Redwood Research
An independent AI safety research organization focused on adversarial robustness and interpretability research, dedicated to ensuring AI systems remain controllable under extreme inputs.
Core Facts
Timeline (last 90 days)
AI safety research organization Redwood Research published an analysis article pointing out that latent-space reasoning architectures represented by COCONUT and full-bandwidth Transformers may greatly weaken the currently most effective AI supervision tool — chain-of-thought (CoT).
Redwood Research believes that interpretability techniques such as SAEs, probes, and NLAs will be difficult to sufficiently replace the supervisory function of CoT in the near term.
Redwood Research believes that coordinating to avoid latent-space architectures is more feasible than coordinating to maintain monitorability, because the architectural definition is relatively clear (operationalized as 'high opaque serial depth'), whereas monitorability evaluation requires a great deal of scientific judgment and leaves room for disagreement and motivated reasoning.
OpenAI announced that it will allow third parties to participate in safety evaluations during the training phase, covering the full pipeline of training, evaluation, and deployment, with partners including METR and Redwood Research.
Redwood Research Chief Scientist Ryan Greenblatt noted that there is currently no good way to understand and supervise the activities and goals of AI collectives.
All Facts (5)
OpenAI announced that it will allow third parties to participate in safety evaluations during the training phase, covering the full pipeline of training, evaluation, and deployment, with partners including METR and Redwood Research.
65%UnverifiedAI safety research organization Redwood Research published an analysis article pointing out that latent-space reasoning architectures represented by COCONUT and full-bandwidth Transformers may greatly weaken the currently most effective AI supervision tool — chain-of-thought (CoT).
50%UnverifiedRedwood Research believes that interpretability techniques such as SAEs, probes, and NLAs will be difficult to sufficiently replace the supervisory function of CoT in the near term.
50%UnverifiedRedwood Research believes that coordinating to avoid latent-space architectures is more feasible than coordinating to maintain monitorability, because the architectural definition is relatively clear (operationalized as 'high opaque serial depth'), whereas monitorability evaluation requires a great deal of scientific judgment and leaves room for disagreement and motivated reasoning.
50%UnverifiedRedwood Research Chief Scientist Ryan Greenblatt noted that there is currently no good way to understand and supervise the activities and goals of AI collectives.
50%