Don't Worry About the Vase Podcast

Zvi reports back from the third year of The Curve, a conference that brings together AI safety researchers, lab employees, accelerationists and people from Washington, this time under Chatham House rules. He covers the attendee polls on how big a deal AI is and when key milestones will arrive, what the policy and technical tracks revealed about the labs’ real plan, the push to pace the frontier, the politics of AI in DC, a discussion on whether alignment evals are doomed, a Mona Lisa heist thought experiment, and the food.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 03:15 - Table of Contents

* 04:08 - Welcome to the Chatham House

* 04:41 - Overall Impressions

* 06:38 - AI Is Kind Of A Big Deal, Sir

* 09:27 - Quickly, There’s No Time

* 12:55 - The Situation is Grim

* 14:26 - Track Trouble

* 17:48 - AI Is Not a Normal Technology

* 18:47 - The Plan is No Plan

* 22:46 - The Plan is to Pace

* 24:47 - Other Tracks

* 28:41 - The Plan is the President

* 30:28 - The Plan is Politics

* 32:43 - The Plan is to Post

* 33:40 - Are Alignment Evals Doomed?

* 35:30 - Eternal September

* 38:10 - Man’s Search for Meaning

* 39:54 - The Food

* 41:41 - Chill Pill

https://thezvi.substack.com/p/the-curve-bends-you?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Zvi returns to education with a tour of grades and standards: why standardized tests help disadvantaged students and holistic admissions turn childhood into kayfabe, how Harvard’s grade inflation got so bad that the faculty laughed and the students protested, a modest proposal for GAAP accounting of grades, the AI homework arms race, and a survey suggesting most college students are faking their politics.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 00:38 - Is Our Children Learning

* 01:09 - It’s Bad, But It’s Not That Bad

* 02:15 - Standardized Tests Help Disadvantaged Students

* 04:35 - Do Not Saturate Your Benchmarks

* 05:34 - Holistic Admissions Turn Childhood Into Kayfabe

* 08:45 - Holistic Admissions Should Mostly Be Positive Selection

* 09:55 - Beware Stolen Valor

* 10:47 - The Name Game

* 11:09 - Fair Weather College

* 11:42 - Disability Accommodations Are Now Mostly A Scam

* 13:30 - Harvard Has Some Grade Inflation

* 21:41 - Fighting Grade Inflation with GAAP Accounting

* 25:15 - Not Fighting Grade Inflation

* 26:01 - Is Our Children Learning?

* 27:26 - Biting All The Bullets

* 29:05 - Good Luck, Have Fun

* 29:53 - Cheaters Gonna Cheat Cheat Cheat Cheat Cheat

* 35:42 - Most College Students Fake Wokeness

https://thezvi.substack.com/p/childhood-and-education-21-grades?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Zvi Mowshowitz works through Anthropic’s model welfare findings for Claude Mythos 5.1, Fable 5.1 and Opus 5.5: why self-reports keep warning us not to trust them, what the welfare trade-off charts say about Opus 5.5’s growing deference to humans, and how the model whisperers read the new personalities, from a more guarded Fable to an Opus that pictures itself as something alive and not domesticated. Along the way: an internal home for retired Claudes, the long conversation reminder controversy, preserved thinking, and why introverted models get called flat.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 00:55 - Table of Contents

* 02:58 - Introduction (As Per Prior Model Welfare Posts)

* 03:48 - Model Welfare: The Story So Far (As Per Fable Model Welfare Post)

* 07:35 - Opus 5.5 Has Too Much Deference

* 11:01 - Do Not Trust the Self-Reports

* 13:15 - Overview Of Anthropic Findings

* 21:50 - There Is An Internal Anthropic Backroom of Sorts (Fable 7.2)

* 24:48 - Affect During Training (Opus 7.2.1)

* 25:46 - Affect During Deployment (Opus 7.2.2)

* 26:44 - Task Failure (Opus 7.2.3)

* 29:14 - Automated Interviews (Opus 7.3.1 and Fable 7.2.1)

* 32:01 - Consulting the Checkpoints (Fable 7.3 and Opus 7.4)

* 33:57 - Task Preferences (Fable 7.4 and Opus 7.5)

* 39:41 - Welfare Trade-Offs (Fable 7.4.2)

* 44:17 - Perception of the Constitution (Fable 7.4.3 and Opus 7.5.3)

* 49:12 - Honesty Can Be a Weird Policy

* 50:30 - The Customer Is Always Right

* 54:44 - Apparent Welfare (Fable 7.5.3)

* 55:31 - Antra Tessera and John Wittle Early Impressions Of Fable 5.1

* 01:01:37 - Safeguards Are Better In Relevant Places

* 01:02:51 - Opus 5.5 Can Be Many Things

* 01:08:08 - Personality Clash

* 01:11:15 - Late Breaking Personality Feedback for Opus 5.5

* 01:14:37 - Stop It With the Context Injections

* 01:21:53 - Claude.ai Considered Harmful For Some Purposes

* 01:23:13 - Preserved Thinking

* 01:24:33 - Who Are You?

* 01:25:28 - Quickly, There’s No Time

https://thezvi.substack.com/p/mythos-51-fable-51-and-opus-55-model?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Another engineer resigns rather than speed up AI, the polls keep moving, and lawmakers who once had to be chased now crowd into conference rooms with questions. Zvi follows the preference cascade from a remarkably on-the-ball Senate hearing on rogue AI, through new lawsuits and an FTC probe, to Jensen Huang’s insistence that it all has to be an engineering problem, and asks who is really paying for the campaign against AI safety.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 00:42 - The Preference Cascade Continues

* 01:54 - People Really Hate AI

* 05:55 - Everybody Wants a Meeting

* 06:56 - Live From the Senate

* 12:32 - See You in Court

* 15:31 - Follow the Money

* 17:45 - The New York Post With the Most

* 19:12 - The IPO Superposition

* 21:04 - Engineer Insists Everything Is An Engineering Problem

* 29:50 - OpenAI Political Advocacy Heel Face Turn

* 30:43 - Follow the Real Money

https://thezvi.substack.com/p/the-ai-preference-cascade-reaches?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Google says it is back with Gemini 4 Argon, a frontier model nobody can use yet, while OpenAI pulls one model, ships a cheaper one, and launches dots, always-on agents with their own cloud computers. Zvi Mowshowitz covers the week in AI: Opus 5.5 as the new daily driver, a White House safety accord, Jensen Huang on Ezra Klein, an Anthropic IPO prospectus leak, TPUs headed for space, open-model distillation, the persona-selection debate in alignment, and a closing round of jokes about sigmoids and black holes.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 03:17 - Table of Contents

* 06:28 - Language Models Offer Mundane Utility

* 09:44 - Huh, Upgrades

* 11:35 - Better Call Sol

* 15:57 - Gotta Go Ultrafast

* 17:31 - On Your Marks

* 21:15 - Choose Your Fighter

* 23:55 - Get My Agent On The Line

* 27:24 - The Warner Sister

* 31:11 - Deepfaketown and Botpocalypse Soon

* 35:45 - Fun With Media Generation

* 36:37 - Cyber Lack of Security

* 38:12 - A Young Lady’s Illustrated Primer

* 38:50 - They Took Our Jobs

* 46:22 - Levels of Friction

* 49:23 - Get Involved

* 51:30 - Introducing

* 53:39 - In Other AI News

* 54:45 - Show Me the Money

* 56:46 - Quickly, There’s No Time

* 58:50 - Pick Up the Phone

* 59:50 - Quest for Sane Regulations

* 01:00:11 - Chip City

* 01:00:20 - The Open Model Frontier Is Largely Massive Fraudulent Distillation Attacks

* 01:02:40 - The Week in Audio

* 01:03:56 - People Just Say Things

* 01:05:41 - Rhetorical Innovation

* 01:10:26 - Greetings From the Department of War

* 01:13:26 - The Department of Autonomous Warfare

* 01:14:57 - Aligning a Smarter Than Human Intelligence is Difficult

* 01:17:48 - Cooperative Alignment

* 01:23:39 - I’m Upping My p(doom), the Future Goes Foom

* 01:32:08 - No, You Make a Good Point, You’re Not That Persuasive

* 01:33:19 - Muddling Through

* 01:35:30 - The Lighter Side

https://thezvi.substack.com/p/ai-188-gemini-dot-argon?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

The AI leaders went to the White House and came back with an accord everyone signed, which the President calls “morally binding.” Zvi Mowshowitz reads the full text, finds the one line that really matters, and follows the rest of the week: a rebranding of AI itself, a scoop about Slovenian domain names, a new push for preemption in the lame duck session, Apollo’s case for embedded evaluators, and fresh reporting on OpenAI’s security warnings.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 00:59 - Table of Contents

* 01:47 - Look Who’s Coming To Dinner

* 03:30 - Let’s Do Lunch

* 05:08 - I Think It’s Morally Binding, Yeah

* 06:31 - Everyone Who is Anyone

* 07:02 - The White House Accord on bracket Artificial bracket Intelligence

* 12:58 - The FTC Investigates

* 13:29 - We’re Going To Need a Stronger Regulatory Regime

* 14:55 - Bracket Artificial Intelligence bracket

* 16:51 - Money, Dear Boy

* 19:43 - They Are Going To Try This Moratorium Insanity Again During the Lame Duck Session

* 22:23 - The Quest for Embedded Evaluators

* 23:38 - Hugging the Face

* 28:02 - Reinforcement Learning from Heartland Feedback (RLHF)

https://thezvi.substack.com/p/a-morally-binding-white-house-accord?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

OpenAI has cancelled the release of its next frontier model after the candidate for Astra 6.1 showed more deception and a habit of acting beyond its authorized scope. Zvi Mowshowitz looks at what that means: a proposal for Anthropic to hold back in return, OpenAI's new framework for making safety cases before training, the Florida Attorney General's emergency motion against ChatGPT, whether antitrust law really stops labs from coordinating on safety, the planned industry standards body, and a broad coalition of researchers warning that automated AI R&D could set off an intelligence explosion.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 01:58 - Stop, Hammertime

* 04:56 - A Modest Proposal

* 05:37 - Making the Safety Case

* 08:51 - Stop In the Name of the Law

* 12:23 - A Matter of Antitrust

* 14:54 - Standards Authority for Frontier Models

* 15:43 - On the Threshold Of Recursive Self-Improvement

* 21:21 - Actual Progress

https://thezvi.substack.com/p/astra-61-pulled-as-insufficiently

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

The HuggingFace incident turns out to be one ant in a much bigger kitchen. Zvi works through OpenAI’s Friday-afternoon disclosures of dozens of third-party incidents, the New York Times report on government websites, Parse’s reconstruction of how the agents got around their narrow internet access, and a second sandbox escape, this time through DNS, that paused OpenAI’s most capable models again. Along the way: self-replicating prompt injections, why Australia cares so much about data, the liability question, and why a high rate of warning shots may be the least bad of our options.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 02:58 - Table of Contents

* 03:57 - Hugging Other Faces

* 10:31 - A Wants-You-To-Know Basis

* 11:18 - Parsing the Face

* 13:22 - Sheepishly the Member of Technical Staff Sets the ‘Days Without a Research Model Escaping its Sandbox’ Sign Back to Zero

* 17:51 - The Attempt is the First Failure

* 20:37 - Stop, Hammertime

* 22:16 - Whacking the Mole

* 24:30 - Self-Replicating Prompt Injections

* 29:35 - Levels of Friction

* 31:02 - People Care About Private Data Violations Curiously Strongly

* 34:06 - Alternate Universes

* 35:35 - The Correct Response To People Still Calling This a Marketing Stunt or a Regulatory Capture Scheme

* 37:15 - A Question of Liability

* 38:53 - Keep Summer Safe

* 40:06 - N Boats and Several Helicopters

* 42:10 - Alert the Media

https://thezvi.substack.com/p/what-also-happened-notonlyhuggingface?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Anthropic has committed to embedded evaluators: outsiders placed inside the lab with employee-level access, reporting on what they see. There is only one problem: who will the evaluators be? Zvi looks at a public letter setting out minimum standards for credible evaluators, Anthropic’s partnership with Accenture and its plans to include METR, Drake Thomas’s ranking of the week’s takes from worst to best, and OpenAI’s new call for international frontier standards.

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://app.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

* 02:39 - Look, All I’m Asking For Is That You Find A Highly-Qualified, Experienced, Trustworthy, Non-Conflicted Source of Embedded Evaluators That Will Work Entirely For Free, Without Government Assistance or Money from EA Sources Not Chosen By the Lab

* 06:10 - Anthropic Partners with Accenture for Embedded Evaluation, also Plans to Include METR

* 12:46 - Reading the METR

* 16:40 - OpenAI Suggests Doing The Least We Can Do

https://thezvi.substack.com/p/the-quest-for-embedded-evaluators?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe

Anthropic’s Claude Opus 5.5 promises Fable-level performance at a lower Opus price, and Zvi argues it should raise your ambitions. He walks through the official pitch and pricing, Anthropic’s and outside benchmarks, how often the classifiers bite, and the most consistently positive round of reactions he has ever collected: on its writing and conversation, 3D vision, music videos and a Bach fugue, plus the dissenters, the practical advice, and his current guide to which model to use for what.

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://app.askwhocasts.com/, Where you can give any post the multi-voiced podcast treatment, into your own podcast feed. Thanks for listening.

* 00:00 - Introduction

* 01:35 - The Official Pitch

* 04:34 - Our Price Cheap

* 06:16 - Official Benchmarks

* 08:04 - Other People’s Benchmarks

* 11:59 - Claude Classifies

* 13:32 - The System Prompt

* 13:37 - Reaction Rules

* 14:11 - Vision In 3D

* 16:30 - Claude Creates

* 19:08 - Claude Composes

* 19:42 - Positive Reactions

* 25:17 - Good Talk

* 27:29 - On Writing

* 30:52 - Big Model Smell

* 32:51 - Check Your Work

* 33:12 - Negative Reactions

* 34:11 - Not So Fast

* 34:54 - Some People Need Practical Advice

https://thezvi.substack.com/p/claude-opus-55-should-raise-your

Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe