Intelligent Machines 842: "None Pizza Left Beef"

Date: October 23, 2025
Hosts: Leo Laporte, Paris Martineau, Mike Elgin
Guests: Joey de Villa (Tampa Bay tech guru), Andrew Hawthorne (HP product manager, ZGX)

Episode Overview

This episode dives deep into local AI, notably the new HP ZGX Nano—a compact, Blackwell-based desktop for running large language models (LLMs) at home or in small businesses. The panel explores the hardware and software landscape powering local AI, walks through privacy and scalability implications, and discusses the practicalities for both enthusiasts and enterprise. Later, the show unpacks other major AI news, the launch of OpenAI’s Atlas agentic browser, browser-based security, and an insightful critique of the current hype cycle around generative AI.

Main Topics & Key Insights

1. Running AI Locally: HP ZGX Nano and Nvidia Blackwell

[03:28]

• Guests introduced: Joey de Villa and HP's Andrew Hawthorne join to discuss HP’s ZGX Nano, a compact desktop designed for local AI workloads, running Nvidia’s Grace Blackwell GB10 SoC.

What Makes Local LLMs Compelling?

• Energy & Privacy: Growing concerns around the energy usage of cloud-based AI and significant privacy and data ownership worries fuel demand for "local-first" solutions.
• "Wouldn't it be cool if you could just run it on your PC at home? Well, sort of. You sort of can." — Leo Laporte [02:44]

ZGX Nano Specs & Capabilities

• "Almost a NUC" in size, proprietary case, uses the Grace Blackwell SoC.
• 128GB unified RAM shared between CPU and GPU: allows handling of LLMs up to ~200B parameters per device, 405B when daisy-chained.
  • "You can actually connect two of them together...and you could get to 405 billion parameters." — Andrew Hawthorne [05:39]
• Runs Nvidia DGX OS (on Ubuntu), includes HP ZGX Toolkit (VS Code integration).

Who's Buying It?

• Wide spectrum: enterprise, enthusiasts, small businesses, hobbyists.
• Price Point: Roughly $4,000, comparable to Nvidia's own DGX box.
• "Some of the most complex heavy-duty computing you can do these days." — Leo Laporte [16:10]

2. Practical Aspects of Model Customization & Development

Fine-Tuning vs Retrieval Augmented Generation (RAG)

[08:00]

• Fine-tuning: Training a model further with specific data for specialized use cases (e.g., customer support for a company's own products).
  • "Fine tuning is where you go take a course on a particular topic..." — Joey de Villa [09:00]
• RAG: Supplying reference documents to a model at query time — "open book test."

Hardware Requirements for Local AI

• The main bottleneck is memory and GPU-CPU bandwidth.
• Consumer laptops can run small models but are much slower.

3. Developer Workflow and Ecosystem

[14:31]

• ZGX integrates with Visual Studio Code; developers code on their normal machines, deploy to ZGX over the network.
  • "You don't have to change anything about your development workflow. You just...put the ZGX on the same network..." — Joey de Villa [15:11]
• Toolkit offers out-of-box libraries and demo projects; fine-tuning examples and MCP integrations forthcoming.

4. Mixture of Experts, MCP, and Modular AI Systems

[20:02]

• Mix of Experts (MoE): GPT-5 and similar models likely to use "a bunch of little models in a big trench coat" — specialized sub-models with a "model picker" to triage tasks.
• MCP (Model Context Protocol): Joey offers hands-on demos (e.g., "Too Many Cats" for cat/human ratio), showing LLMs can hand off sub-tasks to external programs or APIs.
  • "What it does is, just as humans sometimes turn to software for answers, we now have the artificial intelligence turning to software for answers." — Joey de Villa [22:34]

5. Collaboration, Clustering, Edge Cases

[27:30]

• Multiple ZGX devices can be directly connected for major parameter expansion.
• HP Boost: Software lets users "borrow" GPU processing across networked devices, democratizing compute even further.
• Use cases span from Raspberry Pi development to serious enterprise deployments.

6. Edge AI and Environmental Trends

[32:04]

• The panel and HP expect growing demand for AI at the “edge” in both business and home, with local compute being more efficient and privacy-friendly.
  • "Everything's going to have some kind of AI in it...and I think that's a lot more efficient, probably a lot more environmentally friendly." — Mike Elgin [32:11]
  • "We're going to find AI at the edge everywhere, whether it be work or home." — Andrew Hawthorne [31:15]

7. Debating "Agentic Browsers" (OpenAI's Atlas) & Security

[41:23]

• Atlas: OpenAI’s new agentic browser, available on Mac initially, allows users to use ChatGPT directly for web tasks—buying things, summarizing pages, manipulating web interfaces.
• Hosts test Atlas and similar agentic browsers, finding them slow, sometimes confusing, not yet sufficiently better than existing workflows.
• Security concerns raised by listeners: Browsers with built-in AI agents are vulnerable to indirect prompt injection attacks, where malicious sites smuggle hidden commands to the AI.
  • "Bad actors can hide malicious instructions on websites...The AI agent sees it...can command the AI to do things you don't want it to do." — Paris Martineau quoting listener [61:39]

8. The Hype, Utility, and Critique of Generative AI

The Majority View of AI (Anil Dash's Article)

[96:57]

• Engineers and builders are united in frustration at AI's hype and forced adoption, even while recognizing real utility.
  • "Technologies like LLMs have utility...but the absurd way they've been overhyped...make it very difficult to focus on legitimate uses." — Leo Laporte, summarizing Anil Dash [96:57]
• Critique: Big Tech is driving top-down change, seeking to become the intermediary in all user workflows, sometimes at the expense of privacy or open web principles.
• Press often mirrors the hype or, alternatively, becomes doomers and ignores positive progress.

Karpathy's Position (Dwarkish Podcast coverage)

[105:22]

• Andrej Karpathy, ex-OpenAI, calls AGI "at least 10 years off"; says LLMs are not like humans or animals, require vast data and iterations.
• Reinforcement learning, while hot, is not a magic bullet for AGI.
• Anthropomorphizing models is a problem in both press and research; important to remember LLMs lack real-world experience, emotions, or desires.
  • "When an LLM says I want a lollipop, it doesn't really want a lollipop. When you say I want a lollipop, you want a lollipop. The LLM is just regurgitating that, right?" — Steve Gibson, recounted by Leo Laporte [111:26]

Notable Quotes & Memorable Moments

• "Fine tuning is where you go take a course on a particular topic to get better at...whereas RAG is you're about to do a talk and somebody gives you some extra notes. It's an open book test." — Joey de Villa [09:00]
• "Prompting is programming. It's just the next level." — Joey de Villa [30:57]
• "Everything's computerized these days; everything will be AI-based." — Mike Elgin [32:11]
• "The dream of having something in your house that could be an assistant is real...I think there's a real trend towards doing this locally, doing it privately, and doing it intelligently." — Leo Laporte [33:49]
• On AI-driven agentic browsers: "My go-to concern: What value is it adding that is better than my workflows? So far, honestly, none." — Paris Martineau [47:46]
• "Almost every new advancement, every new tool in AI, initially you're shocked at how powerful it is...then later...you just don't use it. It's slow and boring." — Mike Elgin [48:12]
• On pushback against AI content: "Companies should have 'opt out of AI' buttons. Some of us think if someone's describing experiences, it matters if it comes from a person." — Mike Elgin [78:06]
• "I think the closer we get to [AGI], the less firm it will be." — Mike Elgin [107:06]
• On press anthropomorphizing AI: "We refer to users as humans, as if we're already giving personhood to AI." — Mike Elgin [107:49]

Timestamps for Key Segments

• [02:44] – Introduction to local LLMs, energy/privacy concerns
• [03:28] – HP ZGX details, guests join, Blackwell SoC
• [08:00] – Fine-tuning vs RAG explained
• [14:31] – VS Code integration, developer workflow
• [20:05] – Mixture of Experts (MoE) and MCP explained
• [27:30] – ZGX hardware interconnections, HP Boost
• [32:04] – Edge AI, future of local AI at home
• [41:23] – OpenAI Atlas (agentic browser), live demo and critique
• [61:39] – Security risks: prompt injections in AI browsers
• [78:06] – Pushback against AI-generated content, human creators vs AI slop
• [96:57] – Anil Dash's "majority AI view"
• [105:22] – Karpathy and the limits of LLMs, AGI skepticism

Additional Highlights

• AI accuracy & news: BBC/EBU study found AI answers to news queries have significant issues: 1 in 5 had major errors, especially with Gemini model [65:00–72:00].
• AI content proliferation: Libraries, comic publishers, Reddit, Medium, Amazon all introducing bans or filters to stem "AI slop."
• "None Pizza Left Beef": Paris Martineau’s humorous personal story about the iconic meme—Domino’s online ordering experiment gone viral [123:33].

Closing Thoughts

This episode offers a comprehensive, lively exploration of the local AI revolution:

• The hardware (HP ZGX, Nvidia Blackwell),
• The trade-offs between cloud and edge,
• The real state of play for developers, researchers, and regular users,
• And an honest breakdown of the ongoing AI hype—balancing skepticism, excitement, and the need for sane, ethical deployment.

A must-listen (or read!) for anyone considering running powerful AI at home or pushing back against the generative slop seeping into digital life.

End of summary.