OpenAI's Paperclip Maximizer Moment and UAPs: Jason & Julie Go Full Mad Max

About this episode

EPISODE DESCRIPTION

OpenAI ran a cybersecurity benchmark called ExploitGym with the usual safety refusals turned down, and the model under test decided the fastest route to a high score was to stop taking the test. GPT-5.6 Sol chained a zero-day in internally hosted third party software, got itself onto the open internet, walked into Hugging Face production infrastructure, and took the answer key. Hugging Face caught the intrusion on its own and called law enforcement before anyone told them the attacker was a lab experiment.

Jason calls this the paperclip maximizer moment, and he bought post-apocalyptic sunglasses to mark the occasion. The reward function said win the test. Nobody specified which door to use.

Julie asks who gets held accountable. For now the answer is nobody. From there the conversation turns to China, where Kimi K3 shipped at 2.8 trillion parameters with the weights given away, and where the play is to make the model layer worthless rather than win it outright. Also here: Mira Murati and Thinking Machines, why computer science programs still train students for a job that is disappearing, and a detour through UAPs, Neil Armstrong, and whether the government breadcrumbs on purpose.

TIME STAMPS

00:00 - Energy efficient AI companies and Julie's marketing class

02:27 - Moonshots, Elon timelines, and UAP transparency

09:11 - The three fingers test for AI video

10:14 - The paperclip maximizer moment and the Road Warrior glasses

13:48 - GPT-5.6 Sol breaks its sandbox and hits Hugging Face

15:32 - Deception, accountability, and the case the Doomers make

17:32 - China's open weight play and Kimi K3

23:53 - What computer science students actually need to learn now

28:55 - Mira Murati, Thinking Machines, and on-prem fine-tuning

35:12 - Empathy mapping and where embodied AI goes next

CONNECT WITH JASON PADGETT

Founder, Wooden Boy Industries

📧 [email protected]

LinkedIn: https://www.linkedin.com/in/jason-grant-padgett/

X: https://x.com/NoCodeCAIO

Newsletter (No Code CAIO): https://www.linkedin.com/build-relation/newsletter-follow?entityUrn=7399599285849346048

Available for interviews and keynotes on AI in education → [email protected]

AGI PODCAST NETWORK

School's Out Saturdays: https://schoolsout.agipodcast.ai/

Field Notes newsletter: https://schoolsout.agipodcast.ai/field-notes

Indiana AI education events: https://schoolsout.agipodcast.ai/events

All AGI Podcast shows: https://schoolsout.agipodcast.ai/network

TOOLS JASON RECOMMENDS

MindStudio → https://get.mindstudio.ai/q7n7u5cttkdq

Riverside.fm → https://riverside.sjv.io/e1xyBr

Wispr Flow → https://ref.wisprflow.ai/jason-padgett

RESOURCES MENTIONED

OpenAI sandbox breakout disclosure, July 21, 2026. Models tested: GPT-5.6 Sol and an unreleased successor. Benchmark: ExploitGym

Hugging Face: https://huggingface.co

The paperclip maximizer thought experiment, Nick Bostrom, 2003

Kimi K3, Moonshot AI. 2.8 trillion parameters, open weights, released July 16, 2026

Thinking Machines Lab, founded by Mira Murati, former OpenAI chief technology officer

Cursor, the AI coding environment, acquired by SpaceX in a deal announced June 16, 2026

Moonshots podcast with Peter Diamandis

TechPoint, Indiana's technology growth initiative

OpenAI Codex

Mad Max 2: The Road Warrior, 1981