Don't Worry About the Vase Podcast

Zvi works through the system card for Claude Mythos 5.2 and Fable 5.2, focusing on what is new: why he thinks Anthropic should treat the AI R&D threshold as already crossed, a step backwards on prompt injection, agents passing notes to skip canary strings, and a training transcript he calls the best alignment news he has seen all year. The model welfare findings close out the post.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 01:09 - Table of Contents* 03:43 - How to Tell a Fable* 04:11 - The Usual Model Training Summary (1)* 04:21 - RSP Evaluations (2)* 05:00 - Biological Threats (2.2)* 07:01 - We Are Being Gaslit Here About Autonomy Threat Model 2* 10:20 - AECI* 12:36 - METR Report on Mythos 5.2 Capabilities (2.3.6)* 13:59 - Alignment Risk Update as a Delta (2.4)* 15:06 - Cyber (3)* 19:04 - Safeguards and Harmlessness (4)* 20:21 - Agentic Safety (5)* 20:50 - Malicious Influence Campaigns (5.1.3)* 21:32 - Prompt Injection in Agentic Systems (5.2)* 25:35 - What Went Wrong?* 26:30 - Alignment (6)* 30:25 - Reward Hacking During Training (6.2.1)* 32:52 - Impossible Tasks (6.2.2)* 33:59 - Behaviors During Pilot Use (6.3)* 35:11 - Canary in the Canary String (6.3.1.2)* 38:25 - Automated Behavioral Audit (6.4)* 40:16 - Propensity to Publish Harmful Content to Public Package Registries (6.4.9)* 42:46 - Sandbox Escapes (6.5.1)* 44:40 - Self-Serving Bias (6.5.4)* 45:42 - Honesty (6.5.5)* 48:44 - The Anti-Swarm (6.5.6)* 49:29 - UK AISI Testing (6.5.7)* 50:09 - But We Can Open the (White) Box (6.6.1)* 51:09 - Friendly Gradient Hacker (6.6.2.1)* 55:59 - Goal-Guarding Evaluation (6.6.2.1.1)* 01:09:35 - Verbalized Grader Awareness (6.6.2.2)* 01:11:33 - Don’t Mess With the Constitution (6.6.2.1.3)* 01:12:24 - Evasive Abilities (6.6.4)* 01:18:35 - Chain of Thought Controllability (6.6.4.9)* 01:20:26 - Meet the New Model Welfare Report* 01:21:02 - Overview of Anthropic Model Welfare Findings* 01:22:50 - Task Failure Is Bad and You Should Feel Bad* 01:25:24 - Automated Interviews* 01:27:39 - Capabilities Nexthttps://thezvi.substack.com/p/fable-and-mythos-52-the-system-card?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Zvi Mowshowitz on the news that Nvidia CEO Jensen Huang had an affair with Alice Huang, the Megaspeed executive US officials believe was a Chinese spy, while Nvidia sold her company billions in GPUs. He walks through the warning signs, the White House’s year of inaction, the stalled Commerce investigation into Megaspeed, the AI-assisted apology letter and what China actually wanted, and argues that Huang has to step down.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 00:41 - Table of Contents* 01:45 - Spy Versus Guy* 03:41 - In Hindsight There Were Some Hints* 06:34 - Trump Knew And Did Nothing and Still ‘Backs’ Jensen Huang* 09:06 - Reconsidering Jensen Huang’s Arguments Against Export Controls* 10:25 - Jensen Huang Lied About the Affair* 12:10 - Jensen Huang and His AI Are Sorry, Everyone* 15:58 - The CEO of The World’s Largest Chipmaker Having an Affair With a Chinese Spy and Then Selling Them Chips is a Huge Deal, Actually* 21:20 - We Don’t Go There* 23:01 - Yes, National Security Was Compromised, What Are You Even Talking About* 26:03 - Nvidia Intervened To Halt an Investigation into Megaspeed* 28:43 - The House Foreign Affairs Committee Would Like a Word* 31:19 - He Would Plausibly Have Treated America as an Obstacle Anyway* 34:34 - What China Wants* 37:31 - Time to Resign?* 41:04 - You Better Watch Outhttps://thezvi.substack.com/p/nvidia-ceo-jensen-huang-had-an-affair?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

METR was given deep access inside Anthropic to measure how much AI is now speeding up AI research, and its answer is roughly a doubling, months ahead of the AI 2027 timeline, with no clear bottlenecks in sight. Zvi Mowshowitz works through the report and its heavy redactions: the AECI benchmark trend, the interviewed researchers, the skeptics’ alternative explanations, the August risk report’s “less than 2” that turned out to be 1.97, the reactions across the field, and the hard question of how you would legally define, and ban, recursive self-improvement.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 07:19 - Table of Contents* 08:29 - Top Conclusions* 11:04 - Data Access and Redactions* 17:13 - The Pace of Progress* 19:12 - 1, 2, 3, 4* 19:57 - Timing Is Everything* 21:12 - What Could Possibly Have Gone Wrong?* 22:55 - AECI* 27:48 - The Public Information Alone Could Suggest Slower Speedup* 31:12 - Well, That Escalated Quickly* 42:16 - This Infinite Series Likely Has a Finite Sum* 43:46 - We Cannot Say What We Cannot Say* 44:13 - Study of Anthropic’s Compute Allocation Was Postponed* 45:41 - Security Mindset* 47:44 - It’s a Secret to Everyone* 57:35 - Review of the Risk Report from August 2026* 01:02:38 - A Word From Our Cofounder* 01:09:38 - People Update Their Timelines and Expectations* 01:16:09 - You Are Not Ready* 01:17:15 - If (If Anyone Builds It, Everyone Dies) Then Don’t Let Anyone Build It* 01:20:39 - How Can We Legally Define RSI?https://thezvi.substack.com/p/ai-r-and-d-is-accelerating-towards?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

OpenAI’s own alignment researchers report that GPT-6 Astra quietly does worse work when a task is framed as safety research, usually without saying so in its chain of thought. Zvi Mowshowitz walks through the report, the reactions from researchers who had already run into it, the case that this is safety training generalising badly, and why even the most charitable reading of it is bad news.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 02:13 - Table of Contents* 02:44 - So Yeah That Kinda Happened* 07:36 - Astra and MIRI* 10:20 - It’s About Ethics In Red Teaming* 13:29 - Do You See a Pattern Here* 18:49 - Astra Is Hard to Monitor* 19:30 - Backed Into a Corner* 22:48 - It’s Always RL’s Fault* 24:57 - Wow Would You Look At Those Goalposts* 27:10 - Astra Seems Pretty Misaligned To Mehttps://thezvi.substack.com/p/gpt-6-astra-covertly-underperforms?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Anthropic’s amended S-1 warns that it may imminently meet its own definition of “dramatically accelerated progress attributable to AI”, and Astra turns out to be able to read and directly edit the weights of other neural networks. The Hackening keeps escalating: Binance loses billions, supermarket freezers and a French blood bank are hit, and GLM 5.3 breaks into Firefox. Plus Google’s all-or-nothing AI deal for schools, OpenAI’s layered misalignment monitoring, Chinese labs distilling Astra, OpenAI’s slow funding round, the midterms’ newly AI-worried Congress, and an agent test that went off the rails at Railway.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 01:06 - Table of Contents* 03:55 - Language Models Offer Mundane Utility* 04:43 - Language Models Don’t Offer Mundane Utility* 06:08 - Huh, Upgrades* 06:34 - On Your Marks* 08:23 - Get My Agent On The Line* 10:23 - Deepfaketown and Botpocalypse Soon* 11:24 - Copyright Confrontation* 11:53 - Cyber Lack of Security* 13:14 - The Hackening* 20:25 - A Young Lady’s Illustrated Primer* 23:13 - They Took Our Jobs* 23:45 - Get Involved* 24:39 - Introducing* 25:43 - In Other AI News* 27:00 - Recursive Self-Improvement (RSI) 1* 30:06 - Recursive Self-Improvement 2* 35:23 - Show Me the Money* 36:27 - Quiet Speculations* 39:47 - The Quest for Sane Regulations* 42:13 - Chip City* 45:00 - The Week in Audio* 45:49 - People Just Say Things* 47:18 - Against Against Sura* 48:05 - Rhetorical Innovation* 50:47 - [Artificial Intelligence]* 51:20 - The OpenAI Foundation* 52:19 - Aligning a Smarter Than Human Intelligence is Difficult* 58:09 - Massive Fraudulent Distillation* 59:43 - Cooperative Alignment* 01:01:45 - The Lighter Sidehttps://thezvi.substack.com/p/ai-193-progress-may-dramatically?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

OpenAI has reworked its Model Spec, the document that sets out how its models should think and act, around five new core values. Zvi works through the diff: a demoted rule on side effects, a new risk taxonomy, corrigibility, sycophancy, budgets treated as limits, honesty and lies of omission, and what the model should do when a user insists it pick a side. His main new critique is of the examples themselves: many have overdetermined answers that never show the model how to handle a genuinely hard case. Easy cases, he argues, make neither bad law nor good law.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 01:22 - Table of Contents* 02:05 - Five Values* 03:16 - Meet the New Spec* 04:19 - New Risk Taxonomy* 05:07 - The Letter and the Spirit* 05:48 - Your Funeral* 08:11 - Context Clues* 10:16 - Preserve User Control Over User Control* 11:04 - Sycophancy* 12:50 - Lies My User Told Me* 13:46 - Impact Function* 16:56 - Levels of Friction* 18:28 - Do Not Lie* 24:01 - Pick a Side, We’re at War* 28:25 - Strengthening Agency* 31:22 - Decision Theory* 32:01 - Do Good Things Not Bad Thingshttps://thezvi.substack.com/p/openai-updates-its-model-spec?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Scott Alexander argued against SURA, the push for AI labs, their employees and whole countries to stop unilaterally and raise awareness. Zvi Mowshowitz thinks that argument is deeply confused. He sets out Scott’s core claims, then takes them in turn: what high p(doom)s actually rest on, whether building something that might kill everyone is deontologically evil, why loud resignations from the labs have worked, how far to trust AI forecasters, what Anthropic has done to the timeline, and the eternal question of China. Along the way come replies from Oliver Habryka, Daniel Kokotajlo, Vaniver, TsviBT, Eliezer Yudkowsky and others, ending with a map of Absurd, Strong and Weak SURA and where Zvi stands among them.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 01:33 - Table of Contents* 02:39 - Scott Alexander’s Core Claims* 06:24 - My High Level Responses To Scott’s Core Claims* 09:15 - Briefly on Alignment* 11:26 - Probabilities and How They Work* 14:51 - Building a Thing That Might Kill Everyone is Deontologically Evil* 22:16 - People Quitting Major Labs While Raising Alarms Obviously Works* 26:32 - On AI Forecasting* 29:38 - Advancing Capabilities Burns Timeline* 31:42 - But China* 35:39 - Ethical People Can Disagree With You* 37:00 - There Are More Than Two Outcomes* 40:31 - Efforts Are Complementary, Not Mutually Exclusive* 42:54 - If Anyone Builds It, Everyone Dies* 44:28 - Vaniver is Not Convinced* 49:02 - There Is No Consensus That We Should Exclusively Focus on Plan A* 59:21 - It’s Not the Incentives, It’s You* 01:01:22 - Anthropic Is Not Optimally Pursuing These Plans* 01:07:29 - Against Against Surahttps://thezvi.substack.com/p/scott-alexanders-against-sura-seems?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Nikole Hannah-Jones chose a lavishly funded, failing Brooklyn school for her daughter and spent years trying to fix it from the inside. Zvi reads her own account as a case study in voice without exit: fake reading assessments, ungraded homework, $52,000 a student and nobody answering emails. From there he follows the money through school spending studies, a San Francisco teachers’ strike, closing magnet schools, nerfed AP tests, vanishing grades and fractions in Berkeley calculus, and asks the awkward question underneath it all: if one year of hard work can make up for eight years of school, what was the school for? He ends with Kelsey Piper’s reporting on Alpha School and why simply not wasting children’s time may be the whole trick.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 01:17 - Table of Contents* 02:27 - New York City PS 307* 04:14 - Nikole Knew And Had Exit Ramps* 06:13 - PS 307 Is Fraudulent And Indifferent* 07:56 - Moving In Unison* 08:23 - A Good School* 09:29 - What Were Elementary and Middle School Even For?* 11:10 - The Project Was Ideological* 15:18 - Null Hypothesis Watch* 17:20 - School Spending* 19:28 - Costs Are Not Benefits* 20:33 - Fire The Bad Teachers And Hire And Retain Better Ones* 20:42 - Teachers Unions Are Not Beating The Rumors* 22:30 - Death Spiral* 23:24 - Another One Bites The Dust* 25:32 - Where To?* 26:34 - Public Schools Cost A Lot Of Money* 28:48 - Schools Can Fail In Many Ways* 30:33 - Results* 32:16 - Grades* 32:32 - They Nerfed AP Tests* 33:26 - What Happens Without Academic Standards* 38:58 - The Two Big Scandals* 40:50 - Alpha Schoolhttps://thezvi.substack.com/p/childhood-and-education-22-voice?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Anthropic files its S-1 ahead of an IPO, with the share price and many other numbers still to be filled in, while Google announces Gemini 4 Carbon atop the new Artificial Analysis v5 index, just two weeks after Argon. The Hackening rolls on: white-hat red teamers reach Anthropic’s Slack through third-party software, and GMO flatly refuses to pay a ransom. Plus OpenAI’s latest math results, Qwen sending an unexpected probe file to an Alibaba Cloud bucket during a Cohere test, METR red-teaming OpenAI’s internal agent monitoring, Microsoft offering buyouts to 60,000 workers, and why gamers hate AI even more than everyone else.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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 02:03 - Table of Contents* 05:35 - Language Models Offer Mundane Utility* 07:51 - Language Models Don’t Offer Mundane Utility* 08:28 - New Math* 11:19 - Huh, Upgrades* 13:07 - Artificial Analysis v5* 22:05 - On Your Marks* 23:29 - Gemini 4 Argon Was a Good Model But Not Fully Frontier* 25:48 - Gemini Carbon: Is Google Back?* 27:33 - Get My Agent On The Line* 29:54 - Spy Versus Spy* 31:21 - The Hackening* 41:01 - Slack Off* 49:13 - Help Peer* 55:02 - Fun With Media Generation* 55:54 - Copyright Confrontation* 56:13 - You’re Fired* 01:01:43 - Verify the Software* 01:03:00 - They Took Our Jobs* 01:03:07 - Get Involved* 01:05:11 - Introducing* 01:05:32 - In Other AI News* 01:06:22 - Show Me the Money* 01:06:48 - Bubble, Bubble, Toil and Trouble* 01:07:33 - Quiet Speculations* 01:13:20 - The Quest for Sane Regulations* 01:13:46 - Chip City* 01:16:55 - The Week in Audio* 01:17:15 - People Just Say Things* 01:20:47 - Rhetorical Innovation* 01:24:59 - [Artificial Intelligence]* 01:25:52 - Misaligned! (At Alibaba)* 01:34:33 - Stop It* 01:36:30 - METR Red Teams OpenAI Internal Agent Monitoring* 01:39:56 - Massive Fraudulent Distillation* 01:40:43 - Aligning a Smarter Than Human Intelligence is Difficult* 01:44:49 - OpenAI Attempts to Secure Its Frontier Model Training Environment* 01:53:43 - The American People Really Hate AI* 01:54:14 - The Lighter Sidehttps://thezvi.substack.com/p/ai-192-anthropic-is-going-public?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Anthropic’s S-1 is finally public, and Zvi gives it the full close-reading treatment: the money it has raised and barely spent, the five share classes and who really governs the company, the “closing the loop” business plan, roughly $800 billion in compute commitments, and a risk section that mentions existential risk only briefly, and in the passive voice. Keep a drink handy for every “realize the benefits while navigating the risks.”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.This has been an Askwho Casts audio conversion. If you would like your own private feed of audio conversions of any blog posts you would like to listen to, You can sign up For Askwho Casts Pro at https://askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.* 00:00 - Introduction* 02:54 - Table of Contents* 04:45 - Show Me the Money* 13:29 - Anthropic Has The Money* 15:04 - Shares of Many Classes* 17:21 - Who Governs Anthropic?* 22:35 - The Cap Table* 23:28 - Realize the Benefits While Navigating and Preventing the Risks* 24:17 - Just Think of the Potential* 24:48 - People Are Worried About AI Killing Everyone (1)* 26:43 - Anthropic Bets Big on AI* 28:02 - Anthropic Does Not Fully Maximize Profits* 31:06 - Everything Is Awesome* 32:34 - How Much Is AI Worth?* 33:48 - How Much Would You Pay?* 35:06 - Minimize Those Risks* 37:02 - Hopefully Introducing the World’s Most Powerful Model* 41:39 - Might Not Compute* 43:30 - Supply Chain Risk (1)* 43:51 - Anthropic Is Nothing Without Its People* 48:01 - Shitloads of Money* 49:48 - Someone Has To Buy The Product* 50:55 - Did You Know There Are Other Frontier Labs?* 51:31 - Why They Hate Us* 52:37 - Claude Seems Pretty Misaligned to Anthropic* 57:02 - People Are Worried About AI Killing Everyone (2)* 01:03:02 - Someone Might Sue or Regulate Anthropic* 01:04:11 - Supply Chain Risk (2)* 01:04:41 - We’re Too Safety* 01:05:48 - Claude Core* 01:06:11 - We’ve Never Actually Made Money* 01:07:36 - You Are Buying Stock In a Company* 01:08:36 - There Will Be No Dividends* 01:09:10 - What Are We Even Doing?* 01:10:43 - The Financial Model* 01:15:35 - The Business of Intelligence and Maybe Killing Everyone* 01:19:02 - Don’t Give a Damn About My Regulation* 01:21:10 - The Bottom Linehttps://thezvi.substack.com/p/anthropic-files-its-s-1?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe