A Private LLM for the people

If you have been following the dramatic goings-on in the AI world recently since the leaking of Anthropic’s entire private toolchain repo, you will have noticed a few big names dropping some “Sit up and take notice” type of tools, namely xAI’s Grok Build and Google’s Antigravity IDE both releasing in May.

A contender entering the arena you may have missed is former world’s most subscribed YouTuber, Pewdiepie, releasing his Odysseus AI Self-Hosting LLM harness as a fully open source project, and inviting contributors to submit improvements.

What Pewdiepie has done is fairly unprecedented, not only in the world of YouTube personalities, but in the cutthroat, venture-capital funded world of AI and LLMs, what he has done is fairly unthinkable. He has released his own, personal, first-party set of agentic tools, that he has built painstakingly over the past couple of years, for free, into the public domain and invited contributors to get on board. Pretty insane stuff.

So what is Odyssey, what do we get if we install it, and why would you build a dedicated server just to run it on?

Well, ultimately it’s an LLM Agent Harness, much like a Cursor or a Grok Build or a Claude Code (where most of the inspiration is clearly drawn from) that lets you use LLMs as agents, carrying out tasks in the real world, for their owners. Unlike those other tools however, this doesn’t have a team of developers or a billion dollar company behind it. Pewdiepie built these tools, by himself, to help him manage his own fleet of GPUs, test out different models and check how they run on his hardware, automatically configure local LLM hosting and model optimisation and so on. All pretty geeky, “AI for the sake of AI” type stuff which is common in the hobbyist space.

However, he’s also done some really, really interesting stuff as well, stuff that you can only really do if you fully trust what you’ve built, and have built it yourself. Like an entire email client that can check and read your emails for important messages, or a “deep research” tool that runs web searches in its own locally hosted search engine, so you never leave a cookie trail and don’t have to read a hundred websites looking for the best deal in a product or the best source of information, it finds it for you. It also has a “Memories” feature that reads your chat history and automatically picks up on important facts it should remember about you. These are features I want, like right now. It’s frankly astonishing, and shows what one determined and technically-skilled person can do with an LLM assist.

If you haven’t already, please see his excellent YouTube video for the full details. Suffice to say it looks extremely impressive, and I want one, like yesterday.

I’m all in. I was already thinking about building something similar myself for locally hosted models anyway, and now I don’t have to. Odysseus is already beyond what I was even imagining.

The problem:

Local LLMs need a lot of computer to run. My gaming PC won’t cut it, I’m going to need to build a new machine.

Pewdiepie has spent over $20,000 on his, and it is a monster capable of running enormous models.

I can’t afford that. I can’t find a Mac Mini for sale with over 24GB of RAM anywhere in the world either, so my options are very limited.

The solution:

I have set myself a budget of £2,000, I will aim to buy used parts wherever I can and it makes sense, and see what I can build with that budget.

I’ve decided on 4 graphics cards because that means I can use a normal consumer case and motherboard, and I’m limited to about 64GB total VRAM by my budget, since we’ve all decided to go temporarily mad, and prices of everything are through the roof.

My 1000W PSU is the first part to arrive. I could have gone higher, I could have gone fancier, but this one was on offer at £100 with free shipping so how could I refuse. I bought this new, since the PSU is not a part I want to buy used or cheap out on.

Spent: £100

Budget Remaining: £1,900

I already have regrets. I could have just blown my budget straight away and bought a 128GB DGX spark for £3,200 in the time it’s taken me to write this blog post, let alone the time it’s going to take me to put this box together. It would arrive tomorrow and I’d have twice as much space for running AI models in, straight off the bat. But my wife would probably kill me, and somehow this seems like cheating.

It’s very easy to blow a lot of cash on this hobby without really trying. I need to remind myself that this project is going to look amazing on my CV and the profile I build on it will help me plan my entire life… So it’s an investment, not a ridiculous toy 😵