Yes, you will hit the limits of this machine. 8GB of RAM and a phone chip will see to that. But the limits you hit on the Neo are resource limits — memory is finite, silicon has a clock speed, processes cost something. You are learning physics. A Chromebook doesn’t teach you that. A Chromebook’s ceiling is made of web browser, and the things you run into are not the edges of computing but the…
Does the model “understand” anything? That’s a philosophical question, but mechanically: no magic is happening. The model is a big math function that maps input tokens to a probability distribution over the next token. During training, the parameters are adjusted to make the correct next token more probable. Whether this constitutes “understanding” is up to you, but the mechanism is fully…
Conversational interfaces are a bit of a meme. Every couple of years a shiny new AI development emerges and people in tech go "This is it! The next computing paradigm is here! We’ll only use natural language going forward!". But then nothing actually changes and we continue using computers the way we always have, until the debate resurfaces a few years later. Link
The psychology of LLMs use worries me deeply. It’s a system that could not be better designed to encourage harmful behaviors if you tried, because it mixes a couple of very potent psychological traps. And it’s further being pushed by a large amount of venture money, and being co-opted by grifters with vested interests. Please try to be vigilant for these risks as you engage with these tools, both…
It’s not even black-and-white, it’s gray-and-gray. Link
My first reaction when I read about the plans of building datacenters in space was "wait, are we sure this is going to work? I don't think it's realistic". The author knows more about the topic than I do (to be honest that's a low bar) but their opinion is the same. I suppose this is just about possible if you really want to do it, but I think I've demonstrated above that it…
I agreed with Marc on all the questions he asked. The bottom line is that error handling in systems isn’t a local property. The right way to handle errors is a global property of the system, and error handling needs to be built into the system from the beginning. Getting this right is hard, and that’s where blast radius reduction techniques like cell-based architectures, independent regions, and…
Machine learning is amazing if: The problem is too hard to write a rule-based system for or the requirements change sufficiently quickly that it isn't worth writing such a thing and, The value of a correct answer is much higher than the cost of an incorrect answer. The second of these is really important. Link
I changed a few things in my homelab since the last post so here we are again with a new post. Hardware The mini PC I use (A Fujitsu Esprimo G5010) is still way over-provisioned for my needs. I found a good deal on a 32GB RAM module so I upgraded it and now have plenty of CPU power and RAM to spare for anything I might want to experiment with. Operating system I have been running NixOS on a…
But here’s the problem: Writing code is easy. Reading it isn’t. These advanced features are like salt: a little bit can enhance the flavor, but too much can ruin the dish. And advanced features have a tendency to overcomplicate things and make readability harder. Software engineering is all about managing complexity, and complexity creeps in when we’re not looking. We should focus on keeping…
I started with every page in my bookmarks. Then I went on to index every link I'd posted on my social media account. Maybe next I'll include the links posted by people I follow, or the highlights from my browser history? Right now I'm hyper-personalizing for myself because you have to start somewhere, but hopefully future steps can build this out into something that could work for…
The app was implemented in two pieces: The kernel , written in C, implemented the database, networking, inter-application communication, various built-in data types, script compiler and evaluator and debugger, and so on The scripts used the kernel and implemented most of the actual app behavior Since it was an app, it had plenty of UI — menus, contextual menus, buttons, larger UI components, and…
I recently made the Space Invader Generator for Creative Coding Amsterdam code challenge. I made it for fun of course... and galactic domination too! You can see how it looks below and in this post I'll show you how it works using an interactive animation. Link The interactive animation does a great job at showing how the generator works step by step.
I write the content on this website for people, not robots. I’m sharing my opinions and experiences so that you might identify with them and learn from them. I’m writing about things I care about because I like sharing and I like teaching. I spend hours writing these posts and AI spends seconds summarising them. I'd much rather people read the whole thing, take it in, digest it and have…
Right now, LLMs are an extremely immature technology. I personally believe they’re not going to get much better than this, but a breakthrough innovation could change that. Either way, it doesn’t matter to me. If the technology is a fad and completely implodes, I couldn’t care less. If it’s not a bubble and LLMs actually turn out to be the new best thing, I can easily adopt them into my own…
By “surfing the internet” I don’t just mean going online. I mean exploring the internet solely by following hyperlinks from page to page, with no clear destination except for that one wonderful, as-yet-unknown website that will amaze and enthrall you when you find it, the one that will seem like it’s been waiting for you your whole life and which you can’t get enough of. Link As soon as I read…
That’s the real danger of shortcuts: they make you feel powerful and productive, but can lull you into a false sense of security. Without the intuition and habits forged by experience - knowing when to be wary, when to back up, when to double-check - you’re at risk of costly mistakes. These lessons aren’t learned by skipping to the answer; they’re earned by wrestling with real problems. Link…
Rust today is the result of significant investments made by forward-looking institutions and the efforts of thousands of individuals who shared a belief in the long-term payoff of what they were building. It may seem like 10 (or 15 or 20) years is already a long time for Rust to have existed, but C++ is 40 years old and FORTRAN is almost 70. Rust might last that long, or longer. Link The article…
Local-First is a neat approach. The article describes its pros and cons in a complete and objective way. I found it interesting to learn who their customers are: While most of RxDB is open source, there are also premium plugins that help sustain RxDB as a long-term project. Because people purchase these plugins, we gains insights into how developers are using local-first features: About half of…
One of the most common retorts I get after showing these examples is some form of the statement “but those tasks are easy! Any computer science undergrad could have learned to do that!” And you know what? That's right. An undergrad could, with a few hours of searching around, have told me how to properly diagnose that CUDA error and which packages I could reinstall. An undergrad could, with a…
The personal site isn't dead. It's just been forgotten in the commercialised, capitalist web of today. We owe it to ourselves to rediscover this lost art. We can still be creators for the sake of creating. We can still post content without someone else making money from it. So, once again my digital call to arms: build your own website. Make it fun. Make it pointless. But most…
A healthy software ecosystem doesn‘t arise by specifying how software should behave and then assuming that implementations will do the right thing. Rather we plan on having Roughtime servers return invalid, bogus answers to a small fraction of requests. These bogus answers would contain the wrong time, but would also be invalid in another way. For example, one of the signatures might be incorrect,…
AI-news are a rollercoaster these days. Some days companies claim that they're going to replace engineers with AI. Then people actually try to use AI tools that promise to replace programmers and this is the result: Working with Devin showed what autonomous AI development aspires to be. The UX is polished - chatting through Slack, watching it work asynchronously, seeing it set up environments…
If you’re in your early thirties or older, you remember the breakneck pace at which computers were improving in the 90s and 2000s. If you brought home a shiny new 400MHz Pentium 2 system in 1998, it was literally outdated and replaced by a faster model with higher clock speed and more RAM/bigger harddrive by the time you had set it up and gotten familiar with it. But that’s not the case…
This story is as a cautionary tale reminding us the value of critical skills. The story had been found – because of a librarian using critical skills (and thereby not giving equal weight to each factor), an archive, and a catalogue/index. Verily: librarians, archivists, cataloguers, and indexers are the Noble Professions. For they organise information in a manner in which humans actually…
#hachyderm costs about $1600/mo to run. this is up somewhat, as we've started to add some infra as part of our resilience plan announced in nov. we currently have about: 55000 users 9700 MAU 3.7M toots yielding: $.03/user/mo $.16/active user/mo $.0004/toot from a raw compute & storage perspective. again, this is based on 100% volunteer work. today, our mods and…
There is a lot to like in this essay that starts describing a very personal experience and continues with reflections that go deep into anthropology and philosophy territory. The attempted simplification of the world into these two clear categories – natural and unnatural – is both false, and a source of endless confusion, conflict, and poorly made decisions. That’s because “nature” and “culture”…
Sharing interesting links with commentary is a low effort, high value way to contribute to internet life at large. This post , in particular the quoted sentence, inspired me to add again a link section to my website. This time each link is going to have its own entry with quotes and a short commentary. I often come across interesting posts, video and articles so I should have plenty of material to…
For a couple of years I've done Advent of Code but that's no longer my thing. This year I came across December Adventure and that resonates more with what I'd like to do: a little progress every day on a project rather than having to solve a couple puzzles every day. Lately I've been working on my own tool to optimise the decision on if/when/how to charge our home…
I self-host many services . Some of them need to be up 24/7, others are used only occasionally. I tend to choose lightweight services so I could keep them running all the time without too much trouble, but if I just did that I would never have written this post. Instead, let's see how we can setup a service to start when it receives a request and to have it shutdown after a period of…
It's been a while since I last wrote about my homelab setup. The hardware is still the same Fujitsu Esprimo G5010 mini PC with an Intel Core i7 10700T and 8GB of RAM. The setup based on Proxmox VE running NixOS LXCs I described in this post stood the test of time and I'm very happy with it. On the software side I made many changes, mostly additions, since last year . Infrastructure I…
Due to family commitments, work, and also a few days not feeling great I had less time than I would have liked to use my old computer. Still the challenge was pretty successful and I'm definitely going to keep this old computer out of retirement and using it more as my "home computer" now that I have it fully set up in a way I like. What I did with that machine I've been able to do…
A while ago I came across the old computer challenge on Mastodon and it immediately resonated with me. So here I am finishing setting up an old computer for personal use for a week (and likely longer than that). Hardware I no longer have my first computer (a Pentium 2 from the late '90) nor my first laptop (a 2005-era Acer, bought when I started university). Both would have been great…
I run a few NixOS LXC containers in Proxmox. This setup works very well and it's pretty resource efficient. I've seen a few people struggle getting everything setup, hopefully this post helps. Generate the CT template Let's start by generating the most minimal CT template which we can use to start the container. I assume you want to keep a long lived LXC and run nixos-rebuild (or…
I self-host many apps and I’ve realised that the ones I like the most all share these features: single binary single process fully configurable programmatically uses SQLite if it needs a database frugal with resources My goal is to have an app that can be deployed anywhere trivially and that requires very low ongoing maintenance. Self-hosted apps typically have a very small number of users (often…
About a month ago I explained what wasn’t working well with my homelab. The root cause is that my scale is as small as possible: a handful of users at most, with everything running on a single machine. Container orchestration tools like Kubernetes or Docker Swarmd are designed for a much bigger scale and were mostly just adding complexity to my setup with very limited practical benefits. The…
I have been running the setup I described in August and while I have been fairly happy with it I think I haven't got everything 100% right: I have a stateful VM that don't really fit well with the rest of my model. I was aware of the issue from the beginning, I decided to leave it as something I to look into it later and I guess "later" has arrived. Let's review all the layers of my…
I own a Growatt solar inverter and I decided it would be nice to collect some of the data it emits. At the moment the goal is to have more charts in Grafana but this also opens up the possibility of integrating this data with weather forecasts and energy prices to optimise usage patterns. By default Growatt inverters send data to their own server and data is accessible via a web ui and mobile…
I've enjoyed playing with my homelab for a few months and now it reached a state that I am comfortable sharing. The setup is working well for me, it’s very stable and flexible. Hardware I run everything on a Fujitsu Esprimo G5010 mini PC. It’s a couple years old and I bought it used. It follows the TinyMiniMicro trend and I'd definitely recommend it. With an Intel Core i7 10700T and 8GB…