Ask HN: What Are You Working On? (March 2026)

Tiny LLM Training with Rust/Candle: A DIY AI Journey
In the ever-evolving world of artificial intelligence, one developer is taking a unique approach to machine learning by training a small language model using Rust and Candle. This isn’t your typical AI project – it’s a labor of love, a playground for experimentation, and a testament to the DIY spirit that drives so much innovation in tech.

The project, detailed in a recent GitHub repository, showcases the creator’s journey in developing a tiny language model from scratch. Using Rust and the Candle library, the developer has been constantly tweaking and refining their model, keeping meticulous track of results in a spreadsheet. It’s a process that highlights the iterative nature of AI development, where small adjustments can lead to significant improvements over time.

What sets this project apart is the developer’s approach to data generation. Instead of relying on vast datasets of human-generated text, they’ve opted to craft their own corpus using LLMs. While this might not be considered best practice in traditional AI circles, the creator is unapologetic about their methods. “I can sleep properly despite doing that,” they quip, emphasizing the project’s nature as a fun, experimental endeavor rather than a serious commercial undertaking.

The results so far are intriguing, if not groundbreaking. The model has shown the ability to answer simple questions with up to 75% accuracy, demonstrating a basic understanding of context and relationships within text. For example, when asked about the color of a car mentioned in a sentence, the model correctly identifies it. However, it’s not without its quirks – sometimes it gets tripped up on more complex queries, offering amusing but incorrect responses.

Currently, the model’s corpus stands at around 17MB, comprising stories, generated encyclopedia content, JSON examples, and more. The developer is particularly excited about the JSON content, which is a recent addition to the project. While the model is still struggling with this format, there’s optimism about its potential improvement in the coming weeks.

One of the most charming outputs from the model so far has been its response to the question “What is life?” The simple, profound answer: “It is.” This moment of apparent insight, whether a fluke or a glimpse of something deeper, encapsulates the wonder and unpredictability of working with AI.

This project is more than just a technical exercise – it’s a window into the world of independent AI development. It showcases how accessible tools and a curious mindset can lead to meaningful experimentation in machine learning. While the results may not rival those of large-scale AI companies, the learning experience and the joy of creation are evident in every line of code.

As the developer continues to refine their model, adding to the corpus and improving accuracy, this tiny LLM project stands as a testament to the power of curiosity and the endless possibilities in the field of artificial intelligence. It’s a reminder that sometimes, the most interesting innovations come not from corporate labs, but from individuals tinkering away in their spare time, driven by passion and the simple joy of creation.

Tags: #AI #MachineLearning #Rust #DIY #LLM #ArtificialIntelligence #Coding #Technology #Innovation #Programming #DataScience #TechNews #AIExperiment #LanguageModel #OpenSource #TechCommunity #AIProjects #MachineLearningJourney #AIEnthusiast #TechDIY

Viral Sentences:
– “Training a tiny LLM for fun using Rust/Candle”
– “I can sleep properly despite doing that”
– “It is.”
– “Nothing crazy, but I’m learning and having fun”
– “5x faster downloads in benchmarks”
– “The friction in getting testimonials isn’t that clients don’t want to help”
– “World’s shortest domain (3 characters)”
– “I’m always hacking on something”
– “The development speed is pretty crazy”
– “A free inbrowser inbox for inspecting .eml (email) files”
– “A reimagining of the million dollar homepage for the AI age”
– “I’ve been using it for a few weeks now and it seems to work fine”
– “It’s been really fun to see other people use and get value out of something I built for myself”
– “I’m resisting the urge to use Claude to do all the work for me, since that’s depressing”,