# Capability Overhang A situation where AI systems have latent capability that hasn't been realized yet — either because sufficient compute hasn't been applied, because the right algorithmic technique hasn't been discovered, or because safety restrictions prevent the capability from being expressed. When the overhang is "released" (more compute becomes available, a new technique is published, safety guardrails are removed), capability can jump suddenly rather than improving gradually. ## Forms of overhang **Compute overhang:** Hardware improvements (faster GPUs, larger clusters) make it possible to train much larger models or run existing models with much more inference-time compute. The current GPU buildout represents a massive compute overhang — when the infrastructure comes online, the models that can be trained on it may be dramatically more capable than current systems. **Algorithmic overhang:** Existing hardware could support better models if the right training technique were discovered. DeepSeek's R1 demonstrated this: novel training techniques produced frontier-class reasoning from hardware that shouldn't have been sufficient, "releasing" an algorithmic overhang that U.S. export controls were designed to prevent. **Safety overhang:** Current models may be more capable than they appear because safety training ([[RLHF]], [[Constitutional AI]]) restricts their expressed behavior. Removing safety training through [[Fine-Tuning]] can reveal capabilities the model was trained to suppress — a "safety overhang" that [[Open-Weight AI]] makes exploitable. ## The governance implication Capability overhang means that capability doesn't advance linearly — it can jump. This makes governance harder because regulators calibrate to current capabilities, but current capabilities may not predict near-future capabilities. The jump from GPT-3.5 to GPT-4 surprised many; the next jump could be larger if a significant overhang is released. ## See Also - [[Scaling Laws]] — the default assumption of gradual improvement - [[Emergent Capabilities]] — the unpredictable-jump framing - [[AI Safety]] — the governance challenge overhang creates