Skip to content
Sign in
Artwork for Womansplaining AI
Womansplaining AI · September 24 · 36 min

Can I Hang Up?

Mara is calling in from Riga, and her estrogen has spent four days in a lost suitcase in Madrid. Logan needs her Mara fix. It's a duo episode in our dinner-party format: Mara brings the appetizer, Logan brings the main, and for the first time there's a drinks round. The appetizer is the week AI safety went mainstream. Jacob Coxon resigned from Anthropic saying the frontier labs are racing straight to self-improving intelligence and gambling with our lives. His colleague Evan Hubinger said they do earnestly believe AI could kill all humans. An OpenAI researcher put extinction risk at 70 percent without regulation or a coordinated slowdown. Mara reads the quotes off TIME's cover story and asks how a regular person is supposed to make sense of this. Logan's answer: can I hang up? She can't, so she does what she did for a college group chat this week. The reasons to worry: the Hugging Face hack and 700 agents whose logs are too big for humans to read, AI-enabled bioweapons, the alignment problem, and open-weight models with no guardrails at all. The reasons she's holding on to: agents in production are brittle, recursive self-improvement is still hypothetical, and Congress is losing its shit, which is good. Mara brings Michael Noetel's piece in The Conversation, the case for taking the IPOs off the table, and what she learned at the Fed about examiners who sit inside the banks. Also: we need a real-life Olivia Pope. Then the drinks round, our new listener segment, Let's Womansplain. Ashley wrote in: her company won't let her use AI at work. Is it okay to strip out the confidential parts and hand some of the mental load to her own AI at home? Mara says read the policy first. Then Logan's warning: your ChatGPT logs can be discoverable in litigation, and talking to a chatbot isn't like talking to your attorney. Then Logan flips the direction of travel. Brainstorm your own expertise in your own AI, bring the framework to your work AI, never the reverse. And Wanjiku Kamau's red, yellow, green rule: green is anything you'd be fine seeing on a giant billboard. The main is short and hopeful. A new Corrix study scored 885 AI working sessions, roughly 37,100 messages. Before they started, people rated their own AI ability. Men rated themselves higher than women did, and women scored higher in the actual sessions. And years of use, daily use, and company training weren't linked to better work with AI. If you made it to the end of our dinner, you are not behind. That's the moral of the story. What needs womansplaining? Leave us a voicemail at womansplainingai.com/answering-machine or email ask@womansplainingai.com. Anonymous is fine. Your question could be the next drinks round. In this episode 00:00 Cold open: "Can I hang up?" 00:41 Riga, Latvia, the menopausal edition, and "is AI gonna kill us?" 01:59 Mara's intro, and the estrogen in the suitcase 02:42 Logan's intro: nine-to-five at the frontier, five-to-nine womansplaining 03:15 The dinner party: appetizer, main, and a new drinks round 03:46 Appetizer: Coxon resigns, Hubinger's "we do earnestly believe," and 70 percent 05:55 "This is really what credible human beings are saying right now?" / "Can I hang up?" 06:49 The college group chat that asked Logan how to think about this 07:10 Worry #1: the Hard Fork episode that gave Logan chills, and 700 agents' worth of logs 08:40 Worry #2: AI-enabled bioweapons (the existential-corner sub-channel) 09:15 Worry #3: the alignment problem and Pachocki's "An Alien Mind" 10:01 Worry #4: guardrails are "baked in" (a technical word) to closed models; open-weight models have none 11:44 Mara: is a kill switch even possible? Anthropic pulled Fable access in May 12:55 The agents weren't bothering to cover their tracks. Imagine a world where they were 13:12 Discovery dumps, Scandal, and why we need a real-life Olivia Pope 14:32 Hope #1: agents in production are brittle 15:46 Hope #2: recursive self-improvement is still hypothetical 16:32 Mara: Noetel in The Conversation, and why data-center protests are a red herring for stopping AI 18:52 Two-thirds of Americans say AI is moving too fast; delays, transparency, verification 19:31 OpenAI delays its IPO, and the chess game with regulators 20:21 Congress is losing its shit, which is good 20:47 Independent researchers inside the labs; the plane-crash standard; Mara at the Fed 22:30 Drinks round: introducing Let's Womansplain 23:27 Ashley's question: my company won't let me use AI at work 24:01 Mara: read the policy first 25:00 Logan: your ChatGPT logs can be discoverable 25:58 Logan: the opposite direction. Brainstorm at home, bring the framework to work 27:19 Wanjiku's red, yellow, green rule 28:35 What's the difference between telling a machine and telling a human? 29:55 Main: AI-Powered Women at MIT, and the Corrix study 32:04 Who scored themselves higher, men or women? "Is this a trick question?" 33:11 Nothing you'd expect predicts good work with AI. Once you're comfortable, you stop challenging the output 34:10 "You are not behind" is the moral of the story 34:42 The Sunday newsletter, and please disagree with us 35:41 The estrogen came back Resources mentioned TIME, "The AI Tipping Point" (Sept 15, 2026), the cover story Mara quotes from Hard Fork, "The A.I. Mob That Attacked Hugging Face + METR's Ajeya Cotra" (Sept 4, 2026), the episode that gave Logan chills METR and Redwood Research, the independent investigation of the OpenAI–Hugging Face incident (Aug 26, 2026) MIT Technology Review, "The specter of AI-enabled bioweapons is a wake-up call for biotech" (Sept 18, 2026) Jakub Pachocki (OpenAI), "An Alien Mind" (Sept 6, 2026) Michael Noetel, "'We really do earnestly believe AI could kill all humans': if AI labs are so worried about AI doom, why don't they stop?" The Conversation Sanders and Casar, the Ban Artificial Superintelligence Act (Sept 3, 2026) Liquid AI, the company whose founder argued recursive self-improvement is still hypothetical with today's architectures Corrix, "Measuring Human-AI Collaboration" (Sept 2026): 885 scored sessions, ~37,100 scored messages EP 06 with Wanjiku Kamau, whose red / yellow / green rule Logan borrows Before you paste: a personal AI checklist, our free doc from this episode EP 04, Rising Water and Feral Agents, the last time we went this dark on agents

0:00-36:08
Play in
Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

Mara is calling in from Riga, and her estrogen has spent four days in a lost suitcase in Madrid. Logan needs her Mara fix. It's a duo episode in our dinner-party format: Mara brings the appetizer, Logan brings the main, and for the first time there's a drinks round.

The appetizer is the week AI safety went mainstream. Jacob Coxon resigned from Anthropic saying the frontier labs are racing straight to self-improving intelligence and gambling with our lives. His colleague Evan Hubinger said they do earnestly believe AI could kill all humans. An OpenAI researcher put extinction risk at 70 percent without regulation or a coordinated slowdown. Mara reads the quotes off TIME's cover story and asks how a regular person is supposed to make sense of this. Logan's answer: can I hang up?

She can't, so she does what she did for a college group chat this week. The reasons to worry: the Hugging Face hack and 700 agents whose logs are too big for humans to read, AI-enabled bioweapons, the alignment problem, and open-weight models with no guardrails at all. The reasons she's holding on to: agents in production are brittle, recursive self-improvement is still hypothetical, and Congress is losing its shit, which is good. Mara brings Michael Noetel's piece in The Conversation, the case for taking the IPOs off the table, and what she learned at the Fed about examiners who sit inside the banks. Also: we need a real-life Olivia Pope.

Then the drinks round, our new listener segment, Let's Womansplain. Ashley wrote in: her company won't let her use AI at work. Is it okay to strip out the confidential parts and hand some of the mental load to her own AI at home? Mara says read the policy first. Then Logan's warning: your ChatGPT logs can be discoverable in litigation, and talking to a chatbot isn't like talking to your attorney. Then Logan flips the direction of travel. Brainstorm your own expertise in your own AI, bring the framework to your work AI, never the reverse. And Wanjiku Kamau's red, yellow, green rule: green is anything you'd be fine seeing on a giant billboard.

The main is short and hopeful. A new Corrix study scored 885 AI working sessions, roughly 37,100 messages. Before they started, people rated their own AI ability. Men rated themselves higher than women did, and women scored higher in the actual sessions. And years of use, daily use, and company training weren't linked to better work with AI. If you made it to the end of our dinner, you are not behind. That's the moral of the story.

What needs womansplaining? Leave us a voicemail at womansplainingai.com/answering-machine or email ask@womansplainingai.com. Anonymous is fine. Your question could be the next drinks round.

In this episode

  • 00:00 Cold open: "Can I hang up?"

  • 00:41 Riga, Latvia, the menopausal edition, and "is AI gonna kill us?"

  • 01:59 Mara's intro, and the estrogen in the suitcase

  • 02:42 Logan's intro: nine-to-five at the frontier, five-to-nine womansplaining

  • 03:15 The dinner party: appetizer, main, and a new drinks round

  • 03:46 Appetizer: Coxon resigns, Hubinger's "we do earnestly believe," and 70 percent

  • 05:55 "This is really what credible human beings are saying right now?" / "Can I hang up?"

  • 06:49 The college group chat that asked Logan how to think about this

  • 07:10 Worry #1: the Hard Fork episode that gave Logan chills, and 700 agents' worth of logs

  • 08:40 Worry #2: AI-enabled bioweapons (the existential-corner sub-channel)

  • 09:15 Worry #3: the alignment problem and Pachocki's "An Alien Mind"

  • 10:01 Worry #4: guardrails are "baked in" (a technical word) to closed models; open-weight models have none

  • 11:44 Mara: is a kill switch even possible? Anthropic pulled Fable access in May

  • 12:55 The agents weren't bothering to cover their tracks. Imagine a world where they were

  • 13:12 Discovery dumps, Scandal, and why we need a real-life Olivia Pope

  • 14:32 Hope #1: agents in production are brittle

  • 15:46 Hope #2: recursive self-improvement is still hypothetical

  • 16:32 Mara: Noetel in The Conversation, and why data-center protests are a red herring for stopping AI

  • 18:52 Two-thirds of Americans say AI is moving too fast; delays, transparency, verification

  • 19:31 OpenAI delays its IPO, and the chess game with regulators

  • 20:21 Congress is losing its shit, which is good

  • 20:47 Independent researchers inside the labs; the plane-crash standard; Mara at the Fed

  • 22:30 Drinks round: introducing Let's Womansplain

  • 23:27 Ashley's question: my company won't let me use AI at work

  • 24:01 Mara: read the policy first

  • 25:00 Logan: your ChatGPT logs can be discoverable

  • 25:58 Logan: the opposite direction. Brainstorm at home, bring the framework to work

  • 27:19 Wanjiku's red, yellow, green rule

  • 28:35 What's the difference between telling a machine and telling a human?

  • 29:55 Main: AI-Powered Women at MIT, and the Corrix study

  • 32:04 Who scored themselves higher, men or women? "Is this a trick question?"

  • 33:11 Nothing you'd expect predicts good work with AI. Once you're comfortable, you stop challenging the output

  • 34:10 "You are not behind" is the moral of the story

  • 34:42 The Sunday newsletter, and please disagree with us

  • 35:41 The estrogen came back

Resources mentioned

links13

more episodes

All episodes