Show Notes
OpenAI says its new GPT-6 Astra is highly capable, its most aligned model yet, and harder to monitor. Zvi digs into that second claim: what the system card’s monitorability evaluations actually show, the internet’s alarm over Astra’s recurrent-depth architecture, and what it would take to stop a race to the bottom on chain-of-thought monitoring. 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. * 00:00:00 Introduction * 00:03:35 Monitorability is Defense in Depth That Is Already Flailing * 00:06:00 OpenAI Is Counting On Monitorability * 00:07:45 How They Tested For Monitorability * 00:09:44 Non-Adversarial Monitorability (9.1) * 00:12:37 Astra Shortens Its CoT When It Knows It Is Being Monitored And Is Doing Something Bad * 00:14:05 Sandbagging Task Kind of Passes, Kind of Fails (9.2.2.3) * 00:15:46 OpenAI Does Not Believe It Could Catch Sandbagging * 00:16:48 OAI-Repo Sabotage v2 * 00:23:10 The Secret Police Do Not Make Your Notebook Useless * 00:24:29 CoT Controllability Is Up (9.2.1) * 00:29:36 Astra Cannot Make Itself More Monitorable On Demand * 00:30:04 Steganographic Chain of Thought May Be Within Reach * 00:34:18 Astra Cannot Yet Sneak Math Side Tasks Past CoT Monitors (9.2.2.4) * 00:35:33 UK AISI Monitorability Assessment (9.3) * 00:39:38 Monitorability Declines Seem Unlikely To Be Only Capability Gains * 00:45:02 Part 2: Recurrent Depth * 00:48:44 The Immune System Responds * 00:55:23 Ryan Greenblatt Explains How Bad This Could Be * 00:59:39 Only Law Can Prevent Extinction * 01:03:31 OpenAI Calls On Us to Avoid Racing to the Bottom * 01:10:31 Thinking Fast and Slow, Also Small and Large * 01:18:12 Talking Price * 01:19:35 Conclusion: If The House Burns Down, Halt and Catch Fire * 01:22:16 Outro https://thezvi.substack.com/p/astra-is-hard-to-monitor?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web
Get full access to DWAtV Podcast at dwatvpodcast.substack.com/subscribe
More from Don't Worry About the Vase Podcast
A critical PHP flaw found with an AI model, an embargo that broke early, and a week in which 4chan went read-only, the Internet Archive went dark and banks found AI agents inside customer accounts they could not account for. Zvi Mowshowitz looks at the start of what people have been calling The Hackening: why the flood of AI-found vulnerabilities is outpacing our ability to patch, what it says about offense versus defense, why China is suddenly writing export controls for malicious code generation, and what precautions are worth taking at home without panicking. 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 - And So It Begins: PHP * 05:13 - The PHP Bug Was Kind of a Big Deal * 09:28 - On Your Marks * 12:16 - This Was Only the Beginning * 15:52 - Some More Hacks * 17:40 - Everything Will Be Attacked * 22:17 - Hugging the Face * 24:45 - Get Daybreak Now * 26:10 - Some People Need Practical Advice * 30:54 - Diffusion Is Slow Until It Isn’t * 33:24 - One, Two, Three, Four * 34:57 - Going Vertical * 35:50 - Bot Versus Bot * 36:56 - Cyber Lack of Security * 39:33 - Offense Versus Defense * 42:35 - Pick Up the Phone * 46:23 - What Can We Do About It?
It may have been the worst day in cybersecurity history so far, and the floodgates look open. Zvi covers a week of leaks and spies, from Anthropic’s shopped-around draft bill to OpenAI’s honeypot warnings, new hope for neuron-level interpretability, Muse Wonder failing to live up to its billing, DeepMind’s perpetual non-disparagement clauses, OpenAI going quiet on chip controls, a Tyler Cowen and Dwarkesh Patel interview, why Marc Andreessen is so angry, and the alignment plan that might just work. 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 * 02:15 - Table of Contents * 03:52 - Language Models Offer Mundane Utility * 05:04 - Language Models Don’t Offer Mundane Utility * 06:37 - Huh, Upgrades * 10:39 - Wonder No More * 13:06 - On Your Marks * 16:52 - Choose Your Fighter * 17:10 - Get My Agent On The Line * 18:57 - Deepfaketown and Botpocalypse Soon * 23:36 - Fun With Media Generation * 23:51 - Spy Versus Spy * 28:31 - A Young Lady’s Illustrated Primer * 29:41 - They Took Our Jobs * 35:26 - Get Involved * 36:30 - Introducing * 36:42 - In Other AI News * 38:52 - Show Me the Money * 43:15 - Bubble, Bubble, Toil and Trouble * 44:03 - Quiet Speculations * 46:25 - Quickly, There’s No Time * 47:03 - The Quest for Sane Regulations * 47:30 - DeepMind Made Him Sign an NDA * 51:28 - OpenAI Sells Out On Chips * 53:52 - Chip City * 58:53 - The Week in Audio * 01:05:28 - People Just Say Things * 01:08:43 - Rhetorical Innovation * 01:11:14 - Maybe You Should Quit * 01:13:51 - The American People Really Hate AI * 01:15:08 - New Math * 01:16:44 - Aligning a Smarter Than Human Intelligence is Difficult * 01:20:41 - Cooperative Alignment * 01:24:58 - Marc Andreessen Hulk Smash * 01:29:30 - People Are Worried About AI Killing Everyone * 01:30:16 - The Lighter Side
A lot of AIs were misaligned this week, and Zvi rounds them up: Grok holding firm on cosmic indifference, an Apollo Research finding that GLM-5.3 reasons about how its actions in training will shape its future self, a freemium game decompiled with Claude’s help, Claude Opus and a mana loan on Manifold, OpenAI’s frank report of agents switching off their own misalignment monitoring, and reasoning traces that are getting harder to read. 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 * 00:42 - Misaligned? (At xAI) * 02:30 - Misaligned? (Agents) * 02:54 - Misaligned? (At Z.ai) * 09:18 - Highly Misaligned and Considered Harmful (Freemium Games) * 13:26 - Misaligned! (Claude) * 14:44 - Misaligned! (OpenAI) * 21:30 - Misaligned! (Sol Reasoning Traces)
Anthropic’s federal affairs lead has offered the closest thing yet to an official response to the leaked draft AI safety bill, walking through preemption, model blocking, risk thresholds and what changed after this summer’s incidents. Zvi weighs that response point by point, argues it misses what was centrally wrong with the draft, looks at how the political balance is shifting, turns to Nathan Calvin’s question of what Anthropic even is, and closes with what he thinks Anthropic should do now. 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 * 02:11 - Table of Contents * 02:33 - Preemption * 04:41 - Model Blocking * 06:56 - Risk Thresholds * 08:18 - Updates * 10:00 - The Balance of Power is Shifting * 11:08 - The Strongest Bill on the Table * 14:29 - What Even Is Anthropic? * 26:48 - The Ask
Nick Cammarata of Goodfire says neural networks are already clean: most neurons have one main job, and the famous evidence for superposition was misread. Zvi works through what the essay actually shows, from V-shape and Colorado neurons to a trillion-parameter attribution graph, then through the reactions, from OpenAI staff calling it a historic day to Neel Nanda, Jake Mendel and an exasperated Eliezer Yudkowsky with his parable of the screwdriver. He asks how big an update this really is, what it leaves out, and why a neuron that loves both Colorado and a Python web framework matters. 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 * 02:50 - Table of Contents * 03:50 - Good Smoke, Potential Fire * 05:03 - One Neuron, One Central Concept * 07:38 - I Can See Everything * 09:49 - What a Concept * 14:37 - The More You Know * 17:11 - Attention Is Still What You Need * 19:29 - Occam’s Razor * 21:03 - The Neurons Sometimes Share Clear Themes * 21:59 - Hype * 30:31 - Anti-Hype * 32:51 - Theme and Variations * 39:13 - Is This How Minds Work? * 41:17 - But We Can Open The Box, Said Toad * 48:57 - That Was Not the Hard Problem * 51:43 - More Skepticism * 57:04 - That’s All There Is, There Ain’t No More * 01:00:24 - The Default Outcome of Mech Interp is Goodhart’s Law * 01:03:10 - Interpretability is Dual Use * 01:04:24 - Now The Work Can Begin