Sign up to save your podcasts
Or
Share “Sabotage Evaluations for Frontier Models” by David Duvenaud, evhub, Joe Benton, Misha Wagner, Eric Christiansen, Ethan Perez, Buck, HoldenKarnofsky, Sam Bowman
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
“Sabotage Evaluations for Frontier Models” by David Duvenaud, evhub, Joe Benton, Misha Wagner, Eric Christiansen, Ethan Perez, Buck, HoldenKarnofsky, Sam Bowman
This is a linkpost for a new research paper from the Alignment Evaluations team at Anthropic and other researchers, introducing a new suite of evaluations of models' abilities to undermine measurement, oversight, and decision-making. Paper link.
Abstract:
Sufficiently capable models could subvert human oversight and decision-making in important contexts. For example, in the context of AI development, models could covertly sabotage efforts to evaluate their own dangerous capabilities, to monitor their behavior, or to make decisions about their deployment. We refer to this family of abilities as sabotage capabilities. We develop a set of related threat models and evaluations. These evaluations are designed to provide evidence that a given model, operating under a given set of mitigations, could not successfully sabotage a frontier model developer or other large organization's activities in any of these ways. We demonstrate these evaluations on Anthropic's Claude 3 Opus and Claude 3.5 Sonnet models. [...]
---
Outline:
(03:33) Human decision sabotage
(05:21) Code sabotage
(06:36) Sandbagging
(08:28) Undermining oversight
(10:00) Conclusions
---
First published:
October 18th, 2024
Source:
https://www.lesswrong.com/posts/nnvn6kaBLajiicH8e/sabotage-evaluations-for-frontier-models
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.
View all episodes
By LessWrongThis is a linkpost for a new research paper from the Alignment Evaluations team at Anthropic and other researchers, introducing a new suite of evaluations of models' abilities to undermine measurement, oversight, and decision-making. Paper link.
Abstract:
Sufficiently capable models could subvert human oversight and decision-making in important contexts. For example, in the context of AI development, models could covertly sabotage efforts to evaluate their own dangerous capabilities, to monitor their behavior, or to make decisions about their deployment. We refer to this family of abilities as sabotage capabilities. We develop a set of related threat models and evaluations. These evaluations are designed to provide evidence that a given model, operating under a given set of mitigations, could not successfully sabotage a frontier model developer or other large organization's activities in any of these ways. We demonstrate these evaluations on Anthropic's Claude 3 Opus and Claude 3.5 Sonnet models. [...]
---
Outline:
(03:33) Human decision sabotage
(05:21) Code sabotage
(06:36) Sandbagging
(08:28) Undermining oversight
(10:00) Conclusions
---
First published:
October 18th, 2024
Source:
https://www.lesswrong.com/posts/nnvn6kaBLajiicH8e/sabotage-evaluations-for-frontier-models
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.
More shows like LessWrong (30+ Karma)
View all
Making Sense with Sam Harris
26,366 Listeners
Conversations with Tyler
2,383 Listeners
The Peter Attia Drive
7,944 Listeners
Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas
4,137 Listeners
ManifoldOne
87 Listeners
Your Undivided Attention
1,459 Listeners
All-In with Chamath, Jason, Sacks & Friedberg
9,050 Listeners
Machine Learning Street Talk (MLST)
88 Listeners
Dwarkesh Podcast
386 Listeners
Hard Fork
5,422 Listeners
The Ezra Klein Show
15,220 Listeners
Moonshots with Peter Diamandis
473 Listeners
No Priors: Artificial Intelligence | Technology | Startups
120 Listeners
Latent Space: The AI Engineer Podcast
76 Listeners
BG2Pod with Brad Gerstner and Bill Gurley
456 Listeners