I have a habit of starting technical projects by hoarding every scrap of documentation I can find before I write a single line of code or flip a single switch. I hate the idea of breaking anything, so I am very careful beforehand. That works fine until the pile gets big enough that finding anything again turns into its own project. Using a PS4 as a spare CPU/GPU has become my white whale, but it is so much easier since I spent a month feeding my research material to NotebookLM, which made my life easier.

Turning a game console into a computer starts with a messy pile of notes

Digging through old exploits and code is easier with NotebookLM

Honestly, the start of any project like this is just messy, and turning a regular PS4 console into a Linux compute cluster was no different. Before I could even get to the point of pulling things together or judging which software to use, I had to start by dumping a giant, messy pile of technical documentation into NotebookLM just to set my baseline, but that’s what I pay for.

The stuff I was working with was scattered all over the place: GitHub repos, deep-dive security blogs, sprawling Reddit threads, and I had to go through and organize all of it by hand. The first big chunk of material I fed in was about the firmware jailbreak side of things.

To figure out how to actually get into the PS4’s proprietary Orbis OS network stack, built on FreeBSD 9, I loaded NotebookLM with a ton of documentation on exploits and more information on jailbreaking.

I also grabbed the raw Python scripts people use to automate this through a router. Since the jailbreak scene is such a patchwork of community work, I also had to put together compatibility charts covering firmwares from 7.00 all the way to 11.00.

These were already too much for me to keep track of on my own. I needed NotebookLM, especially when I realized I had to learn more about Linux bootstrapping. That meant saving a bunch of scattered tutorials on the kexec system call, which is what lets you skip past the PlayStation’s normal boot process and load your own Linux kernel instead.

I fed NotebookLM the raw details on how to set up MS-DOS partitions on USB 3.0 drives, the formatting rules for FAT32 versus EXT4, and exactly where the compiled kernel image (bzImage) and the initramfs archive need to go. A good chunk of this research also went into hardware quirks and just so much more.

It gives you the files and links you need without the noise

Once I had my project’s baseline variables locked in and the raw documentation cleaned up, I started actually putting NotebookLM to work. Since NotebookLM keeps itself focused on your source material, it keeps the structure and meaning of the original documents intact, so it rarely makes things up.

One of the things that impressed me most was how well it cut through all the outdated homebrew forum noise. The PS4 jailbreak scene feels like one of the messiest subjects I’ve looked through. Reddit threads mix old firmware 5.05 exploits from years back with newer 11.00 or 12.52 payload talk. I’ve seen people use it in the same thread.

When I added sources to the notebook that I may have missed, it skipped past the complaints, dead links, and speculation, and just pulled out good links. So if I asked something like “what are the exact formatting requirements and file placement for the Linux USB boot drive,” it gives me what I need without me needing to press Ctrl+F on a bunch of documents and sites.

I especially like the citations because I always want to double-check where the information is from. If I did that by hand, it’d be annoying, and I might lose where I read it exactly.

Other than just asking it questions, I also had it build out full outlines to help make sense of the whole pipeline. I had it pull together themes across sources to generate a workflow, and it put together a clean outline breaking the whole process into stages.

This made me feel like I still wasn’t ready, but at least I knew this before starting. If I had only used my own notes and looked at them myself, I wouldn’t have realized it until it was too late. With NotebookLM, I can see it all come together.

Large projects will quickly show you where the system breaks

Big files fail to index and cause the AI to make stuff up

NotebookLM’s huge context window has been a lifesaver for wrangling all the messy documentation on this project, but once you push it hard, you start running into some real problems with how it handles data.

When you load it up with dense technical manuals, it will just fail to index a document well, or it’ll flat-out tell you a file you uploaded doesn’t exist. Officially, the limit is 500,000 words per source, but in practice, anything creeping up toward 200,000 words can quietly fail to index without any warning.

When retrieval breaks down like that, NotebookLM has a bad habit of papering over the gap instead of just saying it doesn’t have the information. It’ll make up a plausible-sounding answer or mash together info from totally unrelated documents. You end up getting a hallucination dressed up as a confident answer.

There are also some practical limits to how NotebookLM handles sources that make life harder on bigger projects. If you’ve got one notebook for network infrastructure and another for kernel compiling, you can’t ask a question that pulls from both without duplicating files between them, which gets old fast. And if the AI starts hallucinating or skipping over documents mid-query, the only real fix is to manually uncheck every other source so it’s forced to focus on just one file.

That kind of defeats the whole point of having a multi-source research tool in the first place. To actually keep things accurate, you really have to be disciplined about how you prep your data beforehand.

This worked, but not the whole time

NotebookLM earns its keep for the first stretch of a project, when you mostly need quick answers pulled from a stack of documentation you haven’t memorized yet. The citations alone save real time compared to digging through tabs to confirm where a detail came from. Once a project grows past a certain size, the issues start popping up. I like it, but I can see why some people don’t.