AI Agents, Local Models and Cybersecurity with Christian Johnson – HGG693

What can a desktop from 2012 teach us about the future of AI? Christian Johnson joins me for HGG693 to talk about keeping useful hardware alive, relaunching Maple Grove Partners’ hosting platform with AI-assisted development, running models locally and protecting our computers as security updates accelerate.

Chapters / Timestamps

[0:03] A Backup Machine Built to Last

[1:49] Brazil, Travel, and the Show

[3:26] Maple Grove’s New Hosting Tools

[9:50] Building with AI

[14:33] The Hardware Behind Local AI

[26:54] GPU Prices and Future-Proofing

[30:49] Making AI Work Portable

[39:12] Better Habits for AI Development

[45:17] The Surge in Security Updates

[49:32] Can AI Write Safer Code?

[53:26] Poisoning AI’s Sense of Truth

[1:02:51] Can AI Do the Math?

[1:09:31] Staying Safe with Rapid Patching

[1:12:52] What’s Next for Maple Grove?

Guest Context

Christian Johnson is a longtime Home Gadget Geeks contributor and the person behind Maple Grove Partners, which hosts The Average Guy Network’s website and podcast audio. In this episode Christian describes the latest Maple Grove customer portal, hosting tools and his own local AI experiments.

Maple Grove Partners: https://maplegrovepartners.com

Maple Grove Hosting: https://maplegrovehosting.com

This Week on Home Gadget Geeks

• Why Christian’s long-running backup desktop is still useful.

• Maple Grove’s rebuilt web/email control panel and managed application containers.

• Agentic coding, better specifications and ambitious project goals.

• Local AI on a Ryzen desktop with 128GB RAM and a 24GB RTX 3090.

• GPU costs, reuse versus upgrading, and portable context across AI systems.

• Security updates, AI-discovered vulnerabilities, model poisoning and tool-assisted math.

Full Show Notes

A Backup Machine Built to Last

We opened with Christian describing an old desktop he originally had in college. That machine has kept working as a secondary backup computer since 2012. It’s a useful reminder that not every task needs the newest CPU or GPU. Christian says that old computer remains his always-on backup system. His goal is to build hardware that lasts about a decade, including his eventual AI machine. The old backup box is a different computer from his Ryzen desktop for local AI.

Brazil, the Show and Fifteen Years of Technology

I had just returned from a work trip to Brazil, and Christian and I caught up before turning to the technology. We’ve been doing versions of this show together for fifteen years. That’s long enough to see server hardware, the way we communicate online and now AI-assisted software development change dramatically.

Maple Grove Partners Rebuilds Its Hosting Experience

Christian has relaunched Maple Grove Partners’ website and customer portal. Instead of relying on licensed hosting-control systems such as cPanel and WHM, he used AI-assisted development to build a custom control panel covering web and email hosting. He credited listener Ken Bracken for beta testing and helping shape requested features. This is a particularly interesting example of agentic coding delivering a practical customer-facing product, rather than simply a demo.

Managed Application Containers: Vaultwarden and NetBox

Maple Grove now offers managed application containers. Christian described how a listener’s NetBox installation ran on its own dedicated VM for about a year. In that example he estimated about $80 per year for a managed container, versus more than $500 a year for the dedicated VM; those figures are from that one case, not a general price list. He also noted that applications packaged into a Docker container can be candidates for the managed platform. Vaultwarden, a separately developed Bitwarden-compatible password manager server, is another early offering. Customers can manage backups and download snapshots. Christian said Nextcloud and Minecraft hosting are possibilities, as is hosting inference capacity for open-weight models, but those ideas should not be mistaken for already-available services.

Building with AI Starts with Better Goals

Christian’s biggest advice for AI software development was not a complicated agent diagram. He starts with a clear goal, requirements and context, often in a Markdown file. He uses a strong frontier model for architecture and early reasoning before considering cheaper or local models for narrower execution tasks. His argument is that the models have improved enough that people can aim for larger deliverables instead of micromanaging every small step. I described how I use ChatGPT to work through a project in conversation before giving instructions to OpenClaw.

The Hardware Behind Local AI

Christian is still using a Ryzen 9 5950X desktop on an X570 platform. Rather than rebuilding everything, he upgraded from 64GB to 128GB of DDR4 RAM and swapped an RTX 3080 with 10GB VRAM for an RTX 3090 with 24GB VRAM. For his experiments, that extra GPU memory matters more than simply owning the newest generation of parts. He also discussed using Unsloth and experimenting with CPU and GPU roles in local inference. He reported running a dense Qwen 27B model at roughly 50 tokens per second on this system; that was his observation, not an independently reproduced benchmark. We talked about why the amount of VRAM can limit which models run practically at home; Christian said he filled up both his old 10GB GPU and now the newer 24GB card.

GPU Prices, Old Equipment and Future-Proofing

I shared my frustration about buying an RTX 3060 when I didn’t expect today’s local AI requirements. Christian talked through the trade-offs of used GPUs, the high cost of newer systems and the danger of buying a huge AI machine while software and hardware are changing quickly. Christian’s guess is that enterprise demand may keep GPU prices high through roughly the end of 2028; he’s thinking about a long-lived build around 2029. The better lesson is to experiment with what you already have, learn what bottlenecks matter, and avoid spending money solely on the promise of being future-proof.

Make Your AI Work Portable

One of the most useful parts of this conversation was Christian’s insistence on keeping context portable. Instead of letting everything live inside a single AI product’s proprietary memory, he records project goals, discoveries, code and progress in files that another model or computer can read. He calls this checkpointing the work. Christian also shares a common Google Drive workspace between Claude and OpenAI, so either system can work with shared files. I described my ChatGPT-to-Claude Drive inboxes, and Christian raised the question of whether those handoffs preserve the tools and outputs, not just the instructions. His point was to keep the work portable even if model providers change their prices. It’s a good distinction for anyone managing longer AI projects.

Better Habits for Agentic Development

Christian described giving agents a list of goals at night, letting them work overnight on maximum reasoning, and checking their results in the morning. He contrasted this with ‘loop engineering’: constantly nudging an agent through little tasks. In his view, more capable models can organize their own researcher, reviewer, developer and skeptic sub-agents. I like the idea of starting with the destination rather than writing an elaborate steering document for every step.

The Surge in Security Updates

We shifted to what I called ‘Updatepalooza’: frequent security updates across browsers, operating systems and applications. Christian discussed how AI tools can help researchers and attackers find weaknesses more quickly. Christian pointed to recently patched PHP remote-code-execution vulnerabilities as an example of why keeping up with security updates matters. He expects stronger AI models to find more long-hidden defects, although his prediction about what happens to vulnerability counts in the future is still an open question. The practical issue for listeners is straightforward: evaluate security updates promptly and keep track of vulnerabilities in the software you actually run.

Can AI Write Safer Code?

A question came up about whether code written by agents is necessarily correct. Christian said no; his expectation is that models may improve at recalling secure patterns as their capabilities increase. He also acknowledged that poor requirements can still produce insecure implementations. We treated the future security effects of AI-written code as a prediction and open question, not a proven conclusion. Christian’s view was that AI-written code has about the same defect rate as human-written code, but agents generate more code and therefore potentially more total vulnerabilities. Poor requirements still lead to problems.

Poisoning an AI Model’s Sense of Truth

Could enough wrong information eventually teach models the wrong answer? We discussed data poisoning, prompt injection and model collapse, including the concern that future models could be trained on an increasing amount of AI-generated material. I used the analogy of making a copy of a copy: each generation can lose something if errors aren’t checked. The discussion highlights why source quality and independent verification matter even when a model sounds confident.

Can AI Just Do the Math?

I described asking ChatGPT voice mode on an older model a few months earlier to calculate pi to fifty decimal places and getting an unreliable answer. It might perform differently today. Christian explained that models can call tools, including calculators or code, rather than relying only on generated text. That distinction matters: accurate numeric output may come from external computation, and Christian interpreted a current model’s instant answer as a calculator call, but I wasn’t convinced that speed by itself proved it. My takeaway: for important calculations, request a verifiable calculation and check the result.

Staying Safe with Rapid Patching

Christian’s advice was to prioritize timely updates. For a computer that has been powered down for a week or two, his recommendation was to bring the system and browsers up to date before doing routine browsing. He also suggested using an AI agent to alert you about relevant CVEs, with any automated patching kept within carefully chosen permissions. Christian specifically cautioned against routine browsing before browser updates finish because browser-level weaknesses can be remotely exploited.

What’s Next for Maple Grove?

Christian is excited about using agentic development to create and host useful applications, including productivity-oriented containers and possible AI workspaces. What I appreciate is that the work is grounded in a community of listeners and real needs. That’s a fitting end to an episode about keeping old hardware useful while deciding where the next wave of software belongs.

Key Takeaways

• Start complex AI projects with explicit goals and portable written context.

• Spend on local AI hardware only after learning your actual RAM, VRAM and workload needs.

• Container hosting can remove the overhead of maintaining a separate VM for every application.

• Keep project artifacts and tools portable, not only AI conversation summaries.

• Don’t assume AI-authored code is secure by default; make your requirements clear.

• Patch systems and browsers promptly, particularly before using a machine that has been offline.

• Treat predictions about the AI market and security future as predictions, not confirmed outcomes.

FAQ

Who is the guest on HGG693?

Christian Johnson of Maple Grove Partners, a longtime contributor to Home Gadget Geeks.

Why does the episode open with a computer from 2012?

Christian still uses an older desktop as a secondary backup machine, illustrating the value of keeping suitable hardware in service.

What is new at Maple Grove Partners?

A redesigned customer portal and custom web/email hosting control panel, plus managed application containers.

Which container applications did Christian mention as available?

NetBox and Vaultwarden were the first examples he described.

Are Nextcloud and Minecraft hosting already available?

Christian discussed them as possible future additions; the episode does not establish their release.

How does Christian start a serious AI coding project?

He outlines goals, requirements and context before asking a capable frontier model to plan the work.

What local AI hardware is Christian running?

He described a Ryzen 9 5950X system upgraded to 128GB RAM and an RTX 3090 with 24GB VRAM.

Why is VRAM important for local AI?

GPU memory limits which local models fit. Christian upgraded from 10GB to 24GB of VRAM to run larger models.

What does portable AI context mean?

Saving useful instructions, progress, files and tools in readable formats so another model or environment can resume work.

Is AI-generated code automatically secure?

No. The episode discusses possible improvements, while acknowledging that clear requirements and human oversight remain important.

What is model collapse?

The concern that training future models on AI-generated material can feed mistakes and polluted information back into later models, like a copy of a copy.

What should I do when I turn on a computer that has been offline?

Install relevant operating-system and browser security updates before routine browsing, as Christian recommended.

Products Discussed in This Episode

Some product links may be affiliate links. If you use those links to make a purchase, The Average Guy Network may earn a small commission at no additional cost to you. Thanks for supporting Home Gadget Geeks.

AMD Ryzen 9 5950X; NVIDIA GeForce RTX 3080; NVIDIA GeForce RTX 3090; NVIDIA GeForce RTX 3060; DDR4 RAM; X570 motherboard platform; Spark systems (mentioned on air as “Spark”); NetBox; Vaultwarden; Bitwarden; cPanel; WHM; Nextcloud; Minecraft; Unsloth; OpenClaw; Claude; ChatGPT; Gemini.

Links and Resources Mentioned

Maple Grove Partners — https://maplegrovepartners.com

Maple Grove Hosting — https://maplegrovehosting.com

NetBox — https://netboxlabs.com

Vaultwarden project — https://github.com/dani-garcia/vaultwarden

Bitwarden — https://bitwarden.com

Nextcloud — https://nextcloud.com

Unsloth — https://unsloth.ai

Artificial Analysis (model comparisons mentioned in conversation) — https://artificialanalysis.ai

The Average Guy Network — https://theaverageguy.tv

Patreon — https://theaverageguy.tv/patreon

Discord — https://theaverageguy.tv/discord

Listen and Subscribe

Listen to Home Gadget Geeks and find the show archive at theaverageguy.tv. Join our community at Discord or support the show at Patreon.