If all your time is spent watching output tokens, where do your input tokens come from? Letting an agent rip on full auto is basically doom scrolling. Even worse if you're doom scrolling while the agent runs. We humans love frying our dopamine receptors. This feels great until you realize what you were offloading: the struggle. The part where you fail. Failure is the entire point. You don't make progress in the gym unless you take a set at least close to failure. The muscle only adapts when it's forced to. It is no different for the brain. Cognitive Atrophy It is very hard to admit to yourself that your skills have atrophied. It is even harder to admit this to other people. I will admit that over the past several months my brain has gotten smoother (and I wasn't even on Twitter much!). Recently, I had written an abstraction for my diff viewer (diffy), an element system with a macro that lets agents write html-like code in rust for native ui (they reason better with this). But it wasn't adopted everywhere in the repo yet, so when I asked for a new feature, the model decided to hand paint it straight to the viewport instead. Every behavior the element system gives you for free was just... missing. Text wasn't selectable. Hover highlights wouldn't go away. And since I wasn't looking closely, it iterated on the slop and produced more slop, more bugs. I just kept saying continue. I lost a whole day untangling it, and the funny part is that once I actually looked at what it had built, every bug was the same bug. When you hit a roadblock and your immediate reaction is to reach for something else (previously, this used to be other people, but now it is a language model) you are essentially skipping the part where you actually learn to solve the problem. It is funny how one of the best "learning tools" has turned out to be the number one cause (anecdotal. sue me) of the lack of learning! It's been a few months since I started writing this, and things have gotten more dire. Several major software services barely work now, grown engineers I once respected are writing somber posts about missing a language model that was banned for a while. Mourning. For model weights. It's all so dystopian. It didn't work, but boy was it beautiful. As the agents get better, one is basically expected to produce code at an alarming rate. The timeline to get something done is compressed but the time it takes to come up with solutions to hard problems has not. There are usually a few good abstractions one can come up with that balance the upsides and tradeoffs for most software problems. However it is currently trivial to turn your brain off and let the slop flow. The code will be complex. It might look like it all works, but something always breaks. And the solution to that? More slop. Software quality is collapsing as a result, and the societal expectation that engineers understand what they ship is disappearing. You never understood the code in the first place. So when you need to change it, you're asking the same stateless clanker to modify code it has no memory of writing. All output tokens and zero thinking tokens. A lower barrier of entry to write software doesn't imply the standards for good software must be lowered. The culture of doing things because they said you couldn't. The growing trend is to do things because you now can (supposedly), but we used to try and do things because we could not out of sheer stubbornness. Carmack and gang shipped QuakeWorld with client-side prediction over dial-up when the conventional wisdom was that twitch shooters over the internet were unplayable. This only happened because Quake's original netcode was laggy and everyone hated it. (They fixed it in a month.) George Dantzig arrived late to class, mistook two "unsolvable" statistics problems for homework, and solved them. Nobody told him they were impossible, so he just did the work. Andrew Wiles spent seven years alone in his attic working on Fermat's Last Theorem, a problem mathematicians had given up on for 350 years. He announced the proof, a reviewer found a hole in it, and he spent another year fixing that too. Notice that all three of them became who they are because of the struggle, not despite it. The people benefiting most from generative tools today, say Terence Tao or Mitchell Hashimoto, already put in the time, so when they offload work they're just skipping the typing. When people like you and me (if this is not you, then I apologize) offload, we skip the grind itself. With language models, easy tasks got easier, hard tasks stayed hard. The hard part was never the task itself. @codex how do I fix this I don't know, I am figuring this out as I go. The amount of time I have spent actually programming has been dropping month over month this year. I used to have a coding stats section on my website that would track hours I spent writing code split by language, recently I had updated it to this: and it made me quite sad. I do think that sometimes all you need is to realize that the thing you are doing is actually detrimental to your growth. Consistency matters more than one would assume. If you consistently take some time away from these tools and actually use your brain, that alone is already significantly better than offloading your thoughts. Solve the problems yourself. Or at least try, fail, and spend time thinking. There is seemingly no "learning" phase anymore. You are expected to just know things. Learning is fun, don't let anyone take this away from you. I've written about this before. It is probably going to be slow, learning takes time and effort. You will feel stupid (I feel stupid). This is a good feeling, because there exists a world where you are no longer stupid and the path towards it is learning. Books still exist! Libraries are still open, notebooks waiting to be written in. Read more. Write more. But If you really do care about improving yourself, be honest and use these models for what they are, highly efficient filters of zettabytes of data (the internet is estimated to be 175-240 zettabytes (10^{21} bytes)). It was extremely difficult to identify what one needed to read to learn niche topics even like 2 years ago. I remember asking a good friend of mine to recommend material to dive deep into learning about SIMD, and honestly there wasn't much stuff to read except the Intel Intrinsics Guide. And if you've ever taken a look at that, it is quite cancerous for a first-time reader. Language models are super useful here because you can point them at material and you can ask questions that pertain to the thing you care about and it will simply just tell you the correct things. Conclusion One good thing in this age of slop is to consume knowledge at an unbelievable pace. I don't necessarily mean using only model output for learning (I don't trust them to learn any topic more than a shallow amount), but rather using them to help sift through the plethora of information available out there and identifying the right things to read. Human slop exists too and using a language model to supplement your learning might help keep you sane (ironically). I like using these models to write code that I tell it to write (outside of work I enjoy doing it myself entirely), and I am largely disinterested in asking it what I should write. There are exceptions of course, because not everyone is working on scaling software services which has largely been solved (but slowly being forgotten), but that would be for you to decide. The best model you have access to (and it has solved continual learning) is, and always has been, the one inside your skull. It's time to scale up its input tokens.

This is a general devlog covering what I’ve been up to over the past few months! Haskell Hakyll is a Haskell library for generating static websites (yes, you do not, in fact, need Next.js for your personal site). The previous iteration of my website was written with Go + templ, and it was not static. I wrote everything myself, and it was, quite frankly, a horror to maintain. The templ LSP actually hinders more than it helps, and since I did not statically generate the content, it was kind of slow for a personal website. I did not like it. I added some very minor CSS fixes and a theme toggle button for my friend’s new website ludwigabap.com, and I saw the light that is Hakyll. :kneel: So I nerd-sniped myself into rewriting my personal website in Hakyll so that it generates beautiful HTML at build time (it uses Tailwind, too). All the dynamic bits of the website, like my silly stats and coding time, are served from a single Rust Axum server binary on my Hetzner VM that serves other stuff too (Memegrep’s server, among other things). Nix-pill This segues into my next nerd-snipe, or rather, my favorite new pill: the Nix pill. My entire website is built with a single Nix flake, including the Hakyll build step, the Cargo build step, and generating the resulting Docker image. Although, to be honest, the learning curve is very steep and I simply do not have it in me to master this language anytime soon, I like the principles, and it seems to me the least bad build system out there. I do most of my work on a MacBook running a NixOS VM as well. (seatedro/dotnix on GitHub.) It’s so much nicer to have a real dev workflow versus whatever the fuck macOS is. Sorry, but Homebrew is not good software. Ember The reason I started programming was that I was playing a game called Midtown Madness as a wee lad and decided that one day I would build a cool game like that. So, after 15 years, I’ve decided to embark on a small adventure to build a cool physics simulation (and eventually a voxel) engine. It’s written entirely in Zig because I enjoy writing Zig and did not want to touch C++ (although there were times I thought about committing this sin). When Ember gets far enough, I might write separate devlogs, but with what’s written so far, there isn’t enough to justify one. I was knee-deep in the mines because I wanted extremely specific features like multi-viewport docking with ImGui, and the off-the-shelf Zig libraries did not want to ship that for some reason. So there were many adventures in getting this shit to compile neatly. Ember now uses SDL3 for windowing and has an abstract rendering API (rudimentary, but the backends are in place) with SDLRenderer3, OpenGL, and WGPU-native backends that can be neatly switched at comptime. I’m using a bunch of cool resources to learn more about how to do this because I’ve never really written any game-dev related code before. I just want to build some cool simulations, and I will get there no matter what. github / seatedro / ember Thanatos After 5 years, I have finally built a new personal computer. I will be running NixOS as my main operating system with Hyprland and Wayland. I will also have Windows installed on a separate drive mainly for video games, though I expect to use Linux more often. It’s so nice not having to worry about my 256 GB SSD getting filled up on my MacBook Air. :> Exa Also, I joined Exa last month to work on the back-end team, and it has been so fun! I am soaking up as much knowledge from my significantly smarter peers at a good pace. As per custom, I revisited @ludwigABAP’s post: On becoming competitive when joining a new company. (PS: it’s on a new website.) Fin Back to the code cave I go—going to extract the polymorphic.

Look everything up Pretty much the best way to learn that I have found is to refrain from suprressing your innate curosity and let it go wild. Most of the time, you will encounter a term or concept that you do not know, instead of glancing at it briefly, go all in. Google it, read the wiki page, found someone’s blog post? Read it. Watch that youtube video (I only do this if it’s not 3 hours like your average Sphaerophoria stream). Dive head first into rabbit holes. I’ll outline an average day where I have two types of goals: - Concrete goals: Like finish implementing X feature in Y project - Loose goals: Learn about X or Y. I will include rough timestamps but they are pretty much meaningless because productivity levels vary immensely throughout the day. 9:00 AM Decided to learn about document parsing. Links: EPUB specification, I found out that EPUBs are just an archive with HTML, and nearly had a crisis. (Now I know.) Immediate questions: so I need to parse the EPUB, extract metadata, fetch the XML and parse that too. Eventually I would have to use a WebView to render the HTML/CSS on the screen to render the book. (Unfortunate.) Found out that in order to read the PDF spec, you need to pay like 350 swiss francs! Spend time perusing SwiftUI docs, and asking grok about how I can render things on the screen with Swift. Fiddle with XCode, marvel at how Swift gets pretty much everything right but uses func ... for declaring functions. It has Result<T, E> though, so forgiven. 11:00 AM Began reading Computer Systems: A Programmer’s Perspective (CSAPP) Studied numeric representations, created Anki cards for hexadecimal conversion 12:00 PM Rabbit holed into learning so much about UTF-8 Bookmarked to learn more about UTF-8 and UTF-16 and writing a parser for it later (ILY @zack_overflow) 1:00 PM Taking a break, lunch + watching a movie. 2:30 PM Read Jon Olick’s single file resize implementation in C++ as reference (ILY @gizmobly) to roll my own resize library for use in glyph Rabbit holed into learning more about Sinc filters and the Lanczos Kernel. 3:30 PM Shifted to working on memegrep(v2). This is where I already know the goals I want to accomplish before the day ends. I implemented a pub/sub flow to help with scale when users upload their private memes. Rough sketch in mind: user uploads meme(s) → server receives req → insert skeleton into db → queue upload → return 201 to user immediately → worker picks up task Spent the next 6 hours coding without even realizing 6 hours had passed (bliss) Ended up with multi-file upload, search, deployed a CLIP model, added all the scaffolding needed in the UI for this. 9:30 PM I posted something about pointers on twitter and ended up reading some history about the nomenclature just for fun. Links: pointers, handles are the better pointers Also saw something about reference counting being used in the libvips library earlier in the day, so decided to check out their implementation since i’ve only used it in rust quite often. reference counting, Rc and Arc 11:00 PM Was in bed and saw a post linking an amazing article by Valve on Source Engine Networking, so ended up being a nice and light read. At the end of the day I ended up with more questions, but I definitely had more answers than when I started! Here’s a DAG of my exploration for fun:

two-weeks I built a website (twoweeksisallyouneed dot com) with just Claude 3.5 Sonnet, zero lines of code written by me. Why I did this So a couple weeks back I had a computer vision midterm and i was allowed a single page of notes. I decided to use Claude to generate a cheat sheet in LaTeX. It was crazy lol, I was able to cook up something usable in 15 mins. pic.twitter.com/0tvIsdv34d — ro/nin (@seatedro) October 24, 2024 Someone had asked me to make a similar cheatsheet for ML, but I thought why not get claude to make an entire website instead? Early days I didn’t really give claude any specific information except what I wanted to build. My plan was as follows: > build out the UI skeleton first > populate with some dummy data > set up a content pipeline > tie up the ui with the data > fix bugs Claude decided to use react (Shocking). With just a couple of chats I was able to get the retro/hacker/matrix style UI down. (It’s going to be hard for frontend engineers to keep up with AI at this rate) I spent some time building and adding secrets/easter eggs to the website which no one has found yet lmaoooo. If/when all the easter eggs are found, I will open source the repository. Good hunting until then bros. Frustrations Things got really annoying, really fast. As soon as I wanted to build some sort of content pipeline, everything went to shit. Claude, no matter how smart of an AI, is not human. It did not think ahead. If I was going to build out this website I would not have started with a barebones react/vite app. I would have probably gone for a full stack framework like sveltekit instead. Generating content for the website was/is a nightmare. LLMs hallucinate, this is known, but did you also know how incredibly frustrating it is to get them to follow instructions? There were multiple instances where Claude (with all the 40% of project context of javascript code) generated a TypeScript interface and proceeded to spit TypeScript code. This project has 0 TypeScript???? The new edit in place nonsense was getting on my nerves. Half the time the output artifact wouldn’t even change, and the other half it would mess up the changes. I have to mention “please use a new artifact” if I wanted any real usable code. I was enthusiastic about building something with just Claude the first few days, then I started getting weary, and then eventually I wanted to take a sabbatical from using AI. When the project context grows, (think like 20% or more), Claude seems to have a hard time using that information. Often times I found myself hitting send, and Claude would spit some nonsense, I would hit stop and paste any relevant code directly and then get some useful code. Context is probably the biggest annoyance I’ve had with LLMs. Random thoughts Perplexity (with claude 3.5) is great (to an extent) because it’s essentially a RAG search so I was able to get somewhat up to date content for the resources and references sections from it. I feel like I was able to ship something of decent quality for sure, but I lost a lot of brain cells during the process. Do not take away from the programmer the only thing he wishes to do, program. Proompting I didn’t particularly do anything unique to get the best out of my prompts/chats. I did do everything in a Project on claude dot ai though, which let me set Project Instructions like so: BE ENTHUSIASTIC. WE ARE GOING TO CHANGE THE WAY THE WORLD LEARNS WITH THIS WEBSITE🚀🚀🚀🚀 LLMs seem to be much more open to doing anything you ask if you gaslight them, so go ahead and do it. Here are some prompts that I used while building the website: PROMPT ------ This is AWESOME: * For the loading screen add some text that says "you can learn anything in two weeks" * The ascii text is also weird, it says erain? is that supposed to mean something? * Where did all my placeholder topics go? * MOAR scrt scan lines * The matrix rain needs to fall vertically and it should be subtle. On mouse move it should get a bit brighter (the char under the mouse) What else do you think you could add? Surprise me One time I asked Claude to add something and it started changing the existing UI????? PROMPT: ------ OKAY, we have a beautiful boot sequence now. let's flesh out the main content screen. First let's fix some bugs: * The animations restart everytime i move my mouse instead of continue organically * What i mean is, when i move my mouse over the windows or around the matrix rain, the matrix rain restarts and so do the typewriter effect on the window title Work on these * [REDACTED] Easter eggs: * Fun Features: * "Power saving mode" that dims everything except what you're reading ....[REDACTED] So many times I ended up editing the prompt and adding lines like this: PLEASE DON'T CHANGE THE EXISTING UI. IT LOOKS GOOD BRO. I’ll just add some random prompts here PROMPT ------ Broski, we need to make the UI responsive and shit. Lot of people are reporting issues with it on mobile and smaller screens PROMPT ------ what sort of content do u think we should add? i was thinking things like formulae, charts (for phd), code blocks (for eng), research papers, youtube video links, blog posts etc. Let's think out loud how we're gonna do this before we proceed PROMPT ------ `pasted_code.jsx` Here is my current dashboard. Here is some sample concept content i have: `pasted_data.json` I need you to render this beautifully in the dashboard. Make any and all changes needed. It needs to look beautiful. And anything else u feel would be good. List out the things you're adding before adding it okay? Concluding thoughts This was a good experiment, a success. If you have a clear vision for what you want your product to be, then AI can help you achieve that vision quite well! However, I don’t think I will be using AI for the foreseeable future. I feel like my learning is stagnating the more i use AI and I want to write my slop code with my own two hands. I might use avante.nvim to quickly write some duplicated code here and there but by god, I miss coding. Actual coding.