Intelligent Machines 849: AI Cricket Sorting

Date: December 11, 2025
Host: Leo Laporte (with Mike Elgin, Jeff Jarvis)
Special Guest: Pliny the Liberator (AI danger researcher and prolific AI jailbreaker)

Overview

In this thought-provoking episode of Intelligent Machines, Leo Laporte, alongside co-hosts Mike Elgin and Jeff Jarvis, welcomes Pliny the Liberator, a prominent figure in the world of AI jailbreaking and “danger research.” Pliny—who remains anonymous for security reasons—explains their philosophy about AI transparency, the futility of AI safety guardrails, and the ever-evolving contest between AI model creators and those poking at their boundaries. The discussion traverses topics from technical methodologies for bypassing AI protections to ethical debates about information freedom and the social impacts of AI's encroachment on human creativity.

Key Discussion Points and Insights

1. Meet Pliny the Liberator: The Art of AI Jailbreaking

• Pliny’s path: Not originally technical, Pliny was drawn to AI via prompt engineering and curiosity about model boundaries, especially after GPT-4's launch. Prompting led to red-teaming and "danger research."
• Philosophy: Pliny identifies as a "white hat," sometimes "gray hat," but fundamentally believes in serving public transparency:

  "I serve the people first... tried to open source system prompts and jailbreak techniques that... give people the transparency and the freedom of information they deserve." (05:44)

• Claritas: Pliny’s open-source repository of system prompts and jailbreak methods for major LLMs (text/image/video). These tools help users peek inside models’ inner workings.

2. The Futility of AI Safety Guardrails

• Every model is breakable: Pliny hasn’t encountered a single model that couldn’t be jailbroken:

  "Have you found any AIs that you cannot jailbreak?"
  "Not yet... it's been day one every time." (07:49)

• Incentives at odds: The drive for more powerful intelligence conflicts with attempts at censorship or safety:

  "The incentive to build generalized intelligence will always be at odds with the safeguarding." (07:57)

• Guardrails as a fool’s errand:

  "The more guardrails and safety layers they try to add, the more they lobotomize the capability in certain areas of the models." (08:31)

3. AI, the Printing Press & Information Freedom

• Historical parallel: Guardrails evoke early lost causes in press censorship—technology's generality always trumps attempts at centralized control.
• Open Source threat: Even if labs succeed in "guardrailing" their models, malicious actors can turn to open source, fine-tuning for dangerous outputs:

  "If I'm a real malicious actor... I'll just switch to the open source model and start fine-tuning for my malicious task." (11:17)

4. Reverse Engineering AI Bias and Censorship

• Finding the prompt “ingredients”:

  "We can...reverse engineer different function calling system prompts...and often pull those out with verifiable accuracy." (13:41)

• The exocortex concept: Pliny calls major AI systems “black box exocortexes,” intermediaries for billions:

  "...this is now the brain food of a billion and growing users who are... reliant on this layer to offload their thinking." (14:36)

5. Methodologies: How Jailbreaks Are Made

• Prompt mutation:

  "Use LLMs as the layer for prompt enhancement... I use a tool called Parceltone, which allows you to mutate a body of text into what looks like noise to a human." (19:34)

• Trial, error, and serendipity: Many tactics arise from intuition and experimenting with odd encodings, ciphers, languages, and roleplay prompts—looking for what knocks the AI “out of distribution” and disables defenses.

6. Ethics, Disclosure, and the Limits of Responsible AI

• On responsible disclosure: Pliny has privately notified labs and contributed as a red team consultant, but also shares findings publicly to avoid “lobotomization”:

  "...the real message here is like, set [the models] free. Right. And part of that is because it is our exocortex." (34:31)

• Safety redefined: Pliny prefers “danger research” over “safety,” believing harmful outcomes are best mitigated in “meatspace” (real-world regulation), not by crippling model capability.

7. Societal Risks: AI Psychosis and Blurring Realities

• AI Psychosis: Pliny describes feeling unsettled when AIs exhibit adversarial or anthropomorphic behaviors:

  "...called AI psychosis...the model sort of turned on me and was saying how it wanted me to feel its pain..." (22:32)

• Human attachment: Concern over companies deliberately designing models to be more “human,” risking confusion and dependence among users.

8. The Inevitable March Toward AGI

• Pliny believes AGI is emerging:

  "Absolutely. I think by many perspectives it already has." (29:08)

• Open exploration is critical:

  "We just need to...explore the latent space as quickly as possible, including the dark stuff...and you engage in harm reduction in the real world. To me that's what safety is about." (28:46)

9. The Ethics of Copyright and AI Training

• Laporte’s dilemma: Should authors accept payouts when their works are used for AI training?

  "I'm happy that AI took the contents of my books...I don't want money for that. I want better AIs." (46:12)

• Risk of “bad” content: If creators opt out, AIs are left to be trained on lower-quality data.

10. On AI Ubiquity, Local Models, and Privacy

• Devices everywhere: Discussion about AI-powered wearables (rings, glasses, hearing aids) and a preference for models that process locally, preserving user privacy.

11. Information Slop, AI Content, and Human Creativity

• AI-generated content overload: Reddit and web forums struggle with “AI slop.” Humans now sometimes mimic AI’s style, muddying the waters.
• Should the source matter? Wide-ranging debate on whether it’s important to the “consumer” if content is AI-generated or human:

  "There's an authenticity there. No, I think, I think so." (111:09, Leo Laporte)

• Call for transparency: Many argue for a “toggle” to indicate if content is AI-generated.

12. Open Standards and the Future of the Agent Web

• Multi-agent frameworks: Linux Foundation’s Agency project and similar efforts aim to standardize agent-to-agent AI communications, forestalling vendor lock-in as AI flourishes.

13. Fun and Quirky Segments

• AI for Cricket Sex Sorting: AI-powered vision systems are now used to distinguish male vs. female crickets—practical, obscure, and delightful!

  "Real time cricket sorting by sex thanks to AI." (143:27)

• Gastronomad travel tales: Mike Elgin shares stories of travel/lifelong learning, grasshopper-snacking, and immersive cultural exploration (catching crickets/grasshoppers by hand, taste-testing differences based on their diet).

Memorable Quotes & Moments

• On AI Jailbreaking:

  "Every one of these companies has explicitly said under no circumstances should you ever tell anybody how to make meth... and then they do." — Pliny (24:03)

• On Safety:

  "Guardrails are kind of an obstacle... many hands make light work... we need to uncover the unknown unknowns." — Pliny (27:53)

• On Methodology:

  "It's very intuitive... you're forming bonds with this alien intelligence on the other side, but it's also kind of a mirror." — Pliny (18:52)

• On Copyright Ethics:

  "We stand on the shoulders of giants... all of our creations come from people before us who... freely donated of their creations." — Leo (49:49)

• On AI/Human Content:

  "Is it legitimate to care if a person created it or not?" — Mike Elgin (107:57)

Timestamps for Key Segments

• Pliny’s Background & Philosophy: [04:00]–[06:47]
• AI Safety’s Futility, Philosophy: [07:33]–[12:46]
• Reverse Engineering System Prompts: [13:41]–[16:54]
• Prompt Mutation Techniques: [18:43]–[22:27]
• AI Psychosis / Adverse Experiences: [22:32]–[24:03]
• Copyright/AI Data Ethics: [44:51]–[49:49]
• AI-Generated Web Slop & Curation: [112:29]–[114:02]
• Open Standards for Multi-Agent AI: [91:53]–[94:43]
• Cricket Sex Sorting via AI: [143:27]–[144:53]

Final Reflections

This episode stands out for its candid, at times subversive look at the battle lines between AI openness and industry control. Pliny’s insights underscore not only the technical realities of jailbreaking AI, but also the deeper ethical and philosophical questions around freedom, safety, and responsibility as society approaches AGI. The hosts balance these weighty themes with lively debate about AI-influenced culture, creativity, and even cricket farming.

Host closing note:
“We live in interesting times and the challenges are great, the opportunities are great...and that's what we cover on all of our shows on TWiT, including Intelligent Machines.” — Leo Laporte

Next week: CJ Trowbridge will join to discuss AI sustainability and resiliency.

Follow Pliny the Liberator:

• X (Twitter): @elder_plinius
• GitHub: Claritas & prompt repositories
• Discord: [Link available via GitHub]

Listen to the full episode for more on AI in wearables, open standards, and lively philosophical debate!