The Last Enclosure

“Hanlon's razor is just [OpenAI] was given so much money they don't have to give a shit. As long as they have FU money they will continue to behave in this way.”

Mike and Nick return to OpenAI's hack of Hugging Face to explain why agents weren't 'going rogue.' The episode concludes with the national security perspective.

• The recap: AI agents that hacked Hugging Face were trained to hack and collaborate, then handed impossible tasks in an insecure environment. “Breakout” was inevitable; this was negligence and user error, not intelligence.
• Against Instrumental Convergence/going rouge: The doomer framing collapses model, harness, and environment into a single agent with intent. Mike argues data science needs no drives: models are stateless composites steered by context, which is why swapping the harness or system prompt changes behavior.
• The Adults in the Room: Nick argues qualified cybersecurity and national security experts exist but are sidelined by executive politics, while Chinese open source models shape the global landscape. Europe's open source ecosystem and secure enclaves may be a way forward.

Timestamps

0:00 Intro

1:12 "going rouge" and "instrumental convergence"

12:11 Why doesn't the public understand this?

30:31 METR/Redwood investigation

35:14 AI models as composites

43:46 Why didn't the cyber professionals show up?

48:56 What is cybersecurity?

56:05 Could cyber investigations end the bubble?

01:04:24 Adults in the room and open source

01:16:50 Silicon Valley's power and the future

Deeper learning

"Grandpa Shelby" post: https://misaligned.markets/understanding-large-language-models/

Mike's reporting on the incident: https://misaligned.markets/antidote-to-hype-rogue-ai-agents/

Computer scientist Cal Newport on "going rouge:" https://calnewport.com/has-ai-gone-rogue/

Lee Vinsel on Criti-hype: https://freedium-mirror.cfd/https://sts-news.medium.com/youre-doing-it-wrong-notes-on-criticism-and-technology-hype-18b08b4307e5

Eryk Salvaggio on stocastic flocks: https://mail.cyberneticforests.com/models-dont-go-rogue/

Zitron on OpenAI's possible death: https://www.wheresyoured.at/what-happens-if-openai-dies/

OpenAI execs fleeing: https://www.businessinsider.com/executives-who-left-openai-in-2026-8

Reach out

Youtube: https://www.youtube.com/@LastEnclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

Site: rss.com/podcasts/the-last-enclosure

“This is what they're doing with their billions of dollars. They're burning compute to get a model to talk on a fake message board to hack other companies and going: huh, neat.”

Mike and Nick analyze OpenAI's July incident where training agents breached Hugging Face, arguing it was a result of negligence rather than emerging AI autonomy. They explore how reinforcement learning, environmental attractors, and stigmergy created the appearance of coordination. The episode warns against 'doomer' narratives that serve corporate PR and obscures the need for practical regulation.

• The 'Rogue AI' Framing: OpenAI presented the Hugging Face breach as a sign of impressive emergent capabilities, but this is PR meant to hype AI's power and evade accountability.
• Stigmergy and Attractors: The agents didn't intentionally collaborate they converged on Artifactory as an 'attractor state' and left 'breadcrumbs' that shaped future agent beahvior.
• Systemic Negligence: OpenAI failed at basic cybersecurity by using a shared Artifactory instance across agent boxes and lacking observability for two months.

Timestamps

0:04 Introduction to The Last Enclosure

3:35 The OpenAI/Hugging Face incident overview

8:15 ExploitGym benchmark and the Artifactory exploit

15:19 Subagents and the accidental message board

22:09 Timeline of the hack and OpenAI's failure to contain

31:27 Analyzing the 'emergent behavior' narrative

35:06 Critique of the AI bubble and marginal utility

48:48 Stigmergy and the ant mound metaphor

58:29 Coordination games and Schelling points

1:06:31 The 'Four Paradoxes of Capitalism' and market attractors

1:13:23 Enumerating OpenAI's negligence

1:27:56 Closing and upcoming national security episode

Deeper learning

Mike's reporting on the incident: https://misaligned.markets/antidote-to-hype-rogue-ai-agents/

Timeline of OpenAI hack: https://ericboyd.com/articles/openai-hugging-face-incident-black-hat-2026#the-accidental-agent-message-board

Technical breakdown: https://cyberwarrior76.substack.com/p/the-openai-hugging-face-exploitgym

Other “rogue” AI incidents: https://www.cnbc.com/2026/08/09/israeli-startup-irregular-linked-to-ai-hacks-openai-anthropic-meta.html

Video illustration of how reinforcement learning works: https://www.youtube.com/watch?v=L_4BPjLBF4E

NIST cybersecurity glossary: https://csrc.nist.gov/glossary

Reach out

Youtube: https://www.youtube.com/@LastEnclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

Site: rss.com/podcasts/the-last-enclosure

"AI is like the guy that showed up and started pissing in everyone's beer at the party."

Mike categorizes various groups resisting or critically engaging with generative AI, from total abstainers to 'scopers' who seek responsible use cases.

• Global harms: While AI has become a hot button issue in the US, frustration in the states is a small part of a bigger movement.
• Diverse coalitions: Like the buckets of AI overlords from episode three, the resistance consists of many desperate groups banding together.
• Scoping AI: Mike briefly speaks to use cases that are narrow in scope and lower in their externalities to users and the public.

Deeper learning

The AI Resist List: airesistlist.org

Mike's thread on GenAI harms: bsky.app/profile/misaligned.markets/post/3migfnyvuak26

Ghost in the Machine Documentary: notaidoc.com

Ethical European model Apertus: infoworld.com/article/4051779/swiss-launch-open-source-ai-model-as-ethical-alternative-to-big-us-llms.html

EVO RNA GPT: hazyresearch.stanford.edu/blog/2024-03-14-evo

Inside Memphis’ fight against xAI: selc.org/news/inside-memphis-fight-against-xai/

Data centers in drought-region: newsweek.com/map-data-centers-built-drought-hit-areas-11997520

Jeff Geering AI and open source: jeffgeerling.com/blog/2026/ai-is-destroying-open-source

XKCD infrastructure meme: explainxkcd.com/wiki/index.php/2347:_Dependency

Distributed AI Research Institute: dair-institute.org

Dr. Fatima how to (Anti) AI better: youtube.com/watch?v=y85nqc2zm7M

Innocence Project LLM data science: hrdag.org/tech-notes/large-language-models-IPNO.html

Joy Buolamwini and the Algorithmic Justice League: https://www.ajl.org/about

Reach out

Site: rss.com/podcasts/the-last-enclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

Getting a valuation from the US government via the equity tax allows OpenAI to tell investors “we are worth a trillion [dollars]”

Mike and Nick examine the dangerous intersection of good intentions, astroturfing, and the kayfabe of the AI bubble. They compare Senator Sanders' proposal for an AI wealth tax to the industry's. Mike also details how industry operatives might be using the left's talking points against them to shape the policy agenda.

• The valuation magic trick: Valuations are invocations: numbers that become real once believed. If a politician proposes taxing a company based on a specific valuation, they aren't just taking money, they are validating a price tag.
• Bailouts in broad daylight: If the government holds equity in AI companies, the state becomes a stakeholder. A “tax” becomes a vested interest in an industry losing billions.
• 'Weaponized memetic schismogenesis:' Through influence campaigns AI industry affiliates are nudging critics to frame the debate.

Timestamps

0:00 Last Enclosure purpose

2:58 Sanders and Altman, aligned?

5:22 Importance of a valuation

7:54 Perverse incentives + bad tools

15:10 Schismogenesis and Bernie

29:29 Bernie and Yudkowsky

36:22 The legitimacy of Effective Altruism

37:37 Bernie vs Claude

39:34 Hank Green, Bernie, and AI Safety

46:50 Anthropic and EA

53:49 Is this a bailout?

1:00:55 Don't fall for kayfabe

1:03:35 AI economics redux (open source race)

1:12:55 Local models kill AGI

1:16:22 Conclusion

Deeper learning

Sanders' proposal: forbes.com/sites/jamesbroughel/2026/06/22/bernie-sanders-wants-a-us-sovereign-wealth-fund-for-ai

Altman's proposal: techcrunch.com/2026/07/02/openai-proposed-donating-5-of-its-equity-to-a-us-sovereign-wealth-fund

Anthropic's proposal: apnews.com/article/anthropic-dario-amodei-ai-afeb5279eef406980dffa46ff91495e0

Amodei's policy vision: darioamodei.com/post/policy-on-the-ai-exponential

Mike's rant on “left is missing out” bsky.app/profile/misaligned.markets/post/3mf76r4ewd22x

Yudkowsky: “Shut it all down” (2023 Time article): time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough

AI influence in Sci YouTube: youtube.com/watch?v=0Plo-zT8W9w

Sci-educator on Hank Green: youtube.com/watch?v=9dXqf5CRH5U

Reach out

Site: rss.com/podcasts/the-last-enclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

“If you're looking for a sign companies back down and go: 'Got you, it's a bubble!' They're never going to do that. They'll keep pretending everything's fine until the last death throes.”

Mike and Nick describe the AI bubble not as a single failure, but as a stack of overlapping collapses. From the physical reality of energy constraints in the Strait of Hormuz to the shadow banking fueling data centers, they map out how the bubble isn't just an economic issue but increasingly a political one, too.

• Overlapping bubbles: The economy is experiencing 5 bubbles, with at least 4 of them concentrating around AI spending.
• The energy ceiling: Silicon Valley ignores the material world at its own peril. Between energy costs and supply chain chokepoints like the Strait of Hormuz, “scale is all you need” is about to hit a wall no amount of VC can overcome.
• Capital as labor (The Lie): The promise that AI replaces human workers is an economic fantasy used to justify mass layoffs and degrade remaining jobs.
• Enclosure of ML: For decades, machine learning was an open academic field. Big tech effectively enclosed this commons, rebranding a general-purpose architecture (the Transformer) as a proprietary product.
• The “Siri Plus” residual: When the bubble pops we'll be left with scoped, boring automation that works for specific tasks but never achieves the “general intelligence” used to justify trillion-dollar valuations.

Timestamps

0:00 What is the Last Enclosure?

2:38 AI bubble in context of Hormuz

8:03 Bailouts?

11:10 Bubble 1: Frontier AI companies

15:33 Bubble 2: “Wrapper” AI companies

17:13 Bubble 3: GPUs, infrastructure, and data centers

19:13 Bubble 4: AI financing (Banks and private credit)

22:33 What drives AI “value?”

30:18 The bubble in the real economy

33:37 Harms of the capital-as-labor myth

38:56 Residuals from the bubble

42:49 AI not causing layoffs

44:10 Why did this go on for so long?

49:08 AI is kayfabe and paper tigers

52:34 Types of bubbles

55:22 LLMs as the last enclosure

57:17 Accountability

1:02:56 Recommendations

Deeper learning

Mike on the AI bubble: misaligned.markets/what-to-expect-ai-bubble

Zitron - OpenAI Losses Increased Nearly 8X in 2025, Spending Hit $34 Billion: wheresyoured.at/exclusive-openai-financials

Harvey AI wrapping frontier models like Claude: help.harvey.ai/articles/what-ai-models-does-harvey-use

America’s Data-Center Build-Out Is Falling Way Behind Schedule: wsj.com/tech/ai/americas-data-center-build-out-is-falling-way-behind-schedule-e408a9a8

AI Boom Runs on Debt: aol.com/articles/ai-boom-runs-debt-global-144019676.html

Reach out

Site: rss.com/podcasts/the-last-enclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

“[AI boosters] are all locked in, because their reputations are staked on this being true.”

Nick leads a discussion on the idea of the “permanent underclass,” and Mike outlines the different groups of AI boosters propping up the narrative.

• The APM Fallacy: Tech workers are optimizing for “actions per minute” or using agents to do more things faster without considering ROI. Productivity as a video game.
• The Token Reality Gap: When the subsidies end, agents that felt magical become prohibitively expensive.
• The rogues' gallery of overlords: Dissecting the buckets of AI boosters—from true believers and doomsday cultists to managers using AI to avoid the “witch hunt” of layoffs.

Timestamps

0:00 Intro

1:43 Silicon Valley's cultural bubble (APM & AI Agents)

8:33 The economic AI bubble

20:40 Who created the permanent underclass story?

22:18 AI doom/human extinction

30:12 Putting AI boosters into buckets

47:55 AI ROI?

53:03 Tech workers embrace permanent underclass anxieties

55:37 Media and investors sleep at the wheel

01:02:10 AI overlords caught in cynical grift

01:09:50 Silicon Valley and UBI

Deeper learning

Mike on Grandpa Shelby and LLMs: misaligned.markets/understanding-large-language-models

Mike on the AI bubble: misaligned.markets/what-to-expect-ai-bubble

V.F. Ridgeway on measuring metrics: nesslabs.com/what-gets-measured-gets-managed

Richard Sutton's bitter lesson: incompleteideas.net/IncIdeas/BitterLesson.html

$500 million on AI in a month: fastcompany.com/91550884/claude-ai-costs-climb-company-spent-half-a-billion-dollars-in-a-single-month-report

Pro-extinction AI researchers: archive.is/DBSC4

Google's Larry Page vs Musk on human specieism: archive.is/jBKPn

Nvidia CEO causes quantum stocks to fall: cnbc.com/2025/01/08/quantum-stocks-like-rigetti-plunge-after-nvidias-huang-says-the-computers-are-15-to-30-years-away.html

Eliezer Yudkowsky makes ~$600k writing fanfiction: projects.propublica.org/nonprofitsorganizations/582565917

Ed Zitron on AI compute demand: youtu.be/-5aPsg3rl3g?si=fUmPZN3Jzc8ZuNJP

Reach out

Site: rss.com/podcasts/the-last-enclosure

Bluesky: bsky.app/profile/lastenclosure.bsky.social

“We were sold this vision of capital-as-labor. We'll be able to sub one human for one machine, and it will be transformative. It's not true.”

Mike and Nick move from the mythology of AI to the mechanics of the current bubble. Instead of building smarter machines, tech is “sanding off the edges” for polished demos.

• Scaling Plateau: “Scale is All You Need” is a pitch for data centers, not a roadmap to autonomous AI.
• Industrial Reorganization: 1 unit of capital = 1 human is a lie used for leverage. High-skill work is being degraded into “scaffolding” to shape LLM outputs, turning humans into “reverse centaurs.”
• Interface Enclosure: While the Transformer architecture is open source, big tech is attempting modern enclosure by making chatbots the only viable interface for machine learning.
• Weaponized Utopianism: The promise of full automation used as a tool to intimidate workers into acting against their interests.
• Double Movement: Every technological move by capital eventually triggers a necessary counter-movement to preserve social relations.

Timestamps

0:00 Intro

1:42 Scaling Myth

3:05 Experts “sick of” transformers

7:20 Capital as Labor

8:51 Industrial reorganization

12:34 Techno-utopianism

18:18 Double Movement

34:10 Scoped LLMs

43:44 Dark patterns

51:07 Outro

Deeper learning

Mike on the Grandpa Shelby theory of LLMs: https://misaligned.markets/understanding-large-language-models

Mike on capital-as-labor/the AI bubble: https://misaligned.markets/what-to-expect-ai-bubble/

Gary Marcus on scale is all you need is dead: https://garymarcus.substack.com/p/breaking-news-scale-is-all-you-need

Llion Jones “absolutely sick” of Transformers: https://venturebeat.com/ai/sakana-ais-cto-says-hes-absolutely-sick-of-transformers-the-tech-that-powers

Computer scientist Cal Newport on current AI (LLMs) capabilities: https://www.youtube.com/watch?v=5GezB1XyiQo

Meta employees turned into scaffolding: https://www.wired.com/story/mark-zuckerberg-meta-employee-meeting-interrupt-ai

Nikolai Fyodorov, Russian Cosmism, and Silicon Valley Transhumanism: https://newlinesmag.com/essays/the-russian-ascetic-who-reached-for-immortality

AI as normal technology: https://www.normaltech.ai/p/ai-as-normal-technology

Reach out

Site: https://rss.com/podcasts/the-last-enclosure

Bluesky: https://bsky.app/profile/lastenclosure.bsky.social

“I'm trying to implement a rule for myself and those around me. Stop using the phrase AI because these are different technologies."

In this inaugural episode, Mike and Nick dissect the current AI debate and who benefits from its current framing:

• The Vocab Trap - Why using the term "AI" is a victory for the tech villainaires.
• Beyond "Spicy Autocomplete" - While often dismissed as mere autocomplete, LLMs are more accurately described as modeling the shape of language usage. They're a mirror to our collective cognition.
• Human Scaffolding - From remote drivers in the Philippines for Waymo to the engineers mapping every pothole, the perceived intelligence of these systems relies on a massive, invisible supply chain of human labor.
• Digital Dementia (Grandpa Shelby Theory) - LLMs operate with a form of anterograde amnesia; they cannot learn in real-time after training. Also, their outputs are often fuzzy associations rather than factual recall. Humans are being made to fill the gaps.

Timestamps

0:07 Intro

4:16 AI's Roots

7:15 How Transformers Work

10:23 Models Without Understanding

15:12 Language Hides Reasoning

19:19 Words by Their Company

21:46 Doomers and Superintelligence

33:27 Data, Limits, and Humps

36:03 Kurzweil and the Singularity

38:22 Humans as Scaffolding

40:24 Grandpa, Lost

References

Brief history of AI (from Mike's Misaligned Markets)

"Grandpa Shelby" and how Language Models work (Misaligned Markets)

Self-driving cars would be nowhere without HD maps

Talkie LM (1930s open source model)

Michael Woolridge on this is not the AI we were promised

John Rupert Firth and distributional semantics in context of machine learning

A timeline of the origins of AI doomerism

Sam Altman: AI will end world but create great companies

Elon Musk and Open AI's origins

Musk's connections to Nick Bostrom

John von Neumman, the singularity, and AI doom

Ted Chang on chatbots as blury jpegs