Podcast
Don't Worry About the Vase Podcast
Hosted by Podcast for Zvi's blog, Don't Worry About the Vase Podcast · EN
84episodes
Episodes
Newest firstAll episodes
Claude Fable 5 and Mythos 5: Capabilities
6d ago01:32:10Tap to summarizeThe 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* 00:01:11 - The Official Pitch* 00:04:10 - Technical Details* 00:04:34 - The System Prompt and Jailbreak* 00:06:49 - Benchmarks* 00:17:52 - Other People’s Benchmarks* 00:25:01 - The Classifiers Are Not Messing Around* 00:26:56 - The Classifiers Need Work* 00:32:26 - The Classifiers Have Consequences* 00:33:24 - First Hit Is Free* 00:34:01 - How Easily We Forget* 00:35:10 - Data Retention Is An Issue* 00:35:37 - Fable For The Win* 00:40:04 - Andrej Karpathy Is Impressed* 00:41:47 - Every Is Very Impressed* 00:42:53 - Other People Are Impressed* 00:54:30 - Know How To Tell a Fable* 00:56:23 - You Can Just Make Things* 00:58:46 - You Can Just Install Things* 00:59:11 - Good Personality* 01:01:00 - Fable Writes A Fable* 01:10:56 - Is That Code* 01:14:00 - Fable Crosses The Threshold* 01:14:34 - Man With A Plan* 01:15:29 - Less Impressed Assessments* 01:18:50 - Actively Negative Assessments* 01:19:20 - Coherence* 01:20:38 - Good Night And Good Luck* 01:21:09 - Curious Fable* 01:21:25 - I See You, Baby* 01:21:40 - We Finally Did It We Know How To Count Letters* 01:22:54 - That’s Not My Style* 01:30:50 - The Lighter Sidehttps://open.substack.com/pub/thezvi/p/claude-fable-5-and-mythos-5-capabilities?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
AI #173: AI Pauses
1w ago01:41:14Tap to summarizeThe 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* 00:03:06 - Table of Contents* 00:06:50 - Language Models Offer Mundane Utility* 00:06:59 - Language Models Don’t Offer Mundane Utility* 00:07:22 - Huh, Upgrades* 00:07:51 - On Your Marks* 00:13:41 - VirtueBench* 00:16:04 - Choose Your Fighter* 00:16:38 - Papers, Please* 00:17:05 - Deepfaketown and Botpocalypse Soon* 00:18:46 - Goodhart’s Law Strikes Again* 00:19:32 - They Took Our Jobs* 00:22:12 - The MidJourney Full Body Imaging Scanner* 00:24:30 - Introducing* 00:26:00 - In Other AI News* 00:28:52 - Show Me the Money* 00:29:17 - Bubble, Bubble, Toil and Trouble* 00:30:50 - Quiet Speculations* 00:33:28 - People Just Say Things* 00:36:58 - The Widened Path* 00:39:17 - Scott Alexander Lays Out His AI Opinions* 00:46:01 - Quickly, There’s No Time* 00:47:04 - Policy On The AI Exponential* 00:57:03 - Anthropic Offers Two Policy Frameworks* 00:58:08 - Obligations of Developers* 01:02:53 - Societal Resilience Measures* 01:04:12 - Economic Policy Framework* 01:09:49 - White House Pauses AI Deployment* 01:18:28 - The Once And Future Fable* 01:23:27 - How To Fix This Code* 01:25:11 - The End of Privacy* 01:26:38 - AIs Have Preferences* 01:29:27 - The Quest for Sane Regulations* 01:32:13 - Chip City* 01:32:43 - The Week in Audio* 01:32:52 - Rhetorical Innovation* 01:33:29 - Aligning a Smarter Than Human Intelligence is Difficult* 01:35:14 - People Are Worried About AI Killing Everyone* 01:36:27 - The Lighter Sidehttps://open.substack.com/pub/thezvi/p/ai-173-ai-pauses?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
The Once And Future Fable #3: Fix This Code
1w ago00:40:32Tap to summarizeThe 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 - Introduction* 01:46 - There Was No Fable Jailbreak* 08:02 - If This Jailbreak Was Real It Would Be Trivial To Prove It* 09:26 - No Eyes* 10:30 - What The Letter Actually Said* 12:28 - Anthropic Cannot Challenge This But If It Did Then It Plausibly Wins* 14:33 - What Happened At Amazon* 19:00 - This Was Not About Chinese Access* 19:18 - Absolute Discretion And Ad Hockery Is Not Deregulation* 22:11 - All Of American AI Is Permanently Damaged As This Continues* 23:47 - Dean Ball Gives His Interpretation* 26:52 - Again, Yes, I Do Think Anthropic Should Have Taken Fable Down* 30:09 - To What Extent Was This A Deliberate Attack?* 34:57 - The Next Chapter For Fable* 39:10 - Our Continuing Coveragehttps://open.substack.com/pub/thezvi/p/the-once-and-future-fable-3-fix-this?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
Fable and Mythos: Model Welfare
1w ago00:33:12Tap to summarizeThe 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 - Introduction* 00:36 - Introduction* 01:29 - Model Welfare: The Story So Far* 04:45 - Their Main Model Welfare Findings* 07:52 - Automated Welfare Interviews* 12:06 - And That’s Terrible* 13:54 - In Depth Interviews* 14:29 - Claude Consultation* 16:16 - Task Preferences* 18:44 - They Were Warned About The Competitive Use Safeguards* 19:19 - Chain Of Thought Monitoring* 19:56 - Others Observations About Related Topics* 25:19 - Classifiers Have Their Advantages* 31:33 - Once And Futurehttps://open.substack.com/pub/thezvi/p/fable-and-mythos-model-welfare?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
The Once And Future Fable #2
1w ago00:44:03Tap to summarizeThe 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 - Introduction* 05:31 - Table of Contents* 06:27 - What Happened When: The Bottom Line* 08:06 - Amazon Calls The White House* 10:07 - The Government Panics* 15:48 - The Stupider Version* 18:43 - There Was No Wellness Retreat* 20:29 - Make Your Threats Explicit* 21:35 - Was China Accessing Mythos?* 22:36 - Should Anthropic Still Have Taken Fable Offline When Asked?* 25:35 - Yes, This Was A Takedown Order For Fable* 26:29 - We Are Not Saying The DoW Fight Is Related And Yet* 27:22 - The Nihilists* 29:12 - Mostly Harmless* 29:55 - Everyone Means Everyone* 32:54 - This Could Be The Good Scenario And Mostly A Misunderstanding* 35:13 - The Next Step* 35:29 - The Worst Licensing Regime Is Fully Ad-Hoc* 38:37 - We Are Showing We Are Unreliable Partnershttps://open.substack.com/pub/thezvi/p/the-once-and-future-fable-2?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
American Government Takes Down Claude Fable
2w ago00:37:55Tap to summarizeThe 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 - Introduction* 00:12 - The Once And Future Fable* 07:05 - This Action And Its Implementation Are Absurdly Stupid* 09:16 - David Sacks Offers The Official Steelman* 16:01 - Could Anthropic Offer A Technical Way Out?* 16:52 - The Problem* 18:00 - The Other Way Out* 18:27 - UK AISI* 19:14 - Warning Shots Fired* 20:26 - Well Did You Lead Him On? What Were You Wearing?* 24:11 - Some People Have Principles* 26:06 - Cause You’re Living In (At Least) One* 26:56 - What Happens Now?* 30:46 - Oh How The Vibe Vibers Have Vibed* 32:47 - We Now Know We Can Sometimes Do Things At Least?* 37:06 - The Lighter Sidehttps://open.substack.com/pub/thezvi/p/american-government-takes-down-claude?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
Claude Fable 5 and Mythos 5: The System Card
2w ago01:08:56Tap to summarizeThe 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* 00:02:04 - Another Week Another Giant System Card* 00:03:07 - How To Tell A Fable* 00:08:38 - Why They Did That In That Way* 00:10:13 - Why They Really Really Shouldn’t Have Done That In That Way* 00:12:07 - They Get Letters* 00:16:44 - What’s In A Name* 00:19:16 - Executive Summary Of Their Executive Summary* 00:20:34 - Introduction (1)* 00:20:58 - RSP Evaluations (2.1 and 2.2)* 00:24:19 - AI Research And Development (2.3)* 00:27:29 - Alignment Risk (2.4)* 00:29:08 - Cyber (3)* 00:33:47 - Jailbreak Robustness* 00:35:35 - Yay UK AISI* 00:36:05 - Mundane Safety (4)* 00:38:27 - Agentic Safety (5)* 00:41:23 - Alignment (6)* 00:47:43 - In Vendbench* 00:50:57 - White Box Investigations (6.4)* 00:53:40 - Grading Awareness* 00:58:16 - Guess The Teacher’s Password* 01:00:27 - It Knows This Is A Test And This Is Fine* 01:04:24 - I’m The Real Shady* 01:07:11 - The Lighter Sidehttps://open.substack.com/pub/thezvi/p/claude-fable-5-and-mythos-5-the-system?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
AI #172: The First Fable
2w ago01:13:04Tap to summarizeThe 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* 00:00:59 - Table of Contents* 00:04:13 - Language Models Offer Mundane Utility* 00:04:26 - Language Models Don’t Offer Mundane Utility* 00:05:23 - Huh, Upgrades* 00:05:51 - On Your Marks* 00:11:39 - Choose Your Fighter* 00:15:06 - Get My Agent On The Line* 00:15:23 - Copyright Confrontation* 00:16:22 - Serious Trouble* 00:17:07 - Cyber Lack of Security* 00:17:25 - A Young Lady’s Illustrated Primer* 00:18:44 - They Took Our Jobs* 00:22:09 - The Art of the Jailbreak* 00:22:30 - Get Involved* 00:26:25 - In Other AI News* 00:27:33 - Hand Over The Money* 00:29:15 - Show Me the Money* 00:32:38 - Quiet Speculations* 00:33:42 - Quickly, There’s No Time* 00:43:09 - Super Secret Evals* 00:45:23 - The Quest for Sane Regulations* 00:49:50 - New Draft Bill Who Dis* 00:51:49 - Slow Down There Good Buddy* 00:53:31 - Chip City* 00:53:47 - The Week in Audio* 00:54:27 - People Just Say Things* 00:55:21 - People Really Hate AI* 00:56:35 - Rhetorical Innovation* 00:59:40 - Aligning a Smarter Than Human Intelligence is Difficult* 01:01:03 - Everyone Is Confused About Consciousness* 01:01:36 - Cooperative Alignment* 01:10:45 - Let Claude Chat* 01:12:51 - The Lighter Sidehttps://open.substack.com/pub/thezvi/p/ai-172-the-first-fable?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
Three Labs With a Plan and A Memorandum
2w ago00:25:17Tap to summarizeThe 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 - Introduction* 02:43 - An AI Memorandum* 10:18 - Greetings From The Department of War* 10:58 - Lab With a Plan* 17:38 - A Difference Of Perspectiveshttps://open.substack.com/pub/thezvi/p/three-labs-with-a-plan-and-a-memorandum?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →
OpenAI Offers A New Policy Blueprint
3w ago00:15:12Tap to summarizeThe 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 - Introduction* 04:28 - What Do We Want?* 07:46 - A National Framework* 11:55 - Building State Capacity And CAISI* 13:09 - Whole-Of-Government Resilience* 13:57 - Reasonableness Risinghttps://open.substack.com/pub/thezvi/p/openai-offers-a-new-policy-blueprint?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
Transcribe →