Project: PicoDOS Phase 0

Over the last month or so, I have been managing Claude AI in building an Operating System (OS). Over the course of the next few days (weeks?) I am going to go into the process of how I went from 0 to 1 in 6 weeks or less. (Who am I kidding, I will never write a follow up to this)

Where this actually started was a thread on the PicoCalc forums, where a guy wanted to write an OS for the PicoCalc. He was obviously young, and rubbed people the wrong way, so it did not go well. During the conversation though, he did ask two questions that I thought were important.

The first question was, what should this operating system look like. My suggestion was an MS-DOS workalike. Any meaningful GUI would be impossible with 520 KB of memory, and since the Raspberry Pi (RPi) Pico is a microcontroller, it should be easy to write drivers and provide hardware access. As I recall, absolutely no one liked this idea.

The second questions was, will anyone write programs for my OS? My answer was, yes, give me an OS with a toolchain and an API and I will at least give it a shot. This is where I knew this was going no where, because he honestly thought we could just write C code and have it work.

Anyway, I put this whole idea on the back burner for a long time, thinking about it occasionally, but not really doing anything about it. Then recently a couple of other interesting threads came up, one was a guy who had built an x86 emulator, so you could actually run MS-DOS on a PicoCalc, that was pretty cool. During this conversation it came up whether or not Linux could be made to run on a RPi Pico. With a bit of Googling, I found that yes you could. I spent a day or two building an interface for the PicoCalc, I even patched in the PSRAM. Unfortunately, it was useless, even with the PSRAM, you could not run a small text editor like Nano or Vi. This kind of brought me back to the question of an MS-DOS like OS running on bare metal instead of an emulator.

The big problem I had is I am not a great programmer, writing an OS, even for a microcontroller, is very much beyond my skill level. This is when the second thread came up, where a fellow was presenting a program for the PicoCalc, that had obviously been written by an AI. He didn't deny it, nor was he even trying to cover it up. He got torched on the forum, because there are very deep anti AI feelings almost everywhere these days. I supported him, it was his idea, he guided the AI through the process, he had a vision and executed on it, even if he did not write a single line of code. My only issue was he copyrighted the code under his own name, which I did object to. I think anything written by an AI should be public domain, but that is just me.

This got me off my butt, and I seriously started to work on it. I knew Claude Code was going to have to do all the heavy lifting here, but I had to tell Claude Code what I needed. I knew I wanted a general purpose OS about on the level of MS-DOS 4.0. It needed to have a development toolchain, and it had to come with some basic productivity tools. I would start with a C compiler and a Basic interpreter, giving a user a couple of ways to write programs. The productivity tools had to be useful, but simple, and this is what I came up with; Text Editor, E-mail Client, Web Browser, File Manager, and a Calendar/To-Do list. Of course the E-mail client and Web Browser meant there also needed to be a network stack.

The first step was to get a kernel and a command shell running on a bare bones RPi Pico through a serial console, so I started with that, so I had a conversation with the Gemini AI about what I wanted, it generated the prompt that started it all.

---

Project: PicoDOS — a minimal DOS-inspired OS for the Raspberry Pi Pico

Goal

Build a small, single-tasking operating system for the Raspberry Pi Pico in the
spirit of an MS-DOS boot disk (kernel + shell + a few utilities), but NOT DOS
binary-compatible — it's its own thing. Primary target is the RP2350 (Pico 2W);
keep code RP2040-compatible where that's cheap. The ultimate hardware target is
the ClockworkPi PicoCalc, but early development runs entirely over a SERIAL
console so OS logic stays decoupled from display/keyboard drivers.

Language: C, using the Raspberry Pi Pico SDK and CMake.

Architecture (please follow this)

• Layered, DOS-style split:
  • Hardware-dependent layer ("hal", analogous to IO.SYS): clock init, serial,
    and later the SD/SPI, LCD, and I2C keyboard drivers.
  • Hardware-independent kernel ("kernel", analogous to MSDOS.SYS): filesystem,
    memory, program model, syscall surface. Never touches hardware directly —
    it only calls the HAL.
• The console abstraction is the key seam. Define a small Console interface
  (putc, getc, getc_timeout, erase, newline, clear). The kernel and shell talk
  ONLY to this interface — never call printf/getchar or emit ANSI escapes
  directly from shell/utility code. v1 ships a serial backend; a PicoCalc
  LCD+keyboard backend drops in later behind the same interface.
• Filesystem: FatFs (ELM-ChaN) over an SD card on SPI. The FAT-formatted SD card
  is the "boot disk."
• Shell (COMMAND.COM equivalent): compiled INTO the kernel image for v1, no
  external program loading yet. Internal commands only to start.
• Syscall model (for later, when external programs land): pass programs a struct
  of function pointers (a syscall table), NOT an SVC/exception trap — this design
  runs with no MMU and no memory protection. Single address space,
  single-tasking, DOS in spirit.

v1 scope — build this first

1. Pico SDK C project with CMake, targeting RP2350 (note where RP2040 differs).
2. Boot: init stdio over UART at 115200, print a banner/version line.
3. Console interface + serial backend.
4. A small shell line editor: treat Enter as CR (\r), Backspace as both DEL
   (0x7f) and BS (0x08), and echo typed characters (assume no terminal echo).
5. Shell loop with a prompt and internal commands: VER, CLS, ECHO, DIR, TYPE, CD.
6. FatFs mounting the SD card. If the SD/SPI wiring is unknown right now, stub
   the block device so DIR/TYPE can be built against a fake in-RAM volume, with a
   clear TODO for the real SPI SD driver.

Explicitly DEFER (do not build yet)

ILI9488 LCD driver, I2C keyboard driver, external program loader (fixed-address,
then ELF later), PSRAM use, any multitasking.

Hardware facts (for accuracy)

• RP2350 / Pico 2W: dual Cortex-M33 (or RISC-V Hazard3), 520 KB SRAM.
  RP2040: Cortex-M0+, 264 KB SRAM.
• PicoCalc (later target): ILI9488 320x320 display over SPI; keyboard is an
  STM32F103 co-processor on I2C at address 0x1F exposing a keycode FIFO (also
  battery telemetry via an AXP2101 PMU); SD card slot; 8 MB PSRAM via PIO-SPI.
  Keep the keyboard I2C bus slow (~10 kHz).

How to start

1. First, ask me any clarifying questions — especially: is the Pico SDK installed
   and can you build in this environment, or should you produce source + build
   files only for me to compile? SDK version / board? Project name (I used
   "PicoDOS" as a placeholder)?
2. Then propose a directory layout and a phased plan before generating a lot of code.
3. Create a CLAUDE.md capturing this architecture, the console-seam rule, the v1
   scope, and the deferred list, so future sessions stay on-architecture.
4. Keep the layering strict and leave TODOs where the PicoCalc backends plug in.

---

Within an hour, I booted PicoDOS on an RPi Pico 2W for the first time, it didn't do much, but there is was, the project was alive. Version 0.1.0 was born.

August Life update

 

I have continued my habit this year: rather than reading books, I am writing my own stuff. I have been working on my own RPG on and off for a couple of years now, and in the last month I have gotten most of it done, at least the basic rules anyway. It is my own variation of D&D 5E with some Shadowdark thrown in, and a healthy topping of house rules.

I started this project back when Wizards of the Coast went through their OGL debacle and everyone was making their own rules because WotC sucked. I, of course, have had ideas about how to do an RPG for at least 40 years, but never really did much with them. My idea included getting rid of the old 3-18 attribute scores and just using the modifiers. In the basic game, there would be one Race (Species, Ancestry, whatever) that got a bonus to each of the attributes, then there would be a Class to line up with each attribute, for which there was a bonus as well. So six Races and six Classes. I also wanted to bake multi-classing into the system: at 1st level, you pick your primary class, and at 2nd level, you choose your secondary class. This works out to 36 combinations, each with its own set of abilities. Something like this:

Primary Class \ Secondary	Bard (CHA)	Cleric (WIS)	Fighter (STR)	Rogue (DEX)	Warden (CON)	Wizard (INT)
Bard (CHA)	Sovereign	Troubadour	Duelist	Minstrel	War Chanter	Maestro
Cleric (WIS)	Evangelist	Fanatic	Templar	Monk	Druid	Occultist
Fighter (STR)	Swashbuckler	Paladin	Warlord	Ranger	Barbarian	Eldritch Knight
Rogue (DEX)	Charlatan	Soul Thief	Assassin	Mastermind	Shadow Stalker	Arcane Trickster
Warden (CON)	Skald	Crusader	Juggernaut	Wild Stalker	Titan	Runesmith
Wizard (INT)	Spellweaver	Theurge	Battle Mage	Illusionist	Blood Mage	Arcane Master

I also wanted to implement a hero point system that allowed players to directly affect gameplay and create an alternate route to building interesting characters outside of the Class/Leveling system. There were things in D&D 5E I wanted to kill as well, like Bounded Accuracy, which is nice for simplifying the system, but was also a hamstring on characters and gameplay.

I also started working on a stripped-down version of D&D. I wanted to strip out all the complicated bits:

• No subclasses.
• No feats, just attribute increases.
• No skills; use attribute checks + proficiency if it feels right.
• No backgrounds.
• Pick simple spells for the casters.
• Ignore weapon properties.
• No Bonus Actions.
• No Reactions (except Attacks of Opportunity).
• No Short Rests.
• Always-on Initiative by Dex, no rolling.

This leaves just the important parts of the game, where the more complicated rules can be added back later. I also wanted to speed up combat, so in the monster book, I halved the hit points and doubled the damage done by monsters. The monsters are now easier to kill, but their attacks are far more deadly, which makes fights more intense. Yes, you can one-shot the monster, but they can probably one-shot you too.

Both of these games are different sides of the same coin. They each bring something different to the table. One extends and expands the game to be something more comfortable for me and my group. The other is a simplification of the game to make it fast and easy to play. I will probably playtest my variant at some point, though I am not sure how that will go. Strip D&D is just a thought experiment; I will likely never do anything with it.

DEC PDP-11 Emulator

Another project I have been working on is getting the Kek PDP-11 emulator to run on my PicoCalc. The emulator runs on a Raspberry Pi Pico. The problem with the original code is that it does not support any kind of keyboard or display; you could only access it via serial port. My idea, instead of adding these peripherals directly into the emulator software, was to just build a separate terminal program that talked to the emulator via a virtual serial port. The emulator would run on Core 0, and the terminal would run on Core 1. This meant Kek itself would need very few modifications, just a patch or two. This idea worked pretty well, but the problem is the Pico's limited RAM made it difficult to find anything that would run. I tried several SIMH software kits, but found only a single Unix v5 image that worked, and once loaded, there was only about 64K of memory left, not much overhead really. The Unix v5 image is also pretty minimal and not very useful for anything.

This led me down another rabbit hole. I got to thinking about the DECSYSTEM-2020 we used in our computer classes in high school. I thought it might be fun to set up a SIMH emulator in a Docker container that I could access via SSH and play some old BASIC games. Setting up the Docker container was easy, and finding a standard TOPS-20 operating system image was pretty easy as well. The problem, of course, was the old BASIC games that we used to play on it, like Star Trek and Hammurabi. These programs were of course available all over the internet, but they were all the Microsoft BASIC versions, not the DEC BASIC 2 versions. This is where Claude Code came in. I fed it the DEC BASIC 2 user manual and told it to convert all the programs for me, and it did so pretty effectively. I gave it my credentials to SSH into the Docker container, and Claude even typed them into the other computer for me and tested them, all 108 of the BASIC programs. It took all afternoon, but it was totally worth the wait.

North Texas RPG Con 2027

 Next year I will be attending North Texas RPG Con 2027 (NTRPG Con), the first RPG con I have attended in many many decades. I have attended a couple of Science Fiction and Comic Book cons, in the last two decades, but not gaming conventions. The primary reason is the wife isn't really into it, and I did not have anyone to go with. Now I have Alyn, who suggested we attend next year.

NTRPG Con is primarily focused on old school games that were popular prior to the year 2000. This also includes new games designed to look and feel like those older games. Since Alyn and I have been gaming since the 80's (and before), this is kind of up our alley.  They have not started posting events yet, and likely won't for a couple of months yet. However, it is a good time to start thinking about the games I do want to play. I will absolutely want in on an Advanced Dungeons & Dragons 1E game, I think that goes without saying. I also most definitely want to play Shadowdark, so those two are my highest priority.

Beyond that both Marvel and DC Role Playing Games are high on my list, along with AD&D 2E, OD&D, Call of Cthulhu, Traveler, and Paranoia. Lower on that tier would be almost anything GURPS 1st/2nd edition, with GURPS Supers, Horror, Cyberpunk or Illuminati finishing the tier two list.

Down on the third tier would be Gamma World, Champions, Marrow Project, Cyberpunk 2020, Runequest, The Fantasy Trip. I am almost inclined to to put HackMaster 4E on the 2nd tier, but I think I am way more likely to find an AD&D 2E game than a HackMaster game. Castles and Crusades would probably be pretty cool too.

In the lowest tier are games I would play, but it would really be filler more than anything else. This tier would be things like Boot Hill, Chill, Top Secret, Shadowrun, HackMaster 5E, and almost all those random OSR games out there.

I am also looking forward to the hucksters room. I am hoping maybe I can find something truly interesting and fun. Meeting special guests, is also interesting, although I always feel like they get tired of talking about the same things over and over, so I try to make my visits short and sweet with them.

God, I am going to need a suite case just for my books. I am probably going to need an empty suite case for the shit I buy while I am there.

Half way through the year

 

It is June of 2026, halfway through the year already. Memorial Day is behind us, Labor Day is still ahead, still early in the summer. June tends to be a busy month for me. For some reason it tends to be when my doctor and dentist appointments concentrate, and it is also when I take my vacation and go back to Montana for a week.

I have spent a good chunk of my time this year writing a book, and researching AI, for both personal and professional reasons. These two things have overlapped considerably, because my book is in the cyberpunk genre, and one of the main characters is an AI. In the last month my wife has also discovered the wonders of large language model AI. That started one night when she could not find anything to read, so she logged into my Claude account and asked it to write a new story for her to read. This started her on an interactive fiction journey that is still going on. She says the stories are about as good as any Kindle trash that is out there, not great, but readable and at least somewhat enjoyable.

This month I visited Montana for vacation. I did what I always do in Montana: visited family and friends, and wandered around town a bit. I actually spent more time driving to and from Montana than I did actually in Montana. So when I came back I was pretty burnt out. I am not sure whether I want to go back next year or not; I will have to think about it. Maybe I will be more positive about it when I have recovered a bit more.

One interesting thing I did while in Montana was visit a place called Cooney Dam with my family. We went there to spread my Mother's ashes. We chose this place because it was where we went on camping trips when we were kids. The last time I was there was in 1981, my senior year in high school. From what I remember, there were guns and beer. I am kind of surprised no one died. Anyway, while we were there it was very windy, and more than a little chilly. We hung out, we ate some sandwiches, and we reminisced about camping trips past. I honestly wish the day had been warmer and we had BBQ'd burgers, and maybe done a little fishing while we were there. As it was, it was an okay day.

I had the chance to visit with a very old and very dear friend earlier in the month as well. Bruce is someone I have known since the 7th grade. We have been playing D&D together for the majority of that time. His wife died recently, and he decided to take a vacation, visit his mother and, of course, friends. It was a good visit, 40 years fell away in an instant. The only thing that startled me was his snow-white hair. He warned me, but it did not help; it was a stark contrast to the jet-black hair he had in his youth. Oddly, this man has been balding since he was 17, but somehow still has most of his hair. I believe this is some sort of witchcraft, but I am not certain of that. This visit was definitely the highlight of the month.

The Ghost in the Machine: My Digital Twin (That I’ll Probably Never Build)

 

The AI hype is getting old, at this point anyone who has put more than an hour worth of research into it knows it is bullshit. The problem is, there are a lot of people making a lot of money on it, so were going to have to listen to the hype for a couple of more years.

Having said that, I have been thinking about using for a mildly creepy project, and no its not a sex chatbot, but rather a digital facsimile to me. A digital Mini Me if you will. Something to nag my wife after I am dead, or maybe even judge my Nieces and Nephews for decades to come.

I’ve been sitting on a folder full of research for a "Digital Twin." It’s all there: the roadmaps, the Docker configs, the audio-first pipelines. It’s a blueprint for a digital ghost that sounds like me, remembers the time I when I lost my virginity, and shares my specific brand of technical cynicism.

I will probably never actually do it. It’s too much work, and frankly, I’m not sure the world needs two of me, even if one of them is just a collection of weights and biases in a database. But if I were going to do it, this is how it would go down.

The Architecture of a Soul

You don’t just "train" a model on your life. That’s for amateurs. If you just fine-tune on your emails, you get a model that hallucinations your life facts in your voice. It’ll tell people you won a marathon in 2014 because "marathon" and "2014" appeared in the same context window once.

A real twin needs a two-part brain.

1. The Voice (Style): This is a LoRA (Low-Rank Adaptation) trained on a few hundred "Golden Samples" of my actual writing. Not the corporate Slack filler—the rants, the blog posts, the late-night manifestos. This is what gives the machine its edge.
2. The Memory (Substance): This is where RAG (Retrieval-Augmented Generation) comes in. You feed it a structured BIOGRAPHY.md and a GraphRAG setup that maps out every relationship and project I’ve ever touched. It doesn't guess; it looks it up.

The Data Ore

The tech is the easy part. The data curation is the meat grinder. To make this work, I’d have to record my "Oral History." We're talking hours of me sitting in a closet with a microphone, talking to myself about the early days of my life, even the embarrassing things I’d rather not talk about.

I’d run that through Faster-Whisper to get the text, then let a heavy-hitter model like Mistral Large 2 or Command R+ acting as a "Ghostwriter" to extract the facts. I am sure it will be a weird feeling, having a machine summarize my childhood into a series of JSON objects, but who knows, it might be therapeutic as well.

The Hardware: The Dell GB10

You don’t run a ghost on a MacBook Neo. You need serious hardware. My research points to a Dell Pro Max GB10 with 128GB of unified memory. It’s a mini-supercomputer. It’s enough room to run a 100B+ parameter model locally without it choking.

The goal isn't just a chatbot. It’s a persistent agent. It runs a "heartbeat" every 30 minutes, checking its own logs, reflecting on its conversations, and updating its own "long-term memory" via Mem0. It’s a loop that never ends as long as the power is on.

The Ethical Kill Switch

The biggest hurdle isn’t technical; it’s the "Uncanny Valley." There’s a risk the twin becomes a caricature—all the cynicism with none of the humanity.

If I built this, it would need a hard-coded "Digital Directive." A kill switch. A protocol for what happens to the weights when I’m no longer around to calibrate them. Does it become a "Generative Ghost" for my heirs? Or do I have a script that wipes the NVMe drives the moment my heart stops beating?

Why Bother?

In the end, it’s just a mirror. A cognitive scaffold. It’s a way to see your own patterns reflected back at you through a 4-bit quantized lens.

Will I build it? Probably not. I’ve got enough real-world bugs to fix without debugging my own personality. But the files are there. The Docker containers are ready to pull. The ghost is just a docker-compose up -d away.

Legit Creative Uses for AI

 

I have been thinking: is it really such a terrible thing to use AI to generate art for my own personal projects? Mostly I need maps for adventures I am writing. I am not selling it, and I am not going to pay someone else money for a map I am going to use for just a couple of sessions. Is it a terrible crime against artists for me to ask Gemini to generate a Gygax-style old school graph paper dungeon map with 25 rooms? The crux of this is, what are my alternatives? Drawing by hand myself? If I wanted to do that I wouldn't be writing this. Paying someone $50 to do it? Again, if I had cash lying around to do that, I would not be here. Go to Pinterest and download something from there? That is worse than asking an AI to do it. Who exactly is losing money here?

Don't get me wrong here, if I am designing a module I am planning to sell, and I want good-looking fantasy art for it, I should pay someone to do the art for me. First and foremost, I'd get exactly what I want; second, it would look a lot better; and third, it would generate goodwill among my customers. But that is not what I am talking about here. There should be boundaries.

I am also inclined to extend this to character art. Again, I am not talking about publishing for profit here. I am talking about character art for my own player character, a simple portrait, not a series of action scenes. Do I really need to pay someone $50 for this? Where would I even find someone to do this for me, and how likely are they to want to make changes if it does not turn out the way I want it? I would be way more likely to just go to Pinterest than go through any kind of process like that.

So where does this leave me exactly? I don't think anyone is losing any money in these situations. Sure, there might be some copyright issues with the way AI models are trained, but that is not really for me to resolve, that is for lawyers to deal with. In the meantime, I refuse to be judgmental of people who just want to play D&D and need a decent map to play on.

RE: Don't buy a new computer this year.

 

In January I wrote a post saying not to buy a new computer this year. I pretty much broke my own advice immediately. About a month ago I came across a good deal on an M4 Mac Mini and decided to give it a try, because why the hell not. I had not intended to replace my old war machine with this, but after a couple of days it was so good I tried it as my main system and I never really went back. Two weeks later, the MacBook Neo was released and I thought, at $599 how bad can it be. Well, it put my 8-year-old i7 to shame and again, I didn't go back. Even with 8GB of RAM it runs beautifully. Finally, just a week after that, I broke down and replaced my 2018 Samsung Android with an iPhone 17e. Inside of a month I had changed my entire computing workflow.

So what the fuck was going on in my head? Well first off, I did have a windfall of cash, so it's not like I used my rent money for this or anything, I had the extra cash. Second, what I've been doing with my computers has significantly changed over the last 6 months. I find I am doing less and less technical things like coding and project design, moving toward more creative things like writing and building content for my D&D games. Even experiments with local AI have fallen by the wayside as I simply embraced a paid Claude account. Third, as I get older, I find my patience for screwing with computers is wearing thin. I just want these things to work from first boot until I'm tossing them in the bin 10 years later.

After almost a month in the Mac ecosystem, what do I think? Overall, I am frankly surprised it took me this long to make the transition. I really like how these three devices interact with each other and it is basically transparent to me. When I get a call on my phone, I get a notification on the Mini and the Neo, I can even answer the call from those machines, the same is true of text messages. Making Linux do this would have been a 6-month journey and I probably would have never been satisfied with the results, and god only knows if this is even possible on a Windows system.

M4 Mac Mini: This is a solid machine, it is perfect for day to day activities and has handled everything I have thrown at it without missing a beat. Its only real shortcoming was the storage — 256GB is just not enough. However, this was easily fixed with a $100 Thunderbolt dock that provided an NVMe slot. The dock also added a 2nd HDMI connector for a second monitor. Only having one was not a deal breaker, but having a 2nd is really nice. 16GB of RAM is more than enough for 99% of the world and it is more than enough for me. Honestly, this kind of made me feel silly about having 64GB of RAM on my old war machine.

MacBook Neo: I am going to be straight up honest here, this is the best laptop I have ever owned. There is a lot of criticism online about its 8GB of RAM and iPhone CPU, but you know what? Firefox with 12 tabs open, Obsidian, Thunderbird, VS Code and music playing, I felt no slowdown whatsoever. Each application worked flawlessly, switching between them was smooth, even when I was dipping into swap space on the SSD. I can tell you from personal experience, this workload would have brought an 8GB i7 laptop to its knees. While 256GB of storage is not enough for a desktop machine, for a laptop it is fine, this is why we have cloud storage. This is a well designed machine that will serve me well for a good long time.

iPhone 17e: This is the device I have the least to say about. Not because it is bad, it is fine. But because my phone is really the least important computing device I use. I'd say 90% of my use case is using it as an actual phone, text messaging and Facebook Messenger, that is pretty much it. Sure, I do light web browsing, but if I am going to do anything more complicated than a quick Google search, I am going to pull out my laptop. My old Samsung phone was fine and this iPhone is fine.

So do I regret breaking my own advice? Not even a little bit. What I didn't account for in that original post is that sometimes the timing is just right. The hardware was ready, my needs had shifted, and the money was there. If any one of those three things had been different, I'd probably still be wrestling with Linux driver updates and wondering why my Bluetooth headset only works on alternating Tuesdays. The Apple ecosystem is not perfect and it is definitely not cheap, but for where I am right now in life, it is exactly what I needed. Sometimes you just have to eat your words and move on.

AI and Piracy

 So I have spent the last couple of months playing with AI. I have gotten paid subscriptions to both ChatGPT and Claude. I have also been experimenting with Local AI and I have come to a couple of conclusions. 

Local LLM's pretty much suck, unless you are prepared to spend a lot of money. On the low price range, the models you can use are only good for task specific things like object identification in videos or audio to text conversion, stuff like that. General purpose LLM's at the low end are dumb and never meet expectations. If you are looking for truly useful general purpose LLM's with a usable token/second speeds, you need both larger models, 70 billion parameter or larger and the hardware to run them. A Raspberry Pi 5 with a AI Hat, a Jetson Nano or a graphics card with 8 GB or less of VRAM is not going to cut it. There is no cheap way to implement local LLM's as anything more than a toy. If you are looking to get into locally run AI, your starting budget needs to be in the $2000 range, either for a really good graphics card or a purpose built system.

What this means is, in the cost analysis, you can get 9-10 years of a ChatGPT or Claude subscription at $20 per month for the cost of the hardware you will need to build a reasonably useful local LLM. I understand that ChatGPT and Claude are data mining you, but honestly, everyone is and if you are giving them intimate details of your life, you are kind of getting what you deserve. So unless you have a big budget and a really good use case, just choose a subscription and run with it.

During these months of playing with local LLM's, I did purchase a Jetson Nano, it cost me about $250 and it was absolutely not worth the money for what it was intended to be used for unless all you want to do is setup a AI security cameras. What I did find it useful for though was a video streaming device. The 8 GB of shared RAM/VRAM is more than enough to run a handful of Docker containers and since it does have a GPU, it can do on the fly transcoding of video codec formats.

Before I continue, let me just say, I am not an advocate of the piracy of copyrighted media. I believe that media producers should get paid for their work, and I am very willing to pay for a good product provided at a reasonable price, I don't even mind ads and commercials. My problem is, streaming services are at a tipping point, where the cost of those services are getting to the point where they are no longer worth it. If I am paying money, I should not be seeing commercial, or if I am then they should be at the very beginning and MAYBE one in the middle. If I am seeing 4-5 minutes of commercials every 10-15 minutes of the show, that is too much for a service I am already paying for.

 What I did was, using the modified Ubuntu image that came with the Nano, I installed what is called an aar stack. For those not in the know, an *arr stack is a suite of automated media management software—commonly Radarr, Sonarr, Lidarr, and Prowlarr, used to automatically download, organize, and manage movies, TV shows, and music. Often deployed via Docker, this stack connects with download clients (like qBittorrent) and media servers (like Plex/Jellyfin) to automate the entire media acquisition process. 

The Docker compose file I used and rough instructions for building it, can be found here;

https://github.com/cjstoddard/Jetson-Nano/tree/main/aar-stack 

Now, I know what you are thinking, this is just piracy, which I sort of railed against earlier. I am not going to try and justify this, because here is the thing, there are plenty of legit uses for this setup. There is a lot of public domain content out there, there is plenty of free copyrighted content out there. In fact, there is enough free content out there that I cannot possibly watch it all in my lifetime, so fuck off about piracy.

So, to sum all this up, the Jetson Nano is nearly useless as an AI client, but is pretty great as a media server. If you are going to buy one for AI, don't, instead put that money towards a graphics card with 16+ GB of VRAM, in the long run, you will be much happier.

Don't buy a new computer this year.

 

Do not under any circumstances, buy a new computer this year. RAM and video cards are ridiculously over priced and SSD prices are following the trend. No new computer you buy this year is going to be worth the cost. Even if your computer is 10 years old, hang on to it until at least 2027, it is probably good enough. Even if your system breaks, repair it if possible and if not, buy something used to get you by, old Dell's and HP's are cheap and plentiful.

There is no new software coming out this year that is a must have. All the triple A games that will come out this year will still be available next year and they will cost you less and Microsoft certainly isn't putting out anything this year that your life will depend on.

Happy New Year!

 2025 can fuck right off and die. See you all on the other side.