Artificial intelligence is poised to become a transformative technology with massive benefits. It could raise living standards across the world, transform key sectors such as healthcare and education, and accelerate scientific discovery. As with any transformative technology, it also comes with risks. The high rate of change and novelty of impacts make it hard to forecast and mitigate risks of a novel transformative technology. This article is based on Google Deepmind's Paper on 'An Approach to Technical AGI Safety and Security.'
1. Introduction
Considering the many risks, while it is appropriate to include some precautionary safety mitigations, the majority of safety progress should be achieved through an “observe and mitigate” strategy. However, as technologies become ever more powerful, they start to enable severe harms. An incident has caused severe harm if it is consequential enough to significantly harm humanity. “Observe and mitigate” is insufficient as an approach to such harms, and we must instead rely on a precautionary approach.Google's strategy for mitigating severe risks from AGI focuses on technical safety and security mitigations, specifically addressing misuse and misalignment. For misuse, their approach is to evaluate whether the model has the capability to cause severe harm, put in place appropriate deployment mitigations and security mitigations if so, and assess the quality of the mitigations by attempting to break them. For misalignment, their strategy is to attain good oversight using the AI system itself to help with the oversight process, identify cases where oversight is needed, apply defense in depth to defend against misaligned AI systems, and assess the quality of the mitigation through alignment assurance techniques.
2. Navigating the Evidence Dilemma
While severe harms necessitate a precautionary approach, such an approach to a new and quickly evolving technology suffers from an evidence dilemma. Precautionary mitigations will be based on relatively limited evidence, and so are more likely to be counterproductive. Research efforts aimed at future problems may study mitigations that later turn out to be infeasible, disproportionate, or unnecessary.An intermediate approach is possible. Many risks do not seem plausible given currently foreseeable capability improvements but could arise with capability improvements that are not yet on the horizon. These risks are good candidates to defer to the future when more evidence will be available.
3. Assumptions about AGI Development
Google's approach is underpinned by five core assumptions: the current paradigm continuation assumption, the no human ceiling assumption, the uncertain timelines assumption, the potential for accelerating improvement assumption, and the approximate continuity assumption.
3.1. Current Paradigm Continuation
The current paradigm for developing frontier AI systems will continue to represent the dominant approach to increasing AI capability for the foreseeable future. This assumption influences Google's focus on an anytime approach to AGI safety.
3.2. No Human Ceiling for AI Capability
AI capabilities will not cease to advance once they achieve parity with the most capable humans. This assumption implies that Google's approach to safety must leverage new AI capabilities as they become available.
3.3. The Timeline of AI Development Remains Uncertain
The timeline for the development of highly capable AI remains unclear. Given this uncertainty, it is crucial that frontier AI developers have an anytime safety approach to put in place in the case of short timelines.
3.4. The Potential for Accelerating Capability Improvement
The use of AI systems could plausibly lead to a phase of accelerating growth. Google's risk mitigation strategy must also be significantly accelerated through AI assistance to ensure we retain the ability to react to novel risks as they arise.
3.5. Approximate Continuity
General AI capabilities will scale fairly smoothly and predictably with the availability of computation, R&D effort, and data. This assumption enables us to iteratively and empirically test Google's strategies and detect flawed assumptions as capabilities improve.
4. Risk Areas
When addressing safety and security, it is helpful to identify broad groups of pathways to harm that can be addressed through similar mitigation strategies. We consider four areas: misuse, misalignment, mistakes, and structural risks.
4.1. Misuse Risks
Misuse risks describe risks of harm ensuing when a user intentionally uses the AI system to cause harm against the intent of the developer. Google's approach to misuse proactively identifies dangerous capabilities, measures relevant model capabilities, and implements security and deployment mitigations.
4.2. Misalignment Risks
Misalignment occurs when an AI system knowingly causes harm against the intent of the developers. Google's strategy to address misalignment begins with attaining good oversight, using the AI system itself to help with the oversight process, and using it to train the AI system.
5. Addressing Misuse
This section describes measures an AI developer can adopt to significantly reduce misuse risks. Google's objective is to reduce the risks of severe harm occurring via misuse by making it difficult or unappealing for entities to inappropriately access dangerous capabilities of powerful models.
6. Addressing Misalignment
Google's strategy to address misalignment involves attaining good oversight, identifying cases where oversight is needed, applying defense in depth, and assessing the quality of the mitigation through alignment assurance techniques.
Conclusion
The transformative nature of AGI has the potential for both incredible benefits as well as severe harms. To build AGI responsibly, it is critical for frontier AI developers to proactively plan to mitigate severe harms. Google's approach outlines technical mitigations for misuse and misalignment, and we hope that this paper can help the broader community to join us in enabling us to safely and securely access the potential benefits of AGI.
Full Google Article and Paper:
https://deepmind.google/discover/blog/taking-a-responsible-path-to-agi/
Deep Dive Video:
https://www.youtube.com/watch?v=2aenIJ4C6ic
No comments:
Post a Comment