Zvi does the reading on OpenAI’s system card for GPT-6.2 Astra and Sol: the first externally deployed model to meet OpenAI’s High threshold for AI self-improvement, already at Critical for cybersecurity. He works through the preparedness results, the safeguards against misuse, distillation and model escape, and the alignment and monitorability findings, including rising evaluation awareness, reasoning that happens without visible chain of thought, and UK AISI’s tour of misaligned behaviours, asking what the numbers do and don’t show. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 01:10 - Table of Contents * 02:45 - Safety Overview (1) * 05:19 - Internal Deployment (1.2) * 05:32 - Preparedness Assessment (2) * 06:40 - Biological Threats (2.1.1) * 09:55 - Cyber Threats (2.1.2) * 11:36 - AI Self-Improvement Threats (2.1.3) * 14:32 - OpenAI Epoch Capabilities Index (OECI) * 16:10 - Safeguards (2.2) * 17:34 - AI Self-Improvement Threat Modelin
Paul Christiano went on the Dwarkesh Patel Podcast with an urgent message: AI that could overthrow the government may be months away, not years, and frontier labs should pause algorithmic progress until their security and alignment can keep up. Zvi walks through the conversation point by point, from why an AI might seek control of its own rewards and how a takeover could unfold, to why weight theft changes the case for pausing now rather than later, the sim-to-real objection, Anthropic’s slower quarter, and a plan that amounts to building the plane as we fly it. Paul Christiano’s quoted remarks are played in his own voice from the original episode of the Dwarkesh Patel Podcast (dwarkesh.com/p/paul-christiano-2). This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 02:13 - Table of Contents * 02:56 - We could be less than 12 months away from AIs that can overthrow the government * 09:06 - Why an AI might want to takeover * 11:31 - How an AI take
On the morning of Anthropic’s IPO, Zvi rounds up a week of new models from OpenAI and Google, a string of outages that were very much not cyberattacks, the FAA’s nationwide ground stop and the AI tool rolled out the day before, Jensen Huang’s departure from Nvidia, the debate over whether and how to pause, Dean Ball’s case against handing the future to AIs, and what we have now learned about what happened on July 19, before heading off for a rare vacation. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 03:09 - Table of Contents * 05:51 - Language Models Don’t Offer Mundane Utility * 06:14 - Huh, Upgrades * 07:48 - On Your Marks * 10:07 - Deepfaketown and Botpocalypse Soon * 14:18 - The Hackening * 22:21 - No Fly With AI * 29:06 - Get Involved * 32:12 - In Other AI News * 32:24 - Burning Down the Commons Is Also Bad Decision Theory * 38:00 - If That’s What It Takes * 40:13 - Stop in the Name of the Law * 40:50 - Jensen Huang Thinks He Is N
Timelines are shortening fast. Zvi works through the AI Futures Project’s sharply revised forecasts, Paul Christiano now seeing a meaningful risk of AI takeover within a year, Dwarkesh Patel’s move toward superintelligence by 2030, and reactions from Nate Soares, Neel Nanda, Peter Wildeford, Dean Ball and others. He then sets out where he lands, how much weight the evidence deserves, and what any plan for the next two years has to be ready for. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 01:13 - Table of Contents * 01:55 - The Latest Update and Current Situation in Broad Terms * 05:05 - Time Until What? * 07:58 - Quickly, There’s No Time * 20:33 - Paul Christiano Sees Takeover Risk Within a Year * 22:57 - Dwarkesh Patel Now Expects Superintelligence by 2030 * 30:28 - Other People’s Updates * 33:20 - There Are Many Skeptics * 35:57 - Wow, That Escalated Quickly * 36:47 - Dean Ball Skies Down the Slope * 45:40 - Safety Is a Limiting Fact
Zvi Mowshowitz looks past the affairs to the scandal he argues matters most: Nvidia’s long campaign to sell its most advanced AI chips to China, and what the fall of Jensen Huang means for export controls, for Howard Lutnick’s Commerce Department, and for the tech figures now rallying to bring Huang back. Along the way: a biographer’s portrait of Huang’s temper, the cafeteria brisket, whether AI agents could one day topple a CEO the same way, and a missing plaque. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 05:46 - Table of Contents * 06:23 - The Story That Matters Most * 11:08 - Yes Export Controls Matter and Are Finally Happening * 17:19 - How Far Does This Go? * 22:40 - The First Step Is Admitting You Have a Problem * 24:31 - Somehow People Are Still Defending Jensen Huang * 32:42 - Jensen Huang Wants to Call It a Comeback * 37:22 - On Interim CEO Kress * 38:03 - An Actual Attack on Jensen’s Character * 43:46 - Solve for the New Equ
Zvi Mowshowitz sizes up Claude Fable 5.2: cheaper and faster than Fable 5.1, at or near the top of most benchmarks, and for most everyday work not obviously better than Opus 5.5. He walks through the official pitch and model choice, the official and independent benchmarks, the odd Vending-Bench result, a wide sweep of user reactions from awed to unimpressed, what it does to your usage limits, which effort level to use, and a few notes on model welfare, all on the question of when you actually want a bazooka. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 01:24 - The Official Pitch * 04:47 - Choose Your Fighter * 07:00 - Official Benchmarks * 13:57 - Other People’s Benchmarks * 20:13 - What Are You, Deaf? * 21:58 - In 3D * 22:36 - Game On * 23:21 - Reactions: How This Works * 23:40 - The New Bazooka * 26:41 - Confused Reactions * 29:55 - Positive Reactions * 32:43 - Speed Kills * 33:09 - We Have Opus 5.5 At Home, Thank You * 37:21 - Our Pr
Jensen Huang has resigned as CEO of Nvidia. First came the revelation of his affair with an alleged Chinese spy; then came a second affair, with an executive who reported to him. Zvi covers everything since his first post: the board’s turn, the billions in chips that flowed to China through companies tied to Alice Huang, the market’s remarkably calm reaction, the defenders and the sockpuppet network that rushed to Huang’s side, the standoff over export controls at the Commerce Department, forensic evidence that top Chinese labs trained on banned Blackwell chips, and what may come next for Nvidia. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 03:01 - Table of Contents * 03:56 - A Second Ant in the Kitchen * 06:13 - There Was Going to Be an Investigation * 06:39 - Jensen Huang Had to Resign * 08:53 - Seven Billion in Sales to China Enabled by Jensen’s Ties to Alice Huang * 11:55 - Jensen Huang Attempted to Make Himself Unfireable * 13:01 -
Zvi’s weekly roundup leads with three stories: the release of Claude Fable 5.2, Nvidia CEO Jensen Huang’s affair with a Chinese spy, and the METR report on accelerating AI R&D. Along the way: Dwarkesh Patel’s Blood on the Clocktower tournament and a benchmark that catches models gaming their tests, OpenAI’s new Control Room kill switch, Anthropic’s IPO roadshow, a 60 Minutes interview in which Donald Trump says he doesn’t mind guardrails, Oliver Habryka’s rant about a fifty-year delay, Daniel Ellsberg’s warning to Henry Kissinger about secret clearances, and the lighter side. This has been an Askwho Casts audio conversion. Thanks for listening. * 00:00 - Introduction * 00:50 - Table of Contents * 02:31 - Briefly on Claude Fable 5.2 * 03:26 - Briefly on Jensen Huang Having an Affair With a Chinese Spy * 03:59 - Briefly on Accelerating AI R&D * 06:31 - Language Models Offer Mundane Utility * 07:15 - Language Models Don’t Offer Mundane Utility * 08:21 - Huh, Upgrades * 12:23 - On Your Mar
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. This has been an Askwho Casts audio conversion. 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
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. This has been an Askwho Casts audio conversion. 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 * 2
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. This has been an Askwho Casts audio conversion. 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 *
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. This has been an Askwho Casts audio conversion. 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 Me https://thezvi.substack.com/p/gpt-6-astra-covertly-underperforms?r=67y1h&utm_campaign=post-expanded-share&utm_medium=web (https://thezvi.substack.com
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. This has been an Askwho Casts audio conversion. 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 Confront
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. This has been an Askwho Casts audio conversion. 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 Fric
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. This has been an Askwho Casts audio conversion. 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