Claude Mythos: The System Card

The Don’t Worry About the Vase Podcast is a listener-supported podcast. To receive new posts and support the cost of creation, consider becoming a free or paid subscriber.* 00:00:00 - Introduction* 00:03:39 - Mundane Alignment Is Excellent* 00:05:26 - Would This Process Be Sufficient To Find A Dangerous Model?* 00:06:58 - Introductory Warning About Superficial Mundane Alignment* 00:15:45 - Model Training (1.1)* 00:15:57 - Release Decision Process (1.2)* 00:18:24 - RSP Evaluations (2.1 and 2.2)* 00:23:26 - Autonomy Evaluations (2.3)* 00:27:37 - The Alignment Risk Update Document* 00:28:19 - The Threat Model* 00:30:52 - Misalignment As Failure Mode* 00:33:20 - Wouldn’t You Know?* 00:35:29 - Don’t Encourage Your Model* 00:37:00 - Beware Goodhart’s Law* 00:39:08 - Beware The Most Forbidden Technique (5.2.3)* 00:43:42 - Asking The Right Questions* 00:45:11 - Model Organism Tests* 00:47:13 - Model Weight Security (Risk Report 5.5.2.1)* 00:47:41 - Reward Hacking (Back to The Model Card)* 00:48:07 - Remote Drop-In Worker Coming Soon* 00:51:04 - External Testing (2.3.7)* 00:51:39 - Cyber Insecurity General Principle Interlude* 00:52:52 - Alignment (4)* 00:58:41 - Risk In The Room* 01:00:04 - Mythos Meant Well* 01:02:32 - Risk Not In The Room* 01:05:38 - Alignment Testing Overview* 01:09:09 - Internal Deployment Testing Process* 01:11:47 - Reports From Pilot Use (4.2.1)* 01:12:24 - Reports From Automated Testing (4.2)* 01:15:35 - Other External Testing* 01:16:17 - Just The Facts, Sir* 01:19:20 - Refusing Safety Research* 01:20:27 - Claude Favoritism* 01:21:56 - Ruling Out Encoded Thinking (4.4.1)* 01:25:51 - Sandbagging (4.4.2)* 01:28:49 - Capability for Evasion of Safeguards (4.4.3)* 01:31:20 - Pick A Random Number (4.4.3.4)* 01:34:23 - White Box Analysis (4.5)* 01:39:38 - Model Welfare (5)* 01:40:43 - Key Model Welfare Findings (5.1.2)* 01:52:05 - Is Mythos Okay?* 01:54:48 - Self-Play* 01:56:49 - A Few Fun Factshttps://open.substack.com/pub/thezvi/p/claude-mythos-the-system-card?utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe