IDA achieves superhuman performance at coding
Now, the models have become sufficiently good at verifying more subjective things …, allowing the use of IDA to improve the model at many tasks.
Appendix F - Iterated distillation and amplification (IDA) · Archived PDF, p. 49
At a glance
- Assessment: Emerging
- Confidence in assessment: 40%
- Predicted timing: Early 2027
- Primary source: AI 2027 archived scenario: Appendix F - Iterated distillation and amplification (IDA), p. 49
What AI 2027 Predicted
Tracker summary: 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. Archived PDF, p. 49.
Scenario source: archived version and page-number reference: Appendix F - Iterated distillation and amplification (IDA), p. 49.
How We Track This
Test whether a frontier training process uses model-based judgments of subjective work quality as the selection or reward signal in iterated amplification and reinforcement-learning distillation, and demonstrates improvement across multiple non-verifiable tasks on held-out evaluations. Require a technical account identifying the amplification procedure, the learned selection or reward signal, the distillation step, task coverage and a comparison against the prior model or an appropriate ablation. Formal-verifier gains alone do not pass this test. Assess the title’s coding component separately against Superhuman coder emerges; broader subjective verification alone does not establish superhuman coding. These are tracker operational requirements for the mechanism described in Appendix F, archived PDF p. 49.
Current Evidence
Related AI-training techniques supply precursor evidence, but the reviewed sources do not establish the specified broad IDA deployment.
AlphaProof: Reinforcement learning with formal proof verification demonstrates improvement in a well-verifiable domain. It should not automatically be classified as the decomposition, amplification and distillation process specified here.
Constitutional AI and RLAIF: Training from AI feedback is relevant, but the use of model judgments alone does not establish the full IDA cycle or its broad deployment.
RL post-training: Reinforcement learning can improve model capability. RL, distillation and iterated amplification are distinct mechanisms and require separate technical evidence.
The gap: The claim concerns scaling the specified mechanism to harder-to-verify outputs. Research on related techniques does not by itself resolve that claim.
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
- Results in formally verifiable domains cannot establish broad performance on subjective tasks.
- The reviewed sources do not establish the specified subjective-verification breakthrough; they also do not prove there is no viable path.
- 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 mechanism and domain coverage must be explicit; a generic self-improvement label is insufficient.
What Would Change Our Assessment
- Strengthen the assessment: Technical evidence demonstrates the specified amplification-and-distillation cycle on open-ended tasks.
- Confirm: Production use establishes that mechanism and broad task coverage, rather than generic iterative improvement.
- Reassess timing after early 2027: Keep the original window while distinguishing missing evidence from a demonstrated failure.
- Watch for: Better verification of subjective quality, with generalization tested separately.
Update History
| Date | Update |
|---|---|
| 2026-11-07 | Assessment reviewed. Status: Emerging; confidence: 40%. Anthropic’s December 2022 Constitutional AI research demonstrates self-critique, revision and AI preference judgments used in reinforcement learning. This is concrete precursor evidence for learned subjective evaluation. It does not establish the full tracked amplification-and-distillation mechanism across many tasks at frontier scale or the separate superhuman-coding outcome. Clarification: Specified the subjective-evaluation, amplification and distillation evidence needed for the mechanism test and linked the separate coding criterion. Related training techniques alone do not establish broad deployment. |
| 2026-09-06 | Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations. |
| 2026-03-13 | 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. |