with measurable gains
• Upgrade to “confirmed”: Deployed production model uses iterative self-improvement across broad task categories
• Downgrade to “behind”: If by mid-2027, IDA remains confined to formally-verifiable domains
• Would also watch for: Advances in process reward models or outcome-based evaluation that enable verification of subjective quality

Update History

Date	Update
2026-03	IDA-adjacent techniques (AlphaProof for math, RLAIF, RL post-training) are widespread but limited to domains with clear verifiers. The predicted ‘subjective verification’ breakthrough — enabling IDA for open-ended tasks — remains undemonstrated.