BoredomBusted — Find Your Next Favorite Thing To Do
Discover hobbies, activities, places, and ideas that spark joy. Whether you're looking for something creative, active, social, or relaxing, BoredomBusted helps you find your next favorite thing to do.
Browse our hobby guides, things-to-do collections, and place ideas to never be bored again.
AI tinkering isn't about coding skills — it's about your ability to ask the right questions and creatively shape data, revealing unique insights along the way.
Getting started with AI model tinkering as a beginner involves running and experimenting with machine learning models using accessible tools and resources.
You load a model, tweak its settings or training data, and observe how the outputs change.
Unlike prompt engineering – which works within a model's defaults – tinkering means getting under the hood:
In AI Model Tinkering, hobbyists engage in hands-on experimentation by crafting detailed prompts for large language models, running these prompts to generate outputs, and iteratively refining them based on responses. This process often involves creating characters or scenarios, generating visual assets, and integrating them into interactive projects, allowing for a dynamic exploration of AI capab…
AI Model Tinkering fosters rapid skill feedback loops, providing immediate responses to prompts that encourage continuous experimentation and refinement, creating a flow state where engagement is sustained through visible improvements. The novelty of exploring diverse AI applications and the satisfaction of generating personalized content amplify the sense of accomplishment and creative expressio…
You think venturing into AI tinkering demands a computer science degree and a love for scouring endless documentation.
That's just not true. This hobby is more accessible and creatively fulfilling than you expect, and you don't need to be a tech guru to enjoy it.
Sure, the word 'tinkering' might bring coding to mind. But it's really about making choices: selecting data, setting constraints, deciding on outcomes.
The vital skill isn't programming – it's questioning. You've honed this ability throughout your life.
A writer, unfamiliar with code, dedicated two weekends to adjusting an open-source model with her stories. The outcome wasn't a flawless AI clone, but a creative reflection showing hidden patterns in her writing.
Tinkering doesn't just produce output – it reveals insight.
The next thing to think about is what type of tinkerer you want to be – and what you have on hand to create with.
Running a model smoothly looks easy when you watch someone else do it. You expect to follow a few steps and get results. Then your terminal sits blank, and the confident pace of every tutorial you watched feels like a lie.
The error messages start immediately. Three browser tabs open, none of them answering the actual question. tutorials skip exactly the frustration that teaches you most — the part where you don't know what a token is, why a path is broken, or what the error log is even trying to tell you. Getting through one of those moments without rage-quitting is a genuine milestone.
Week one is mostly installing, breaking, and reinstalling — actual modeling comes later. Week two is when a self-configured prompt finally works and the satisfaction hits harder than expected. week three is where one tiny setting change breaks everything, and that single crash teaches more than four weeks of reading.
Llama, Mistral, Stable Diffusion — each has its own error patterns, parameter quirks, and community shorthand. jumping between model families resets your ability to recognize what a specific error actually means. Stick to one family until the errors feel familiar, then branch out.
By week four, you'll have opinions on model sizes and settings you didn't know existed a month ago. The hype you arrived with gets replaced by something more useful — a specific sense of what actually works for you. The next section covers the mistakes that keep people stuck before they ever get that far.
When to start: Morning
Duration: 1.5 hours
Cost to try: $0
Success criteria: if you finished without encountering major code errors, do session 2.
Sorting by newest or largest feels logical — newer usually means better, right? The problem is that base models just predict the next token in a sequence. They don't know what "follow this instruction" even means.
Look for files labeled "instruct" or "chat" — those are the versions trained specifically to respond to prompts, not just complete them.
High temperature sounds like a creativity dial, so people push it up expecting richer output. What they get instead is a model that starts looping, contradicting itself, or producing word salad. Temperature doesn't add intelligence — it adds randomness, and past a certain point that randomness breaks coherence entirely.
Stay between 0.7 and 1.0 for temperature. Pair that with a top_p around 0.9 — that combination keeps the output varied enough to feel natural without letting the model drift into nonsense.
A model that exceeds your VRAM doesn't just run slowly — it offloads layers to system RAM or fails to load at all. You end up waiting minutes per response, which makes it impossible to actually learn anything. Matching model size to your hardware is what makes the difference between a usable tool and a frustrating one.
Check your available VRAM before downloading anything. A quantized version like Q4_K_M gives you most of the model's capability at a fraction of the memory cost.
Packing five requirements into a single prompt feels efficient. In practice, the model satisfies two of them, ignores two more, and partially mangles the fifth. You have no idea which instruction caused the bad output. One instruction at a time means you actually know what's working — which is the only way to improve your prompts systematically.
Add complexity only after the simpler version works. It takes longer upfront and saves you hours of debugging later.
A prompt that works perfectly in Llama 3 can produce garbage in Mistral — not because the model is worse, but because each model expects a specific chat template. When you feed it the wrong format, the model doesn't error out. It just responds strangely, and you spend an hour tweaking the wrong thing. The chat template is part of the model's training — treating it as optional means you're never talking to the model the way it was designed to be used.
Look up the specific template for each model before you start. The model card on Hugging Face usually has it.
AI tinkering happens anywhere you have a laptop and internet. Your kitchen table is fair game, though dedicated spaces like coworking hubs and university maker labs often buzz with more activity.
There's no central governing body here — which means credentials don't get you in the door, but a genuine question always will. Say you're learning, ask what others are working on, and you'll often leave with a mentor and a project.
The gap between mediocre and useful AI output is almost always the prompt. No coding, no installs — just learning how to phrase things well.
Free access gets you started. Spending $5–20 a month on API tiers speeds things up noticeably.
Fine-tuning is what you reach for when prompts stop being enough. Train a model on your own data and it starts to write in your voice or understand your specialized topics.
Best for those who've exhausted basic prompt techniques and want real ownership. Cloud GPU trials run $10–50 for meaningful experiments.
Running models like Llama or Mistral locally means no API fees and no data leaving your machine. Good fit for anyone who cares about privacy and wants hands-on control over setup.
A modern GPU with 8GB+ VRAM keeps things running smoothly. CPU is possible, but noticeably slow.
Chaining models turns them from simple response bots into decision-makers. One model reasons, another searches, another formats the output.
Suited for those with some coding experience who find single-prompt systems too limited.
Multimodal tinkering is where text, images, and audio start working together. Tools like ComfyUI for image generation and Whisper for audio give creatives a very different kind of playground.
Setup can get complex, but community support around both tools is solid.
AI Automation lives in the same world — different mechanics, similar appeal.
Readers who enjoy this often gravitate toward AI Data Projects next.
If you want a related angle, AI Chatbot Building is the natural next stop.
Most beginners focus on model settings like temperature and sampling, expecting results to magically improve. They miss the real issue.
Understanding why a response fails is the true skill that advances your practice.
When you can identify what went wrong, each tweak becomes a test, not blind guessing. No more endless retries for issues that one edit could fix.
Without this skill, you'll keep rewriting prompts while the real problem — a contradiction you wrote two drafts ago — stays invisible.
Plan for six sessions spread over 30 days — about one every five days.
If you find yourself eager to return — especially because something didn't work and you need to fix it — that's the signal. It's the chase for solutions, not just enthusiasm, that means this could be your thing. Start keeping a project log and narrow your focus to one model.
Showing up, running prompts, and feeling indifferent often means you've been relying on existing setups rather than building your own. Add two more sessions and tweak one element without a tutorial.
If indifference persists past that, it's clean data. Explore elsewhere.
If the terminal window feels like homework and configuration tweaks offer no satisfaction, this hobby is process-first — and that's a bad fit if you're wired for concrete results. That's not a character flaw. It's a straight answer.
The real sign: you're deep in a Reddit thread about model tweaking at 11pm with no project on the line and no reason to be there.
Still looking for something to do? Browse things to do when bored for more ideas.
No, you don't need a formal degree, but you should have solid programming fundamentals and basic math skills. Online resources like fast.ai and Hugging Face tutorials make it accessible for self-taught developers. Many hobbyists start with pre-trained models and learn architecture concepts as they go.
You can start for free using open-source models and cloud platforms like Google Colab, Kaggle, or Hugging Face. However, as you scale experiments or use compute-intensive GPUs, costs typically range from $5–50+ per month depending on your needs. Most hobbyists begin free and upgrade only as their projects demand it.
Fine-tuning time varies widely—from hours to days depending on your dataset size, model complexity, and hardware. A simple fine-tune on a pre-trained model might take 2–12 hours on a GPU, while training from scratch requires significantly more. Starting with pre-trained models lets you see results much faster.
Python is the industry standard and has the richest ecosystem for AI work, with libraries like PyTorch, TensorFlow, and Hugging Face. If you're new to programming, learning Python first is highly recommended. Other languages exist but won't give you the same community support or resources.
You can start on a regular laptop using smaller pre-trained models or cloud platforms, but GPU acceleration significantly speeds up training and inference. Most serious hobbyists eventually invest in a GPU-enabled setup (local or cloud-based) to avoid waiting hours for results. Free cloud options like Google Colab offer access to GPUs without upfront hardware costs.
You can build custom chatbots, image classifiers, sentiment analyzers, recommendation systems, or domain-specific tools tailored to your interests. Many hobbyists start with structured datasets (Kaggle competitions), then progress to training on personal data or real-world problems. The possibilities range from creative experiments to practical tools for work or community impact.