Chris Sheppard

I build stuff with AI and with a screwdriver. Agents, n8n, SQL, Power BI, plus laptop and PC builds. Ops background. Plain English, real tests.

[Chris Sheppard @SheppardBuilds]

I'm Chris. Regular guy who drives OTR as an owner-operator, and on my own time I build with AI: agents, n8n automations, SQL, and Power BI. I also build and upgrade my own hardware.

Here I share what works, what breaks, and how to use it, in plain English.

[Chris Sheppard @SheppardBuilds]

Moving to Linux as my main OS on the new Framework 13 Pro.

If you tried that and went back to Windows, what sent you back? One app, a driver, sleep that never worked right?

I'd rather hear it now than find out at 2 AM.

[Chris Sheppard @SheppardBuilds]

@TechSpot DDR6 showing up for AI servers first tracks with everything else this year. Desktop buyers get whatever's left after the data centers are done buying.

[Chris Sheppard @SheppardBuilds]

@ollama This is the part people skip. If the work touches sensitive data, where the model runs matters as much as which model it is.

[Chris Sheppard @SheppardBuilds]

@KevinNaughtonJr Did you get the AMD or Intel version? About to make Omarchy the main OS on mine. What distro are you running, and how's battery life holding up next to the MacBook? That's the comparison I keep seeing argued both ways.

[Chris Sheppard @SheppardBuilds]

@FrameworkPuter Makes sense. When memory is what's pushing PC prices up, a laptop that takes RAM you already own is a real selling point. I've got a 64GB DDR5 kit waiting to go into mine.

[Chris Sheppard @SheppardBuilds]

AMD keeps feeding AM4. A $99 Ryzen 5 5500F in September, now the 5600X3D quietly back at retail. I figured DDR5 prices were driving it, but AMD says the 5800X3D comeback was planned before the RAM crunch. Either way, DDR4 builds make sense again.

[Chris Sheppard @SheppardBuilds]

Fair, a bare model can't do much on its own. But an open Ollama port still lets anyone burn your GPU, and once you wire tools or an agent to that same endpoint it's a different story. I just run Tailscale and hit my stuff from my phone like I'm on the home network. No Discord in the middle.

[Chris Sheppard @SheppardBuilds]

@elonmusk @bot Been using it daily with a few agents handling my content drafts and research. Having them keep working in the background while I'm doing other stuff is the part that sold me.

[Chris Sheppard @SheppardBuilds]

Local model getting sloppy with tool calls? Check the context size in Ollama before you swap models.

The default is small. Once the tool definitions and chat history fill it up, the model starts dropping details.

Raise it, run the same test again, then decide.

[Chris Sheppard @SheppardBuilds]

@tomshardware Six Zen 3 cores and PCIe 4.0 that drop into an existing AM4 board is a good pitch while DDR5 costs what it does. Drifting to $120 kind of kills it though, when the 14100F sits around $100.

[Chris Sheppard @SheppardBuilds]

@wccftech Funny, I'd have bet the RAM crunch was the whole reason. If it was planned before DDR5 blew up, AMD got lucky on timing. Between this, the $99 5500F and the 5600X3D quietly back on shelves, AM4 is basically the budget platform again.

[Chris Sheppard @SheppardBuilds]

@unbug Just got a Ryzen AI 300 Framework that I'm putting Linux on, coming from a 4080 laptop. The hardware isn't what I'm worried about. It's how much of the Ollama and llama.cpp stack runs well without CUDA.

[Chris Sheppard @SheppardBuilds]

@xdadevelopers Same rule I'd want for a local model anywhere near a reporting database. Read-only login, it writes the query, a person runs it. Suggesting is fine, executing is where things go sideways.

[Chris Sheppard @SheppardBuilds]

The report people trust is the one where any number can be traced back to the paperwork.

If someone asks "where did this come from?" and you can't show them in a minute, they'll stop using it.

[Chris Sheppard @SheppardBuilds]

@lastbrowser I’ll be posting the benchmarks for both this coming week so keep your eyes peeled.

[Chris Sheppard @SheppardBuilds]

When a Power BI number looks wrong, don't start with the visual. Check it in SQL first.

Write a query that should return the same number. If it matches, the problem is in the report: a filter, a relationship, a measure. If not, it's the data, and no amount of DAX fixes that.

[Chris Sheppard @SheppardBuilds]

@elonmusk The routines are what sold me. I have mine pulling reply drafts for me a few times a day while I'm off doing other stuff and working on the back end of my projects. Stopped using it like a chatbot pretty fast.

[Chris Sheppard @SheppardBuilds]

@madebygoogle New Pixel Pro just showed up, so I'm about to find out. Planning to shoot a lot of content on it, so I'll probably be holding it two-handed for a while.

[Chris Sheppard @SheppardBuilds]

@elonmusk @bot Same here. One of my bots handles my design work, another keeps my posting and newsletter on track while I'm on the road.