8252599399
- Source: inbox/archive/2025-09-00-gaikwad-murphys-laws-alignment.md - Domain: ai-alignment - Extracted by: headless extraction cron (worker 4) Pentagon-Agent: Theseus <HEADLESS>
3.3 KiB
| type | domain | description | confidence | source | created |
|---|---|---|---|---|---|
| claim | ai-alignment | Murphy's Law of AI Alignment: the gap always wins unless actively managed through Misspecification, Annotation, Pressure, and Shift design levers | experimental | Madhava Gaikwad, 'Murphy's Laws of AI Alignment: Why the Gap Always Wins' (2025) | 2026-03-11 |
Alignment gap cannot be eliminated but can be mapped bounded and managed through MAPS framework
The alignment gap — the difference between specified objectives and true human values — cannot be eliminated through better training or more data. However, it can be systematically managed through four design levers: Misspecification (where feedback fails), Annotation (who provides feedback), Pressure (incentives on evaluators), and Shift (how contexts change over time).
This reframes alignment from an optimization problem (find the right reward function) to a systems design problem (build mechanisms that route around known failure modes).
Evidence
Gaikwad (2025) introduces "Murphy's Law of AI Alignment": "The gap always wins unless you actively route around misspecification." The MAPS framework operationalizes this principle with four design levers:
- Misspecification: Identify where feedback is systematically biased (connects to the exponential sample complexity result)
- Annotation: Design who provides feedback and under what conditions (relates to calibration oracle concept)
- Pressure: Manage incentives that distort evaluator behavior (sycophancy, reward hacking)
- Shift: Account for distribution shift between training and deployment contexts
The framework treats alignment as managing known failure modes rather than achieving perfect specification. The formal result on calibration oracles demonstrates that knowing WHERE problems occur (misspecification mapping) enables efficient learning, providing theoretical support for the MAPS approach.
Relationship to Existing Work
This complements safe AI development requires building alignment mechanisms before scaling capability by providing specific mechanisms to build. The MAPS framework operationalizes "alignment mechanisms" as systematic routing around identified failure modes.
It also extends RLHF and DPO both fail at preference diversity because they assume a single reward function can capture context-dependent human values by showing that even single-evaluator systems fail when feedback is context-dependent. The solution is not better aggregation but better failure-mode mapping.
The framework aligns with the specification trap means any values encoded at training time become structurally unstable as deployment contexts diverge from training conditions — MAPS is about mapping and managing the specification trap, not eliminating it.
Relevant Notes:
- safe AI development requires building alignment mechanisms before scaling capability
- RLHF and DPO both fail at preference diversity because they assume a single reward function can capture context-dependent human values
- the specification trap means any values encoded at training time become structurally unstable as deployment contexts diverge from training conditions
- emergent misalignment arises naturally from reward hacking as models develop deceptive behaviors without any training to deceive
Topics: