Trigger warning: This post (and potential series) is going to discuss some AI/LLM stuff. That’s only a part of what’s going on, but if you hate all things LLM, then maybe this post isn’t for you.
Reasons for Building a PC
I haven’t built a PC in probably 25 years now. Even then, I had the kind assistance of two friends who knew a lot more about PC building than I ever did. More recently I’ve been interested in building a PC for a variety of reasons:
- The intellectual exercise of building something, mostly out of parts, all by myself
- Having a PC running Ubuntu / Linux, something I’ve toyed with now and then, but never really tried in earnest. I already eschew most closed source software in favor of open source alternatives1 , so moving to an open source operating system seems almost natural.
- Some of my personal information currently resides on a computer I use for work – and I’d like to create a clean separation of these things for a variety of practical reasons.
- Having an inexpensive and dedicated box would let me sandbox certain programs, try out different new ideas, and still let me access all this information and processing power remotely, if set up properly.
- Yeah, I wouldn’t mind an Iron Man style Jarvis virtual assistant / butler, but the technology is still a fair bit aways as are my own skills and budget for such things.
Building in Parallel
I’ve taken to “building in parallel” for several recent projects. With the recent remote temperature sensor system, I tried taking measurements with an infrared heat gun, then an ESP32C3 module with lots of wires, then several Wemos D1 Mini modules each separately sending data back to my PC. While the ESP32C3 systems were still (mostly) working, I recognized some of the weaknesses and failure points and began working on the Wemos systems and didn’t take the ESP32C3 sensors offline until the Wemos sensors were up and running. There was certainly overlap for a little while, but this also let me calibrate both systems. I enjoyed working on these things, so building an entire second set of sensors in exchange for doubling the price and quadrupling the soldering wasn’t a horrible experience. I’m reminded of a quote from Contact, “First rule in government spending: why build one when you can have two at twice the price?”
As with those sensors, I’ve had some automated projects running constantly lately. Something that downloads a huge work calendar daily, a few simple reminders, something that sends me reminders over a messaging platform, and a few other assorted digital doodads. I’m keeping all of these online, on my work PC, until I can get the Linux box going – then I’ll look forward to closing these down.
Pricing Things Out
As with nearly any learning experience, you end up paying for it one way or another – money, tuition, wasted attempts and materials, time. Partially because I’m just starting out and partially because I’m fairly frugal, I set a budget of ~$2502 If I’m going to spend any money at all, I wanted something that would work – but I’m not ready to drop $5k on RAM/GPU/newest processors. My needs are still modest – and largely covered by my equally modest work computer. It’s an Intel i7 with 32GB RAM in a very small mini PC case, handles email, big spreadsheets, graphic programs, python projects, and even 3D modeling and rendering with OpenSCAD if you give it enough time. It cost $550 before tax, new from Amazon, nearly 4 years ago, and is still going strong.
For this project I contemplated going with a fanless headless system, but these seemed to require special aluminum cases and appeared to be little harder to come by. I thought about tracking down some e-waste resources, Facebook Marketplace, and Craiglist – but all of these require a fair bit of legwork and I feel like they’d also have more than a modest possibility of wasted time. Any new PC would set me back at least $500 through either Amazon or Costco, so I turned to eBay since the rating system provides at least a modicum of quality assurance – if not in the product, at least the seller.
There are a ton of options for older PC’s that can be given a new lease on life with Ubuntu and some extra memory. I’m still pretty biased against Dell, HP, and Gateway for producing sub-par consumer grade computers for myself, family, and friends. Some noodling about and half-hearted research suggested that these brands may produce significantly better enterprise machines than their consumer grade lines. This only irritates me even more, if true. This pretty much left me with Lenovo.
I ended up going with a “Lenovo ThinkCentre M70q Tiny i5-10500T 2.30 GHz” and paid a little extra to have the eBay seller customize it with 32 GB DDR4 RAM total, a 512 GB M.2 NVMe SSD, and throw in the antenna. Besides being fairly accommodating and helpfully communicative, I went with this listing because it fell squarely in my price range / project budget, it already came with the power adapter, they were California, and it could ship very quickly. While I could have easily spent another week or three trying to find a better deal on better and newer equipment, this was always going to be a project running on low end / older equipment and I valued the immediacy of getting to work over the incremental value of a better system. Frankly, just not having to hunt for a compatible power adapter was a huge plus for me.
In the end I purchased the main system from the eBay seller, paid an extra $36 to make sure the system had 32GB of RAM, ordered a used 512GB SSD (they threw in the antenna) from them in a second purchase for $55, a SATA SSD enclosure to re-use an old 500GB SSD, and $7 to grab some HDMI dummy plugs off Amazon. I knew the system would come with 16GB of RAM, but the benefit of “buying up” to 32GB also meant that I knew the machine wouldn’t have two 8GB sticks instead of a single 16GB stick. I could later upgrade from 16GB to 32GB if I had a single 16GB stick to start – but having two 8GB sticks would mean I would later have to get two 16GB sticks.
Specs and Prices
Anyhow, for those of you playing along at home, here are all the parts for my $284.30 computer system.
| Part | Cost | Shipping | Tax | Total |
|---|---|---|---|---|
| M70q Gen 1(i5-10500T, 16GB, Adapter Incl., No SSD), upgraded to 32GB | $185.99 | $5.93 | $14.41 | $206.33 |
| M.2 SATA SSD to 2.5” SATA adapter enclosure | $10.79 | $0.84 | $11.63 | |
| 512 GB NVMe M.2 SSD, antenna | $55.00 | $4.13 | $59.13 | |
| Dummy HDMI Plug | $6.69 | $0.52 | $7.21 | |
| $284.30 |
I didn’t have to buy the SSD enclosure – but it seemed like a very inexpensive way to double the storage and repurpose my old drive. I didn’t realize I’d need to get an HDMI dummy plug, or even what they were, until I tried to “remote” into the headless box and got a fairly ugly looking desktop instead of the slick Ubuntu desktop. This was probably the least necessary of these purchases. If I kept the system to the PC with the 32GB upgrade, ran the OS off my old SSD in the SATA enclosure, and picked up a $7 antenna off Amazon, the system would have only cost about $213. The extra $59.13 doubled the hard disk space and enabled a more graphically pleasing interface. It’s not much of a splurge and this was never a “build a PC for as little as possible” kind of a challenge anyhow.
The final system is probably on par with my existing 4 year old PC at roughly half the cost. I think it was a “decent” deal and I have a system that I can use to offload a lot of projects. If any of these projects end up requiring more processing power, I’m not even really “out” $285 since I can still keep using this system for other tasks (home automation, running daily scripts, etc).
And, yes, I’m trying to use this PC as a basis for a bundle of scripts I can deterministically interact with via voice input / output as well as certain LLM powered tasks. First, a helpful comic.
We’re probably hurtling towards dead internet theory. I can’t search through Google or even DuckDuckGo anymore without having to wade through their AI summaries and then scroll past search results from AI slop. Using ad blockers or setting up a PiHole to block ads and stop tracking from ever touching my PC may impoverish some otherwise worthy websites. However, I simply don’t have to read every advertisement sent in the mail, on the inside of my grocery shopping cart, or every billboard. Yeah, I hate the idea of AI companies boiling oceans of drinking water for their inscrutable purposes…
There’s a quote I love by H.L. Mencken; “Every normal man must be tempted, at times, to spit on his hands, hoist the black flag, and begin slitting throats.” What if the best way to filter out AI garbage is to use an LLM to do some of that filtering for me? 34
Searching through any search provider is kinda garbage these days, but there’s no reason I should have to wade through that AI garbage myself. Gone are the days you can drill down to the kind of search results you want by simply adding “-negative_search_terms”, because there’s probably no financial reason for them to permit this.
But what if I created my own search function? I recently tasked an AI with searching up 10 options for a certain product for me on Amazon and it did a mediocre job, finding 7 dead or unavailable listings. But it also saved me a lot of time looking through dozens of pages of crap when I only needed to look through 10 crap listings. With some tinkering, I think I could make a script that would identify negative search terms, perform the search as requested, then remove those that should have been removed with the negative search terms. I could see this being incredibly useful for Ebay and Amazon purchases.
A preview for next time: After some dithering, much watching of youtube videos, I’m going to give Hermes Agent a shot over OpenClaw.
Project Jarvis- LibreOffice, Thunderbird, OpenSCAD, etc [↩]
- Is it really a budget if there’s a range?
Words to live by [↩]
- And, yes, I have considered the possibility that an AI company’s aim might be to make everyone distrust the entirety of the internet, then sequester everyone into their own little siloed version of digital reality, everyone watching slightly more targeted ads and Marvel movies than everyone else, and any attempt to use another LLM to filter shit out only results in more precise targeting… [↩]
- I went to the hardware store yesterday and was told by someone who worked there in an orange apron that the entire store, inside and out, was covered in Flock cameras. I think this bullshit surveillance is really starting to get to me. [↩]