NeuraLinux Bringing GenAI to the Linux desktop

Software ages like fine milk. Keeping it from spoiling takes a team of dedicated people and countless man-hours, which is a luxury open-source software often lacks. Meanwhile, Big Tech products brandish the industry's best engineers and colossal neural networks, further widening the gap due to demand for the latter. Nonetheless, we have entered a democratizing era of computing, and I will only stay Team Penguin if they take advantage of the new resources. Now that creativity is the limiting reactant to produce great software, the door is wide open for schmucks like me. So I will bridge this automation gap by first articulating specific usability problems I encounter (in James Mickens-style), and second proposing simple yet cutting-edge solutions that combine underutilized OS features with generative AI models.

This blog presents a series of persuasive arguments to defend those solutions, record the knowledge from my first two years of college, and expand the frontier of OSS. I'm bored of the traditional way humans compute, but I see a clear path to making it seamless. Throughout these articles, you'll find topics from a wide breadth of disciplines that I found really cool. I've poured considerable research into each one to ensure that curiosity is the only prerequisite to comprehension. I hope reading it provides you with fresh perspective on your problems, or recreates an "aha!" moment I had.

If you're interested in training models or developing software, please click this...

Listen, I'm going to keep it straight with you: this website is a help wanted poster. I've narrarated the requirements and design of 17 ideas up to date, so let's run the numbers. Assuming I stop having new ideas and each one needs 100+ hours of development and testing (which is honestly a low ball considering my heuristic is admissible), it will take 1700+ hours of focused work. Assuming I can dedicate on average 1.5 hours/day, I would be done in 1133 days, or over 3 years.

I don't have 3 years, but I do have a proposition. If an idea piques your interest, please implement it yourself. It's a win-win because you get to sharpen your coding knives on a pre-designed insightful project, and I get to play with a new toy at the end of it. For those who want to enter open source development, this is the perfect place to start because there is 0 bureaucracy. If you decide to adopt an idea, then let me know here so I can mark the article and contribute any time I can. To help you decide, I've ternary encoded some metadata for each idea as a number (\(d_2d_1d_0\)) = (progress, time commitment, AI stack position). Pick any 1\(d_1d_0\) according to your availability and interests.

0 1 2
\(d_2\) Article unpublished Waiting for adoption Implementation underway/complete
\(d_1\) Straightforward (50hrs) Challenging (120hrs) Good luck... (250+hrs)
\(d_0\) Traditional, deterministic software Built upon foundational models Requires model training or data collection

Note: This site is under construction. The content that you see is incomplete and will change at any time.


Regulators can never declaw GenAI due to Moore’s law

“AI” has been the Wall Street Journal’s darling buzzword of the past year. If there’s any one topic their technology section adores, it’s how FAANG/MAMAA/Magnificent 7 will break the world selling their “AIs” to every business on its surface, or how Elon Musk claims those “AIs” will lead to our extinction. I can practically feel the business people salivating every time they hear of a newly profitable technology. But I realize it’s time to revisit whether we should actually ready our pitchforks and torches. There’s a growing divide between people who understand the theory behind the technique, and people who...

Identity digital dirt by shifting where your news gets cleaned

Plus, The engineer's guide to fallibilism and pragmatism (022)
In the digital age, the proliferation of generative AI has posed unique challenges to the integrity of information. News publishers have observed the growing disquiet concerning the authenticity and reliability of digital content. I believe there is an urgent need for education and civic engagement to mitigate these challenges. Here,...

A labor-market approach to model inference scheduling in soft-real-time systems

Plus, The beauty of algorithmic transferability (112)
Now that I’ve elaborated how society is screwed due to generative AI, here’s a way to speed up the process via optimal resource utilization! I will show a reduction from model scaling/selection policy to the well-known labor economics problem, for which humanity has fantastic algorithms. But first, I have some...

The desktop is missing crucial quality-of-life features

It was after I asked myself “why does setting the second bit of the keyboard controller’s port to high enable 21-bit physical addressability in x86 real mode?” that I realized my academic education had sunk too deep into the details. I know we need the nitty-gritty architects to construct a solid foundation of knowledge and best practices, but more so we need fresh dreamers to drive innovation beyond the conventional boundaries and find better ways to solve old problems. It’s the reason Nintendo banked on Miyamoto in 1977, a complete outsider to the gaming industry, and now the world has...

Enable the perfect virtual assistant by capturing your life in real-time

Plus, Designing an effective dataset with OS principles (112)
As an iPhone user who talks often about wanting to switch to Android, Apple’s new Journal app has both amazed and terrified me. What a seamless integration of my GPS location, photos, and messages that makes it feels like my phone understands me. It made me remember forgotten but cherished...

Start working where you left off through smart workspaces

Plus, Why the desktop metaphor would make a terrible computer architecture (011)
The battle between GUIs and CLIs was over before it started; people just had bigger things to worry about than scripting their startup procedure and ricing their shell prompt. The desktop metaphor mirrored our ape-brain spacial reasoning well (including its drawbacks)… But fast forward a decade, and desktop apps lost...

Save an hour of battery by dimming the screen when you look away

Plus, Optimizing local vision through compression and sleep (011)
Intel, AMD, and recently Apple have optimized CPU microarchitectures to the point that they do everything at the cost of nothing. Desktop CPUs are the pickleball players who effortlessly play with their left hand while talking a call in their other, while I/O devices are the people struggling on the...

Organize your file system into semantic subfolders and archives

Plus, How ChatGPT makes for the best digital housekeeper (001)
Your Downloads folder is a mess, and you don’t have time to do anything about it. Despite mindblowing advancements in task automation, filesystem organization remains a manual chore—an outdated software convention that hasn’t kept pace with modern needs. Imagine the Downloads folder as the inbox of your file system, a...

Stream any digital media for free without the buffering and shady websites

Plus, The paradox of the subscription model (011)
There’s something universally satisfying about a bargain, and something immensely disappointing about a ripoff. If spending your money is life’s greatest optimization game, then finding techniques that maximize your utility is thrilling! Did you know that by cycling new accounts on a restaurant’s app, you’ll never again pay for guacamole...

System package managers abstract too much from enthusiasts

We digital humans are so detatched from the software we use that it reminds me of the meat industry. Who cares about the chicken who lived for 6 weeks when met with a plate of buffalo wings. Similarly, who cares about the hours of non-profit programming and hilariously pointless culture wars when you’re talking to ChatGPT. But like the vegan will tell you, you should care.

Curate your own Linux distro through autonomous packaging

Plus, Why the best way to "learn Linux" is to make it from scratch (021)
Description coming soon…

The meta-distro reproducible package manager

Plus, Modifying dependency-hell to avoid NP-completeness (020)
Use Linux because of its strong ecosystem of package managers: a uniform interface to the vast universe of software. I remember the seams of downloading installers and zip files manually back when I used Widnows. I would have to navigate every company’s (sluggish) website and click this and that and...

A Linux package set without GNU, X11, nor systemd

Plus, Meet the modular next-generation of Linux software (000)
Description coming soon…

Programmers neglect the non-programming stuff

Aspiring programmers may perceive programming like a playing the guitar to serenade a love at first sight: think of a fun song, write it, and play it. But in reality, it’s a concert. You have to collaborate with other musicians (version control), secure a venue and setting (devops), convince people to attend (marketing), inform attendees to get there (documentation), and hold dress rehearsals (testing). Otherwise, you’re just a bum playing your trombone in the streets with no one listening, save for the disgruntled apartment residents above you that are the source of the tomatoes.

Generate personalized documentation for users and developers of OSS

Plus, How distributing RAG systems can democratize knowledge (011)
Learning new software is hard. Even though we’re in a time where writing software is easier than ever due to ChatGPT and the likes, finding information that’s right for your comfort level and exposure to the project hasn’t become much easier. There’s an abundance of online resources to learn how...

Arrange any song for any instrumentation

Plus, Pioneering a new modality for AI (022)
Music and programming may seem like distant cousins, but they share common ground in their reliance on patterns, logic, and creativity. Yet, while programmers often delve into the intricacies of code, music composition remains a largely unexplored realm, confined to those with extensive training or a natural inclination. It’s a...

Unit test LLM prompts through embeddings

Plus, Coping with nondeterminism in language models (101)
LLMs are non-deterministic by nature, which means our prompts can’t be traditionally unit-tested. Instead, assert that the similarity between the model’s output and your reference output is above your defined threshold. We first calculate the vector representation of both pieces of data, and then take either the Euclidian distance, cosine,...

Linux shells have terrible UX

The UNIX command-line was clearly designed by the hardened programmers fabled in The Night Watch, who worked with rudimentary tools like pointy sticks and teletypewriters. They laid the groundwork of their OS with their bare hands and wrote textbooks when things went wrong. However, today’s command-line users get to wear their non-functional scarfs and sip tea while executing ./run.sh because of the hard work of those systems programmers before us. But boy did they leave us some unresolved baggage while designing these systems.

Toggle the command-line between English and shell

Plus, The history of natural language interfaces (201)
Natural language is the holy grail of HCI, but its principle challenge is precisely describing instructions of high granularity. Think of how many times you’ve failed to get Alexa or Google Home to do a slightly complicated task. This is because the number of possible interpretations is incredibly large. Natural...

Synchronize development environments with aliases

Plus, Solving "it works on my machine" with Occam's razor (201)
In the world of software development, the phrase “it works on my machine” can spell the beginning of a time-consuming, frustrating debugging session. This often stems from inconsistencies between development environments—differences in command setups, dependency versions, or configuration settings. A fundamental aspect of addressing this issue is ensuring that all...

Autoconnect to eduroam during Linux bootstrapping

Plus, Understanding the Linux networking stack (100)
A common chicken-and-egg situation is connecting to institute WiFi during a minimal Linux installation. To download the tools you’re familiar with using, you need to have internet access. Secure networks, specifically 8021.X compliant ones, are notoriously difficult to configure from CLI because CLI authenticators are far behind their GUI counterparts...