Ablation studies are a vital tool in our attempts to understand the internal computations of neural networks: by ablating components of a trained network at inference time and studying the downstream effects of these ablations we hope to be able to map the network's computational structure and attribute responsibility among different components. In order to interpret the results of interventions on neural networks we need to understand how network computations respond to the types of interventions we typically perform. A natural expectation is that ablating important components will substantially degrade model performance (Morcos et al., 2018) and may cause cascading failures that break the network. We demonstrate that the situation in large language models (LLMs) is substantially more complex: LLMs exhibit not just redundancy but actively self-repairing computations. When one layer of attention heads is ablated, another later layer appears to take over its function.

Power-seeking behavior is a key source of risk from advanced AI, but our theoretical understanding of this phenomenon is relatively limited. Building on existing theoretical results demonstrating power-seeking incentives for most reward functions, we investigate how the training process affects power-seeking incentives and show that they are still likely to hold for trained agents under some simplifying assumptions. We formally define the training-compatible goal set (the set of goals consistent with the training rewards) and assume that the trained agent learns a goal from this set. In a setting where the trained agent faces a choice to shut down or avoid shutdown in a new situation, we prove that the agent is likely to avoid shutdown. Thus, we show that power-seeking incentives can be probable (likely to arise for trained agents) and predictive (allowing us to predict undesirable behavior in new situations).

There is considerable uncertainty about what properties, capabilities and motivations future AGIs will have. In some plausible scenarios, AGIs may pose security risks arising from accidents and defects. In order to mitigate these risks, prudent early AGI research teams will perform significant testing on their creations before use. Unfortunately, if an AGI has human-level or greater intelligence, testing itself may not be safe; some natural AGI goal systems create emergent incentives for AGIs to tamper with their test environments, make copies of themselves on the internet, or convince developers and operators to do dangerous things. In this paper, we survey the AGI containment problem - the question of how to build a container in which tests can be conducted safely and reliably, even on AGIs with unknown motivations and capabilities that could be dangerous. We identify requirements for AGI containers, available mechanisms, and weaknesses that need to be addressed.