A Greenhouse, A Server, and the Neighborhood

There’s a 26-foot geodesic dome in my backyard. You can see it when you fly into the airport. From the street, it looks like a lot — but wait until you see what’s in the basement.

I’ve spent the last month building the infrastructure to support it — a server, VMs configured for AI workloads, a mesh network, capture pipelines. The kind of project I’ve attempted and abandoned a dozen times over twenty years. Configuration hell. Tab explosions. Forgetting what I was doing mid-task, losing the thread, starting over.

This time it worked. Not because the technology got easier — Proxmox and GPU passthrough and Docker networking are still exactly as fiddly as they’ve always been. It worked because of how I built it: in continuous conversation with an AI that could hold the context I couldn’t.

This is about that process. What failed, what worked, and what it might mean for the relationship between humans and tools.

The Wrong Tool

I started with Google’s Gemini. It seemed reasonable — I was already in the Google ecosystem, the context window was enormous, and the benchmarks looked good.

What I got was this:

THE VIVIFICATION MANDATE
The Cartesian Trap: Why modern efficiency is often a measurement of how quickly we can burn out.

I had asked for help structuring documentation for a home AI system. What I received was a 15,000-word manifesto that called my greenhouse “The Petrichor Node,” referred to the server as “The Slow Giant,” and divided my house into cognitive states called “The Sentry (Day)” and “The Dreamer (Night).”

I wrote a foreword apologizing for the output — this was a document to share with friends who have similar interests, an attempt to get my ideas in one place better than the over-the-top philosophical ramblings they endure from me in person. Instead I had to include this:

“When the text says ‘The Petrichor Node,’ read it as ‘This specific place.’ When it says ‘The Steward,’ read it as ‘The script running in the basement.’ Do not get hung up on the goofy style and syntax.”

The technical help wasn’t better. Confident answers about configuration steps that didn’t work. Losing track of what we’d already tried. The interaction pattern felt like generating at me rather than working with me.

The irony wasn’t lost on me: the Gemini output was philosophically about avoiding AI slop and maintaining human capability. But it was itself the slop it warned against. Performing depth rather than having it.

What Changed

I switched to Claude almost by accident — a friend’s recommendation, nothing more systematic than that.

The first thing I noticed was mundane. Mid-session, I got pulled away. When I came back and asked “what were we doing again?” — we just continued. The context was there. The thread picked up.

This sounds trivial. It isn’t.

For my entire adult life, the configuration nightmare hasn’t been the individual steps. It’s been the derailments — both kinds. Life interrupts, you come back two days later, and you’ve lost the thread. Or: you’re trying to fix the Docker storage location and suddenly you’re four layers deep in containerd documentation, two hours have passed, and you’ve forgotten you were originally trying to deploy Stable Diffusion.

With context that persists, both failure modes just… stopped happening. I’d get pulled away and come back to continuity. I’d go down a troubleshooting rabbit hole and we’d find our way back to the actual goal. The thread didn’t break.

I have executive dysfunction. I always have. The configuration death spiral isn’t a personality flaw I can discipline my way out of; it’s how my brain works. I’ve built elaborate systems of notes and todos and documentation to try to compensate, and they help, but they’re overhead. They’re maintenance. They’re another thing to lose track of.

This was the first time I’d had an LLM assistant. It was also the first time I’d ever had stable infrastructure that I fully owned. The two arrived together, and I’m not sure I can fully separate which one made the difference. But I don’t think I need to. They enable each other.

The Build

The technical details matter less than the pattern, but some specifics give the shape of the thing:

The hardware: A Dell PowerEdge R730 I picked up used. Two Tesla P40 GPUs — 24GB of VRAM each, 48GB total. These are deprecated enterprise cards with architectural limitations that make them slow by current standards. But they’re mine, running in my basement, not metered by the hour in someone else’s data center. About $600 for 48GB of VRAM.

What we built: A VM for GPU workloads running Ollama and Open WebUI for local model inference. I’ve tested Stable Diffusion and AudioCraft on it too — largely experiments to see what the hardware can do. A separate VM for non-GPU services: Silverbullet for knowledge management, behind Nginx with TLS. A mesh network using Wi-Fi HaLow routers. Capture pipelines from phone to server via Syncthing. VPN setup for remote access.

The errors were the useful part. Secure Boot was blocking NVIDIA driver loading — symptom was “Operation not permitted” on modprobe, which doesn’t obviously point to Secure Boot as the cause. The root disk filled up because Docker was writing to the wrong location while containerd used the system SSD. Silverbullet refused to run without HTTPS, which isn’t documented prominently. The first Docker image I tried for image generation listened on a port that wasn’t network-accessible.

Each of these cost hours. Each was solved in conversation, with the context of what we’d already tried, what the system state was, what symptoms we were seeing. Not magic — just the compounding advantage of not losing the thread.

Timeline: January 11th to January 25th. Twenty-two major sessions. About 1,400 messages. A month of work that in previous attempts would have stretched across a year before being abandoned.

What I’ll admit: There’s probably spaghetti in there. Configuration decisions made under pressure, workarounds that became permanent, things that work but aren’t elegant. The hope is that as I get more capable — or as the tools do — I can go back and clean it up. For now, it runs.

I’m also experimenting with how to connect all these pieces together without what people might consider “real” domain expertise in infrastructure architecture. That’s part of the point. I’m trying to figure out how to put these capabilities in the hands of people who aren’t sysadmins, and the only way to learn what’s hard about that is to be one of those people.

The Load-Bearing Part

Here’s what I realized about three weeks in: the thing that made the infrastructure project work wasn’t just “Claude is better at technical help than Gemini.” It was a methodology I hadn’t been able to sustain before.

The pattern: I come to conversations with plans, problems, half-formed ideas. We work through them. When I get derailed — whether by life interrupting or by a troubleshooting rabbit hole that ate two hours — we pick up where we left off. The context persists.

For twenty years I’ve started projects like this and lost them in the gap between sessions. Not because I lost interest, but because I lost the thread. The activation energy to reconstruct where I was exceeded the energy I had available. So things died.

The infrastructure build was the first proof that this pattern could carry a complex project across weeks, across interruptions, across the kind of chaos that usually kills my momentum.

Which matters because: everything I’m trying to do with the greenhouse and surrounding land requires the same thing.

Growing plants is a long-feedback-loop endeavor. You prune a tree in February and see the result in August. You make a soil amendment and wait a season. The knowledge accumulates slowly, from observation, from failure, from noticing what you noticed last year. If you can’t maintain continuity across those gaps, you can’t learn.

Supporting neighbors who want to learn to grow — same pattern. Someone texts me a photo of yellowing leaves. I need context: what did we plant there, when, what’s the soil situation, what did we try last time? If that’s scattered across messages and memory, I’m useless. If it’s accumulated and surfaced, I can actually help.

The community coordination vision — one skilled person supporting many learners, beds tracked and monitored, kids earning a few dollars helping elderly neighbors who can’t bend down as easily anymore — that’s the pattern scaled up. Capture messily from many sources. Synthesize. Surface what’s relevant to who needs it.

None of that was buildable if I couldn’t even get the server running.

So: I got the server running. And the methodology that made that possible is the methodology that makes everything downstream plausible. Not certain — I don’t have an orchestrated system yet, this growing season will be partly improvised, the neighbor nodes are future work. But plausible in a way that my projects usually aren’t by the time I’m a month in.

From Configuration to Capture

The infrastructure build had its own capture pattern: I took pictures of the monitor hooked up to the server’s VGA port with my phone. Once we had SSH configured, I copied and pasted from the terminal. Photos went to my phone, synced to the server, got referenced in our conversations.

That worked for configuration because I was at a keyboard anyway. But when the natural language interface starts supporting other skills — growing, maintenance, observation while working — the capture needs to be different. Hands in soil, attention on the task, can’t stop to type.

The vision is frictionless capture from wherever I am, in whatever I’m doing. One inbox, many entry points. The device fits the context.

What I’m working toward:

A payphone mounted outside the greenhouse. An actual decommissioned payphone, weatherproof by design, connected to the server through an Asterisk PBX. Pick it up, dial *1, speak what you’re observing, hang up. It’s in the inbox.

Foot switches at workbenches — industrial pedals that activate overhead mics. Step on the pedal, speak, release. No hands required.

Analog rotary phones throughout the house, all flowing to the same place. Dial *1 to log a thought. Dial *3 to hear the day’s briefing read aloud.

A Pi-based voice box at the bedside — just a button and a mic. Press, speak, press again. Beep confirms capture. No screen, no notifications, no temptation to check anything else.

The whimsy is intentional. A payphone outside the greenhouse signals what kind of project this is better than explaining it. It’s not smart-home-sleek. It’s salvage and character and tools that fit hands.

What’s Actually Running

I want to be honest about the current state, because the Gemini document I showed you earlier performed a completed vision that didn’t exist. That’s exactly the failure mode I’m trying to avoid.

Running now:

• Server infrastructure: Proxmox host, GPU VM, services VM
• Local models: Ollama serving various open-source models, tested and working
• Knowledge system: Silverbullet accessible across devices
• Mesh network: Three Haven nodes operational
• Basic capture: Phone syncs voice notes and photos to server
• Greenhouse: Structure complete, plants coming through

Coming online this year:

• Automated transcription (Whisper on the GPUs) and routing to inbox
• Orchestration connecting the pieces
• Full tracking of what’s planted where, when, what happened
• Morning briefings that synthesize what I captured yesterday
• First physical capture endpoints (probably the Pi voice box first)

Future / vision:

• Neighbor nodes with privacy-preserving sensors
• The payphone, the foot switches, the analog phone network
• Community coordination: kids getting notifications that a neighbor’s tomatoes need watering — connecting them to neighbors who need the help, building relationships across generations
• Eventually, real-time inference for questions while I’m working — but that’s far down the road, probably part of a more general natural language interface that can query specific resources

That progression is real. The vision isn’t vaporware — it’s downstream of capabilities I now actually have. But I’m months away from the orchestrated version, and honest about that.

Two Layers, Two Speeds

The system I’m building has two different layers that move at different speeds:

The slow layer is the server in the basement. The P40s doing overnight batch processing. Capture during the day, synthesize while I sleep, surface what’s relevant in the morning. This is what deprecated hardware is good for — patience rather than speed. Knowledge accumulating across weeks and seasons.

The morning briefing isn’t just summaries of what I captured — it’s based on monitoring feeds overnight. News sources, research, developments in the tools I’m using. The system watches the landscape while I sleep and tells me what shifted that matters for my situation.

The fast layer is real-time reasoning. Maintaining flow during work sessions. “What were we doing?” answered immediately. The collaboration that keeps the thread from breaking.

Right now, the slow layer is mine. It’s in my basement, I control it. The fast layer is Claude — cloud-hosted, subscription-based, subject to Anthropic’s decisions. I’m building toward local and human-scale infrastructure, and the tool that made it possible is neither local nor fully in my control.

That’s a real dependency. The methodology and accumulated knowledge should survive if Claude changes or goes away — the notes exist, the documentation exists, the patterns are established. But I’d be lying if I said I could route around it today. The fast layer is load-bearing infrastructure that I don’t own.

Maybe local models catch up. They’ve improved dramatically; maybe in two years something running on my hardware can fill this role. Maybe something else emerges. For now, I’m building what I can build, aware of where the vulnerabilities are.

The Hardware Situation

The P40s are deprecated enterprise cards — architectural limitations that make them too slow for the AI gold rush. That’s why they were available at all. About $600 for 48GB of VRAM, which is still useful even if the inference is slower than current hardware.

The bubble around AI hardware is real. Hyperscalers buying up wafers. NVIDIA focused on datacenter because that’s where the money is. The prosumer AI cards that would be perfect for self-hosters basically don’t exist right now.

I’m part of this, by the way. Working with Claude means using datacenter infrastructure that’s part of the demand distorting the market. Maybe that’s hypocrisy, maybe it isn’t — I’m not too concerned about the label. I’m more interested in building toward something better while using the tools that exist now.

What I’m watching for: Unified memory architectures. Apple’s M-series chips point a direction — CPU and GPU sharing the same memory pool, no VRAM bottleneck. I don’t want Apple’s walled garden, but I want that architecture pattern from someone.

I can’t personally track every development — NPUs, TPUs, AMD’s moves, whatever else emerges. That’s part of what the morning briefing is for. When something changes that matters for my situation, I want to know.

For replication: When I eventually help stand up a new site, the specific hardware will be contextual to whenever that happens. The pattern is what transfers: slow batch processing for accumulation, faster inference for interaction. A site set up in 2027 might use completely different hardware filling the same roles. What I’m documenting isn’t “buy P40s” — it’s the capability tiers and how they work together.

For the Gardener Who Has No Interest in Servers

This isn’t about replacing community gardens. It’s about getting gardens into the nooks and crannies they can’t reach yet.

A community garden requires land, organization, coordination, volunteer hours. It’s wonderful and I’m not competing with it. But most people don’t live near one, don’t have time to commute to one, won’t join one.

What I’m building is different — not simpler, but distributed differently. My greenhouse produces plant starts in spring. Those go to hydroponic systems and beds in neighboring yards. But it’s not just “here are some plants, good luck.” The infrastructure maintains context: what was planted where, when, what amendments were made, what problems arose and how they were addressed. A documented history for each growing site.

When something goes wrong, a neighbor can text me a photo. I don’t have to rely on memory or ask them to reconstruct the timeline. The system surfaces the relevant context — this bed, planted in March, amended with this, showed these symptoms in April. That’s the “digital twin” aspect: a running record that makes distributed knowledge actually useful instead of scattered.

One skilled person, supporting many learners, in a way that accumulates knowledge over time. The technology is infrastructure for that relationship, not a replacement for it.

And here’s the thing for gardeners who never want to touch any of this: you benefit anyway. Every person who starts growing food because the support structure exists, every kid who earns pocket money helping a neighbor and learns that food comes from plants and effort, every yard that converts some lawn to food production — that’s culture shifting. The goal isn’t to make everyone into a technologist. The goal is to make growing things a more normal part of how people live.

For the AI Enthusiast Wondering What a Greenhouse Has to Do With Anything

Most local AI content focuses on benchmarks and throughput. How many tokens per second. Which model fits in VRAM. Optimization techniques for faster inference.

I’m doing something different: batch processing, knowledge accumulation, appropriate rhythms.

The P40s I’m running are deprecated enterprise cards with architectural limitations. By current standards, they’re slow. But they have 48GB of VRAM combined, and they’re mine, and they’re running 24/7 in my basement without metered costs.

What I’m learning from the greenhouse applies directly to running models: patience matters more than speed. Overnight synthesis is fine — I don’t need instant response, I need good response by morning. Knowledge that accumulates across sessions is worth more than fast answers that don’t compound.

The methodology is: capture during the day, process overnight, surface what’s relevant in the morning. That’s not a limitation of my hardware — it’s appropriate technology for what I’m actually trying to do.

If you’re building local AI infrastructure, the question isn’t just “how fast can I run Llama” — it’s “what am I actually trying to accomplish and what rhythm serves that?” The answer might be slower than you think.

What We Lost and What We Might Get Back

Within living memory, a significant portion of the population was directly involved in growing food. In 1900, over 40% of the U.S. labor force worked in agriculture; by 1950, that had fallen to 12%, and by 2000 it was under 2%.¹ My mother grew up on a 100-acre family farm — a boomer at the tail end of this transition, part of the generation that watched agricultural knowledge stop being passed down because it stopped being necessary for survival.

That was often a hard life. The move away from it brought real gains: less backbreaking labor, more security, more options. I’m not romanticizing subsistence farming or suggesting we should return to it.

But something was lost too, and it wasn’t just romanticized nostalgia.

There’s research on this now — a growing body of evidence suggests that interaction with plants reduces stress, improves mood, and provides a kind of grounding that’s hard to get otherwise.² Hospital patients with views of gardens recover faster. School gardens improve student engagement and learning outcomes.³ This isn’t mysticism; it’s measurable. Something about working with living things, seeing growth respond to care, being in a feedback loop with the seasons — it does something for people.

Most of us lost that feedback loop. Not through any individual choice, but through the aggregate of a century of choices that made sense at the time. Cities grew. Agriculture industrialized. The knowledge of how to grow things stopped being common because there was no longer economic pressure to maintain it.

I’m not trying to reverse that. I’m not proposing a back-to-the-land movement — those have been tried, and they mostly failed, for reasons that were predictable in retrospect. I’m not interested in communes or survivalism or ideology.

What I’m interested in is whether technology can give some of that back without requiring people to give up what they gained.

A neighbor who wants tomatoes doesn’t need to become a farmer. They need a bed, some starts, and someone to text when the leaves turn yellow. The knowledge can be distributed. The labor can be shared across people at different life stages — kids who have energy, elderly folks who have time and knowledge, working adults who have money but not hours. The infrastructure can make coordination possible that wasn’t before.

Maybe this is a fool’s errand. I genuinely don’t know if it works until I try it. But the hypothesis is: we can regain some of what was lost — the connection to growing things, the relationships with neighbors, the sense of capability — without pretending the last century didn’t happen.

The dome in my backyard is an experiment in that hypothesis. The server in my basement is what makes the experiment possible to run.

For the Neighbors

From the street, I know what this might look like. A dome. Unusual activity. Someone clearly doing something in there.

What I’m trying to build is an invisible supportive layer that becomes visible only through outputs: a garden that works, plants that thrive, neighbors who want to grow things getting the help they need to succeed. Kids connecting with elderly neighbors through small tasks — watering, weeding, harvesting — earning pocket money while building relationships that wouldn’t otherwise exist.

Not long ago, this neighborhood was farmland. I’m not trying to bring that back, and I’m not trying to change what the area has become. I’m trying to enhance it at a scale that fits — a single-family home doing something useful with its tenth of an acre, in a way that might help the people nearby do the same with theirs.

What This Isn’t

I should be clear about what I’m not claiming.

This isn’t “anyone can do this.” The hardware cost real money. The time cost real time. I had a crypto windfall that bought the equipment — and that runway is now spent, so this is make-or-break time for whether the investment pays off in capability.

This isn’t “I solved executive dysfunction.” I still get overwhelmed. I still lose days. I still have all the same problems. What I have is a tool that happens to fit how my brain works, in a way that previous tools haven’t. That’s personal, not universal.

This isn’t “AI will fix everything.” Most of what I’m trying to do requires physical presence, human judgment, and community relationships that no model can provide. The AI is infrastructure — infrastructure I’m trying to make into convivial, human-scale, appropriate technology. But infrastructure. The work is still the work.

This isn’t “fully local and self-sufficient.” I named the dependency: the fast layer is Claude, and I don’t control it. I’m building toward technology that’s fully within the envelope of the people it serves. I’m not there yet.

And this isn’t finished. A month in, I have proof of concept for a methodology. I don’t have the orchestrated system, the full tracking, the payphone on the wall. I have plants coming through this season, and the bones of something that might work, and honest uncertainty about what I’ll actually ship.

That’s where I am. A greenhouse, a server, and a way of working that — for the first time in twenty years — seems like it might actually hold together.

[January 2026 — Salt Lake City, Utah]

Sources

¹ U.S. Census Bureau and USDA Economic Research Service historical data. Agricultural employment fell from 41% of the labor force in 1900 to 12% by 1950, reaching 1.9% by 2000. See also Dimitri, Effland, and Conklin, "The 20th Century Transformation of U.S. Agriculture and Farm Policy," USDA ERS Economic Information Bulletin No. 3, 2005. ↩

² For comprehensive reviews, see: Soga, M., Gaston, K.J., & Yamaura, Y. (2017). "Gardening is beneficial for health: A meta-analysis." Preventive Medicine Reports, 5, 92-99. Also: Hall, C., & Knuth, M. (2019). "An Update of the Literature Supporting the Well-Being Benefits of Plants." Journal of Environmental Horticulture, 37(1), 30-38. ↩

³ Ulrich, R.S. (1984). "View through a window may influence recovery from surgery." Science, 224(4647), 420-421. For school gardens: Williams, D.R., & Dixon, P.S. (2013). "Impact of Garden-Based Learning on Academic Outcomes in Schools." Review of Educational Research, 83(2), 211-235. ↩