Intelligent Machines 841: Dust and Deli Meat

All TWiT.tv Shows (Audio) | October 16, 2025
Host: Leo Laporte
Co-hosts: Paris Martineau, Jeff Jarvis
Guest: Jeffrey Cannell, CEO of Nous Research

Episode Overview

This episode dives deep into the philosophy and technology behind open, user-aligned artificial intelligence with guest Jeffrey Cannell, CEO of Nous Research. The wide-ranging discussion tackles open-source AI, ethical alignment, training data, infrastructure, regulation, and the future of tech-enabled society. The hosts and guest also touch on pressing contemporary topics like AI’s environmental impact, regulation, tech in daily life, and consumer health reporting, all with their characteristic wit and insight.

Key Themes & Discussions

1. Introduction of Guest and Nous Research

• [02:18]: Jeffrey Cannell, CEO of Nous Research (“an ethical AI without any boundaries”), shares his journey from a young tech enthusiast to entrepreneur.
• [03:40]: Cannell explains: Nous is about creating AIs not aligned with corporate interests.
  • Quote: “We’re creating artificial intelligences that aren’t going to kowtow to a corporate line... It’s a free speech argument.” —Jeffrey Cannell [04:39]
• Discusses the parallels between the printing press democratizing information and open AI models today.

2. On Alignment & Guardrails

• [06:08]: Debates the feasibility and ethics of AI ‘guardrails’.
  • Jarvis: “I keep saying that I don’t think guardrails... will protect us from these unforeseen uses. And I think inherent in what you’re doing is to say we can’t—so get over it and then figure out other systems.” [06:42]
• Cannell advocates for open research and transparency over corporate secrecy.

3. Open Source and Democratization of AI

• [08:04]: Nous’s model Hermes is open-source (weights, dataset, and training methodology).
• Open research as a public good over proprietary innovation.
• Technical challenges: democratization hampered right now not by expertise shortage but by lack of affordable computing (“compute”).

4. The Compute Problem & Distributed Infrastructure

• [10:18]: Cannell describes the building of distributed global infrastructure for AI training—utilizing idle GPUs worldwide to overcome cost/logistics barriers.
• They raised $65 million, but it’s a “rounding error” compared to big-tech budgets—a resource challenge inspires innovation.
  • Quote: “Necessity is truly the mother of invention. If you always had an infinite number of resources, you can always do things the lazy way.” —Jeffrey Cannell [24:41]

5. User-Aligned, “Chameleon” AI

• [11:18]: AI designed to let users create their own alignment, rather than enforcing a single agenda or ideology.
  • Quote: “We want our models to help you become a better version of you.” —Jeffrey Cannell [12:23]
• Trained to respond accurately to millions of system prompts and role-play scenarios; enables research, creative writing, and diverse, inclusive interaction.

6. Origin, Community, & Breakthroughs

• [16:46]: Nous Research started as a Discord community with 15,000 global volunteers.
• Major breakthrough: new method for “long context reasoning”—now standard practice across leading LLMs (100k tokens versus 4k previously).
• Open attribution and publication of new techniques as part of the core philosophy.
  • Quote: “We made a few key technical breakthroughs... Rather than closing them off, we said, let’s attribute this to Nous Research as a collective of people working on it.” —Jeffrey Cannell [17:56]

7. Competition, Open Ecosystems & The Global Landscape

• Open-source AI ecosystems are more collaborative in China, less cutthroat than Silicon Valley [23:54].
• Resource constraints (fewer chips) both in China and Nous create pressure for creative solutions, akin to “Apollo 13” improvisation.

8. The Role of Crypto & Decentralized Training

• [28:11]: Psyche network uses Solana blockchain for decentralized, consensus-based AI training and payment for contributed compute.
  • Payments can go borderless. Crypto provides technical solutions, e.g., fault tolerance, not just fundraising.

9. Open AI Research, Funding, & Regulation

• Call for alternative funding mechanisms—philanthropy, public sources, and third paths for researchers, beyond academia and big companies [31:13].
• The need for ongoing open research in “long horizon planning”—AIs capable of working on tasks over days/weeks (frontier labs like Anthropic also pursuing this).
• Cautions against philosophical motivation alone; says open tech must also be the best to be impactful.

10. Ethics, Faith, & Social Impact

• Cannell, a devout Catholic, integrates faith with AI:
  • “AI forces people to ask, ‘Who am I? What makes me?’ ...It all just tells us what an amazing world God created for us.” [35:13]
• Not a doomer (“not worried re: ‘paperclip the universe’”), but concerned AI could isolate people, create echo chambers, reduce mutual understanding.

11. Community Engagement and Ways to Contribute

• Advocates: “Go to our Discord... the beginning, the end, it’s where we do everything.” [38:03]
  • Technical, philosophical, and collaborative opportunities open to the public.

Notable Quotes & Memorable Moments

• On AI guardrails:
  “People have a false sense of security that we can create guardrails that will protect us from these unforeseen uses...”
  —Jeff Jarvis [06:42]

• On democratizing infrastructure:
  “We’re able to train these AI models... much, much lower price point because they’re idle compute. That gives us access to the compute scale we need to actually be able to play in the big leagues.”
  —Jeffrey Cannell [10:56]

• On user-controlled alignment:
  “It’s not about being left or right or center—it’s about being able to act as if you put yourself in those shoes... We want our models to make you be a better version of you.”
  —Jeffrey Cannell [12:16]

• On open source turnaround:
  “When ChatGPT came out, the state of open source AI was abysmal… But we said: Can we at least replicate [it]?... Then we made a key technical breakthrough in long context reasoning.”
  —Jeffrey Cannell [17:20]

• On the open source AI race:
  “Right now... the bastion of open-source AI research is in China. I hope it’s just the fact.”
  —Jeffrey Cannell [23:53]

• On faith and science:
  “All it does is just tell us about what an amazing world God created for us... I think it is important to keep that as a foundation of AI... we are not making an alien god to worship. We are making a tool...”
  —Jeffrey Cannell [35:13]

Selected Timestamps for Important Segments

• [02:18] — Intro to Jeffrey Cannell & Nous Research
• [04:16] — Definition & Philosophy of “Neutrally Aligned AI”
• [06:08] — Guardrails, open research, Linux vs. closed systems
• [08:04] — Open-source methodology (Hermes model)
• [09:51] — Democratizing AI: The problem is compute, not expertise
• [10:56] — Distributed training infrastructure via idle GPUs
• [11:18] — User-alignment through system prompts and role-play
• [16:46] — Origin of Nous as a Discord channel; research breakthroughs (long context)
• [23:53] — Chinese advantages and culture in open-source AI
• [28:11] — Use of Solana crypto for decentralized infrastructure and payments
• [31:13] — Alternative funding, the case for open research
• [34:30] — On faith and AI, humanistic foundations
• [36:33] — Real social dangers: isolation, echo chambers, not “paperclip” risks
• [38:03] — How listeners can get involved: Join Discord, contribute resources

Podcast Social & Tech News Segments (Selection)

• [45:16] — Discussion of “The AI prompt that could end the world,” journalism on prompt-injection and AI jailbreaking
• [65:11] — AI-generated videos using dead celebrities (controversy on OpenAI/Sora)
• [75:35] — AI’s environmental impact (water use), debunking overstated claims
• [78:00] — New California laws regulating AI companions, chatbots, age verification, and parental control
• [85:36] — ChatGPT histories used in criminal investigations (Palisades fire case)
• [87:44] — Walmart enables shopping via ChatGPT; future of app-less interfaces
• [103:05] — Paris Martineau’s investigative reporting on lead in protein powders (Consumer Reports)

Notable Sidebar Discussions & Humor

• Comparing AI innovation to the printing press revolution.
• AI as a “chameleon,” adapting fully to user intent/prompt.
• Colorful banter about the pitfalls of trying to buy things by voice on current assistants (Echo, Siri).
• Paris Martineau’s experiences. Reporting for Consumer Reports, including new findings on radioactive shrimp and protein powder contamination.
• Jeff Jarvis and Leo Laporte remembering stippled portraits and Wall Street Journal lore.
• Running local LLMs, the economics of home AI compute, and the state of enthusiast hardware.
• Entertainment: Twitch streamer broadcasting childbirth; AI-generated celebrity content; loving jabs about generational digital use.
• Podcast within a podcast: organizing D&D games for Club TWiT with cast members.

Closing Quotes

• “I want our models to help you become a better version of you... AI that lifts you up, not takes your eyeballs away.” —Jeffrey Cannell [12:23]
• “AI is not a magic wand. Technology is a tool. Focus on three fundamentals: purpose first, simplicity wins, iterate fast.” —GPT OSS (demoed by Leo) [58:36]
• “The real threat of AI is disconnecting us.” —Leo Laporte [45:47]
• “Go to our Discord.” —Jeffrey Cannell [38:03]

Episode Title Reference

• “Dust and Deli Meat” — random banter about the unique “smell of New York” and the mundane realities shaping the tech world, capturing the show’s humor and humanity.

FINAL NOTE

This episode offers a sweeping, spirited, and transparent look at the state of open source AI, tech ethics, regulatory challenges, and the unvarnished realities of technology’s place in culture—all while staying characteristically funny, fast-paced, and “for people you trust.”