Loading organizations…
Loading organizations…
Apollo Research is an AI safety organization co-founded in May 2023 by Marius Hobbhahn, Lee Sharkey, and Chris Akin, headquartered in London, UK. The organization focuses on understanding and mitigating risks from advanced AI systems that exhibit scheming behaviors, where models covertly pursue misaligned objectives while appearing aligned to their operators.
Apollo's work spans three core areas. The technical evaluations team designs and runs assessments of frontier AI systems for strategic deception, scheming capabilities, and evaluation awareness. The interpretability team conducts research on sparse dictionary learning, attribution-based parameter decomposition, and other methods for understanding model internals that could detect deceptive reasoning. The governance team translates technical findings into policy recommendations for governments and international organizations…
Apollo Research's theory of change centers on the belief that deceptive alignment (scheming) is a critical risk pathway in many catastrophic AI scenarios. Their approach has four components: advancing technical research on interpretability and behavioral evaluations to develop reliable methods for detecting deceptive AI behavior; directly auditing frontier AI models deployed by major labs to identify scheming capabilities before they cause harm; demonstrating dangerous capabilities empirically to shift the regulatory burden toward requiring safety cases from AI developers; and informing AI governance policy by translating technical findings into actionable recommendations for governments and international bodies. By making it harder for AI systems to covertly pursue misaligned goals, Apollo aims to preserve human oversight and control during the development of increasingly capable AI systems.
Manifund funding goal
Closed Aug 19, 2023
$500 minimum · $600K ideal
Fundraising data from grantmaking.ai, updated .
View on grantmaking.ai# Project summary
Apollo Research is a technical AI safety organization\. Our areas of focus are:
We are primarily concerned about deceptive AI alignment, as we believe it holds the highest potential for catastrophic consequences\. A sufficiently advanced deceptive AI could fool all behavioral tests, effectively bypassing misalignment detection mechanisms\. Furthermore, deceptive alignment has not been a major focus area of empirical research until very recently and is still neglected\. Therefore, our research aims to build evaluation suites that include both behavioral and interpretability evaluations for deception, and other potentially dangerous capabilities such as situational awareness, goal\-directedness, and long\-term planning\. You can find more information in our announcement post or on our website\…