DEV Community: UnitBuilds

I think I found my ideal job

UnitBuilds — Wed, 02 Sep 2026 12:26:08 +0000

Have you ever read a job posting and your heart just sings the song of "HIRE ME PLEASE!!!", yeah, it's more heavy metal, if you were expecting soothing sounds...

I just read a job posting by Fractile and my heart sang out... So much so that I went to LinkedIn to stalk the personnel and send them all DMs to say PLEASE!!! (sorry about that btw)

Essentially, Rust + AI + 'we build the hardware, so you can push it's limits' = Heaven. Literal heaven, that's the kinda job that I'd wake up in the morning and love every second of it.

For those who don't know, I love optimizing and I love saying 'it can be done better' and not just say it, prove it, by building it. That's why a company that says 'hey, push the hardware limits and the hardware department will push back' is music to my ears, it means I can design systems optimized beyond what's technically possible on consumer (or enterprise) hardware, because they tailor the hardware and software to make it faster!

I haven't ever been this excited for a job posting and it shows why, accelerating AI inference has been a hobby of mine, I even experimented a lot with various techniques on how to do it, but this... This is it... For a long time, I believed that Fractals were the key to unlocking faster, smarter, less error prone LLMs and here pops up a company with it in their name! AND they're building hardware and software, specifically for it! If the rest of them didnt have their LinkedIn's private, I'd have personally messaged every person in the building.

Funniest of all, I didnt even submit my Resume, I just wrote them a massive block of text, with GitHub links, dev posts, design ideas on how I'd improve it, etc.

The other thing that sang out to me is 'no tests, but our interview will be in-depth', loved reading that! Test my capabilities, not if I can remember syntax, test whether I can make a system better, not whether I'm literate...

So fingers crossed everyone, I hope I get the job!

P.S. Anyone from Fractile here?

Who needs a health and fitness app?

UnitBuilds — Wed, 02 Sep 2026 07:06:07 +0000

So, my fiance wanted a health and fitness app, to track meals, exercise routines, manage upcoming events that disrupt her routine, accommodates for sick days, keeps data private and of course, has a digital pet.

So I made 1, featuring a digital pet (choose wisely), which evolves with you. The idea being, which creature's nature do you want to embrace? Do you want to be proud like a dragon (boost your self confidence, be strong), or gentle like a panda (handle stress better, lower anxiety). Your choice affects EVERYTHING, from the degree of exercise expected, to how it engages with you. I did it that way, because animals seem to have life figured out, we're still struggling, so embody the spirit of a panda to keep calm, humble and fun.

It includes a personal AI trainer (adjustable kindness, strictness, or dynamic for both, based on your behavior), so you can keep yourself on track and ask questions. If you want to go out partying, ask your trainer, they'll say yes or no and give you a limit.

Pre-emptive discipline. If an event is pre-scheduled, it'll adjust your routine leading up to it, to compensate. Eg. B-day party, you'll eat and drink more, so you estimate the difference and it'll adjust your plan for you leading up to it, so when the event comes, you dont feel guilty, because you dont have to work your way back to the goal, instead, you already accounted for it.

Calorie and macro tracking, that's kinda a given, snap a pic of food, it confirms and enters the stats.

Smart-device linking, like a smart watch or smart scale, so your biometrics are recorded and your pet and trainer will respond if something isnt right, eg. you suddenly get stressed and heartrate spikes, get a nice warm buzz from a panda on your wrist, with a kind message to keep you relaxed and no need to manually enter your weight or BMI, just stand on your scale, let the app track it (visibly, or hidden for the people who get OCD about number tracking).

Household tracking. It's not just you, you have a household too, they affect the groceries you buy and your routine. Either granularly, if both are on a journey, or down to just tracking average grocery consumption.

Meal-prep, either ahead of time for x days, or spur of the moment, based on your macros and your household's macros (if tracking). That's something I found important for it, it tracks your grocery receipts to see what you have in stock, so when you want a recipe for dinner, it can give you a recipe you have the ingredients for! No more "steak and asparagus" during seasons when asparagus is really expensive... It can also prep a grocery list for you, keeping it cheap based on your actual receipts, not just 'this should be cheap'.

Everything is stored locally and optionally backed up to your own cloud backup, I dont want user data, please and thank you.

Still alot I want to add, like a menstrual cycle tracker, routine planner, sleep tracker, tracking buddy, leaderboard, sharing, enhanced mental health, etc. so any thoughts of desired features, lemme know!

Another question, current setup is BYOK, or default to using your device's native AI (gemini for android, siri(gemini) for apple). That good, or would you like the option to pay for my api, so it runs enterprise channels that keeps data private?

Is this something you'd pay for (once off if you go BYOK, I hate subscriptions)? Or if it were open-source, would you still use it? Android is simple, build the apk and run it, but ios dev builds expire in 7 days... Which can be a pain. I want to keep it free, so if it were free and ad-free, would you donate, so I can maintain the apple signing license and put it on the various app stores?

Do we need drafters for credentials?

UnitBuilds — Thu, 27 Aug 2026 20:41:45 +0000

Out of curiosity, how many of you have had concerns regarding AI agents having access to your codebase and credentials in your codebase (eg. appsettings, API keys, etc.). How many of you realized it's a problem after it's saved to cloud and I imagine on a subscription, not API where your data is kept private...

Bump if you think we need a lightweight (eg. 100m param) draft model, that is trained explicitly to filter any sensitive data and scramble it, as a middleware service, to prevent data leaks to cloud. Either that, or a tiny model to run on a codebase to replace all sensitive data with references and an interaction layer for the agentic harness to call the reference, instead of directly calling the credentials from file?

What that would mean, is essentially, your SQL connection string stays safe, your API keys stay safe, your logins stay safe, if you expand, it can keep your banking details safe, while allowing you to use your AI to it's fullest, without slowing it down?

Idk, thought crossed my mind again and I thought I'd actually ask who else has thought of it?

Resume Advice from the soon to be jobless

UnitBuilds — Thu, 27 Aug 2026 19:36:17 +0000

So I just had a convo with a friend

xulingfengFollow

AI testing storyteller. Writing the 36 Stratagems series — ancient Chinese military tactics meet modern AI system meltdowns. 15yr QA → building AI test framework.

and well, he's jobless, I'm on my way there, so here's a hitchhiker's guide to getting hired.

For starters, throw away your Resume as you know it. Nobody cares where you studied, how many years of experience you have, or all the personnel skills you developed. Simply put, those are the bare minimum.

The Hiring Funnel

Hiring is done in a multi-phase funnel.

Phase 1 — Matching Filter

So what are they matching? Exactly what the job asks for. Degree, Years of Experience, skill set. That's non-negotiable — you have to meet those minimum requirements, else your application never even gets to a reviewer.

Phase 2 — Soft Match

This is where they use AI. It essentially takes the job posting and directly runs it against your CV, turning it into a scorecard. If you're not in the top 80%, consider yourself filtered out.

Phase 3 — Say Hi to a Human

This is the first time your resume ever reaches a human, but don't think they're checking your CV. They're just asking the AI "what makes this person special?"

Here's where you should pay attention.

The AI gives the reviewer your life in a single paragraph. That's why overfilling your CV is pointless — the AI won't tell them "Wow, this person is incredible!" They ask it "how good of a fit are they for this role", taking into account the company's industry. If they sell apples, make sure your CV says you sold apples, or pears — not cars or computers.

A broad skillset doesn't set you apart. We all like to think a jack-of-all-trades makes us exceptional, but they're not hiring exceptional — they're hiring exceptionally well suited for a very specific niche role.

Think of it like Cinderella's slipper. If you're trying to be too broad, you're a poor fit — you will step on toes, say bye-bye to getting hired. If your CV isn't tailored to the job's description, you underfit and you end up being seen as "needs work". Surprisingly, you're more likely to get hired as "needs work" than someone who'd step on toes — because teaching skills is easier than teaching you to restrain yourself from stepping on toes. Someone who "needs work" isn't a bad fit, they have room to grow. But you'll only get the job if they're cutting budget or can't find their Cinderella and you're the closest they can get to a perfect fit.

Short-Listed — Now You Shine

Congratulations! They now know you are qualified (Phase 1), you're capable (Phase 2) and you'd be a decent fit (Phase 3). Your CV is officially going to get read.

This is where the top 10 candidates' CVs are cross-examined by a team, who determine if you can actually add value, or if you're just a seat filler. This is where they look at:

Your past experience
The systems you worked on
The niche industries you dealt with
The hobbies you commit to
Open source projects
Released products

They aren't looking for the standard employee anymore — this is where they want YOU. And YOU need to stand out. Hackathons, Dev.to series, book publishing — these are what make you you and make you exceptional for some job!

But if you want to be exceptional for THIS job, you need to tailor your CV to it. Else you bloat their time with irrelevant stuff. Be concise and be relevant.

How to Prepare — The Two Scratch Pads

Here's part 2 of pay-attention and how you should prepare for job searching.

Scratch Pad 1 — Systems You've Worked On

Put all the systems you worked on in the past. Even if it seems unrelated, you never know. I landed my job because I did tech support on Sage Pastel — they happen to be writing an ERP system that interfaces with Sage Evolution. That's experience no degree can make up for. Don't think it's irrelevant — write it down. These are your soft skills that give you an edge.

Scratch Pad 2 — Blocks of Text

This is your fancy talk for what you did, why you did it, and how you did it — with tags for the skills involved (C#, SQL, etc.). This is what they'll actually read. This is your niche and will be the definitive reason why they hire you.

The AI Workflow

Now here's the fun part.

Step 1 — To make sure you get to Phase 3, give your AI access to:

Scratch Pad 1
Scratch Pad 2
Your default template CV (basic education and employment history — no blocks of text, no skills, nothing else)

Now that your AI has sufficient content, feed it the job posting and ask it to compile your CV from your skills and blocks of text to match this job. That way whichever skills are required get cherry-picked, along with the blocks of text relevant to it. When they read your CV, it's concise, well-structured, in your own words, and tailored to the job.

Step 2 — Red-hatting yourself. Feed a new chat the job posting and your compiled CV, then ask:

"Is this person qualified?" (Phase 1)
"How well suited are they for the job?" (Phase 2)
"Why are they an exceptional fit for the job?" (Phase 3)
"If you could score this applicant out of 100, how well suited are they?" (Interview)

If you get 90% or higher on your final score, and reading Phase 3's response makes you really feel "wow, this person would add value" — then you've done all you can to get this far. (Correct me if I'm wrong.)

Disclaimer

Keep in mind, these are my findings from personal job seeking (unsuccessfully, except for the 1 time I mentioned Pastel) and the experience I garnered from my shelved project, TUB (Talent by UnitBuilds), which involved more than a month's process of researching HR policies, hiring conditions, and what makes the system tick.

PLEASE, if you have anything to add — something that worked for you, something you'd adjust, anything that can help people get hired — mention it in the comments. I'll add it to the post. This is meant as a functional post, because at some point all of us will be searching for a job and if it can even help a single person land a job, then it's worth all the effort.

As a community, we watch each other's backs and we help where we can. There's no competition here, only strangers over the internet who have nothing but encouragement and well wishes for each other.

Anyone need an installer?

UnitBuilds — Mon, 17 Aug 2026 18:11:42 +0000

So a bit of context, I've been doing testing for V.E.L.O.C.I.T.Y. Drone and initially, I just copied over a binary, but I wanted it to be a bit easier to setup, so I thought I'd make an installer for it, so it can register as a system tray app. So naturally, I looked up what's the best installer and up popped Inno Setup 7. So I used it and it worked fine I guess, then I saw they apparently charge $155 for individuals, up to $1195 for unlimited users and that locks you to a version, if you want a new version, you need to buy a new license...

So per my usual, I built my own. It's smaller (tool), faster and completely cross-platform, using zstd with adjustable compression ratio, dependency checking, bundling, CI/CD updating, Delta-Updating, adding MSI compliance for managed deployments too and a few more nice to have features.

I'm releasing it under Apache 2.0, so it's actually free and completely open-source, use it commercially, start the next Microsoft and release a billion copies using it, you don't owe me a penny.

My reason for creating it, is like so many other times, I found that the industry gatekeeps actually making money out of software, tooling should be free, so the real products can be made, without any hidden fees.

So my question to everyone is, do you need an installer? And have you been burnt in the past by hidden costs from 'open-source' releases that charge you once you hit a revenue floor?

Who wants a Drone?

UnitBuilds — Thu, 13 Aug 2026 07:40:56 +0000

Not a flying drone (though count to think of it, that might work), I mean a drone system tray app, that secure tunnels to the drone server, which allows your AI to operate via MCP cross-system (think SSH, but for LLMs).

I'll leave details at that, if anyone has any questions, or possible scenarios they'd want to use it for, lemme know in the comments and I'll tell you whether it's capable or not, or if it's a feature I might add to it. Still ironing out some bugs, but would love to hear if anyone else would use it, besides me. Kinda sad open sourcing things that dont get used, so if there's demand, I'll release it. Else I'll just start my AI support services empire myself ;)

Anyone heard of Qoder?

UnitBuilds — Sat, 25 Jul 2026 16:07:41 +0000

With the release of Qwen 3.8 Max and it's insane 98% discount during 'off-peak' hours and 90% discount during regular hours, I decided to download Qoder to give it a go.

For reference, Qoder is essentially Antigravity, but by Alibaba and it blew my mind. Whether it's the fact that they've been throwing freebies at me since I downloaded it (100 usd worth of free SOTA 'ultimate' calls, refreshing every 2 days sofar till the 30th), or the 300 credit 'pro-trial', which goes a LOOOOONG way when you use Qwen 3.8 Max at the 98% discount. Or the fact that Lite mode is free! Like actually free, I've pushed over 20k LOC edits through it today and it hasnt even hit a fair-use cap?

I've been trying out the different platforms lately to see what's the most cost-effective, as my day-job comes with Claude Code, which is great for that, but for any personal projects, that's on my budget, with my accounts. I've been using APIs from various providers to see which is cheaper, more capable, using freebies where possible (eg. Google Startup and Azure Startup free credits) and while they work great while free, reviewing the bill afterwards is why I push them so hard... For perspective, $500 of Azure freebie money for around 50k LOC editing via OpenAI API with the model-router (discounted and routes to the cheapest model that can do the task). $150 in Google freebie money for around 20 LOC editing via Gemini 3.6 flash on Low thinking. Vs 40k LOC on $10 via Qoder (using Qwen 3.8 discounted rate), or $100 for 20k LOC via Ultimate (the expensive one).

Previously, Chinese models were 'decent', but often fell short of their western counterparts, that changed this month... Qwen and Kimi are officially better than anything in their price range, infact they punch high enough that they're the best value for twice their price range at least.

But dont take my word for it, Download Qoder and give it a try yourself. Unlike the western platforms that hide what the actual cost is till you're slapped with a bill and even then they dont actually give you clean breakdowns, Qoder is shockingly open with their free trials, showing you a detailed breakdown of what each task would have cost you. In the modern day of 'gotcha' bills because 'it takes us a while to calculate how much it costs you', Alibaba shows that it's cloak and dagger tactics to nail you with bills you cant predict. Transparent pricing and transparent usages, live. There's no excuse anymore Google and Microsoft, give us live-stats on usages.

Sorry for the poor format, lack of polish, etc. I'm knee deep in work and dont have much time to read or write, but I still wanted to put word out. If you havent downloaded Qoder yet, seriously give it a try. If you like it, you like it, if you dont, it's completely free to try.

P.S. Give the Better Loops feature a try. It's essentially a routine they built that analyzes your codebase for flaws and gives you a breakdown on it with easy to fire scoped 'quests' to fix them. They also have a 'generate wikis' routine, that's actually pretty cool. They treat documenting the codebase as codebase context which is brilliant! Instead of wasting token generating a wiki for people to read, it's actively used to prevent drift.

The 50k LOC diff

UnitBuilds — Tue, 21 Jul 2026 09:22:00 +0000

For those not keeping up, I'm developing V.E.L.O.C.I.T.Y. IDE currently. Built from the ground up to be the definitive Agentic IDE. Last week, I had a MVP, this week I'm 50k LOC deep in rust, building it out, optimizing and added 2 massive features.

a Rust-based native browser built for agentic use. I had build MCP-Lite, which was incredible, along with it's full featured counterpart, MCP-Heavy, to anyone who wants to save a dime, you'll be happy to know, I've fully ported it in... Though I am sick and tired of Chrome eating memory, so I've been hard at work building a Rust-native browser from scratch. And by from scratch, I mean no CDP, REAL from scratch, not just a chromium port.
To the people who read shelved projects, Windows Automata is making a comeback! What started as stripping the sandbox (which worked exceptionally well btw), turned into a full integration now. At this point, the IDE is turning more into an all-rounder than anything else.

And then there's a few that you probably wouldnt know about...

ORC - Orchestrative Version Control, think Git, but built for multi-agent workflows, handling discourse and merge conflicts at write-time, instead of merge time. This will serve as both the internal state-controller, as well as an optional VC integration into Github, given I imagine none of you want to leave the comfort of your caves.
V.E.L.O.C.I.T.Y. Share, my file transfer protocol that hit within 90th percentile of wire-speed, will be the backbone for ORC and will be integral to the Rust-native browser integrated into the IDE to push even more performance out of it. It will also be opened up for any file transfers you need to do securely and for backing up wherever you'd like... Just alot faster and more securely.
V.E.L.O.C.I.T.Y. MCP - the Rust-based MCP that executes in nanosecond scale, instead of milliseconds, while completely stripping JSON and Node.js. This is the core-infrastructure of the entire thing. The models use it for whatever they want to do, be it browser, WA, or development, all tooling is natively faster than any MCP you've ever used.
V.E.L.O.C.I.T.Y. OCR - A rust-based Spline driven OCR engine, will be integrated to accommodate first-pass reading of images and docs. Given it does tasks in milliseconds that take other pipelines 100+ seconds, I think it's fair to say it'll be quite useful, especially for token reduction.
V.E.L.O.C.I.T.Y. NDA - The bread and butter of it all, whether it's for the persistent SiteMap merkle root/changelog, tool definitions, or wherever else JSON would have been used, it's faster, more secure and deterministic triples, so LLMs understand it inherently.
V.E.L.O.C.I.T.Y. Remote - Secured P2P remote desktop, both in and out. That means if you're doing any kind of remote work, the combination of Remote and WA, means you can automate it. Or you know, interface with the IDE remotely on the go. All connections are driven by a secure handshake server, then transition to pure P2P, SHA 256 encrypted comms, built on the NDA standard for auditability, defensibility and reliability.

This IDE is turning into alot more than just an IDE, it's the culmination of months worth of projects, all getting slapped together into a unified product that makes life easier and cheaper. Whether it's remote IT work, Software development, or anything to do with apps or internet, it's capable and it's efficient. Zero-allocation is the mandate and it's strictly adhered to, so expect a minute fixed ram allocation. No more bloated Electron based glorified browser apps, Rust and NDA, so it's reliable, secure and effectively as close to bare-metal as you can get (until V.E.L.O.C.I.T.Y. OS is done and it gets ported to it...)

Sorry for not keeping the game series alive, though you'll be happy to know I'm about 20% of the way done with Dwarven Stronghold, think DF, except built entirely in Rust, so it actually runs decently, with a few more expansions, like the ability to switch to first/third person mode on a dwarf and take control. Never thought 100k LOC would ever be this inept... But it's getting there.

I'll post updates along the way on both as they progress. Both have grown by 150k LOC over the past week, but the 50k LOC each the past 2 days has been night and day in terms of capability.

Disclaimer: My AI is busy, so I wrote this by hand. Except the cover-image, that's NanoBanana...

Life hack?

UnitBuilds — Thu, 16 Jul 2026 22:25:13 +0000

Today I discovered something pretty cool... OpenRouter has free models! And not just baby 2b models, 550b MoE from Nvidia (Nemotron 3 Ultra), along with some really cool ones I discovered along the way, like HY3 by TenCent and North Mini Code by Cohere.

This is a game changer... We all pay subscriptions, or API fees, I built V.E.L.O.C.I.T.Y. IDE to just get a cheaper alternative, using Kimi K2.7. But then I discovered, DeepSeek is shockingly cheap. You get a REALLY competent coding model for sub $1/M tokens in and out... But this... This is the real victory. V.E.L.O.C.I.T.Y. IDE still needs some polishing, but connected with a Cloudflare API key, or OpenRouter API key, it gives you access to any model you want! And that includes free ones.

I'll post an update once it's at a state where it's 'nice' to use, so you all can give it a try.

Why is nobody talking about this???

UnitBuilds — Thu, 16 Jul 2026 13:19:23 +0000

Hy3 is free for the next 5 days? For perspective, that's a nearly 300b model that is currently free. Seriously, get on it! Wire it up via Claude Code, or whatever you use, make the most of it!

From 5,000-Parameter Failures to a 2.5D Dream: My Quest to Out-Simulate Dwarf Fortress

UnitBuilds — Mon, 13 Jul 2026 13:53:57 +0000

For years, I’ve been utterly captivated by the concept of Dwarf Fortress. If I’m being completely honest, the actual execution of the original game felt a bit brutal to play, but the sheer, uncompromising depth of its simulation nature always fascinated me.

Naturally, as a developer with a penchant for optimization and procedural systems, I decided to build my own.

Spoiler alert: I was a bit too ambitious.

🪦 The Graveyard of Good Intentions

My journey to this point is paved with ambitious, half-finished prototypes that fell victim to scope creep and performance bottlenecks:

1. 🏢 The Over-Engineered Engine

A year ago, I tried my hand at a deep world sim. I gave every single entity over 5,000 parameters. The result? Managing anything beyond 10 on-screen elements at once was inconceivably slow. I scrapped it.

2. 👥 The Procedural Identity Crisis

Next, I tried to build a Sims-style game. That quickly evolved into adding Dwarf Fortress-style lore, complex relationship webs, and eventually procedural everything. As a solo developer with limited time, it was an impossible mountain to climb.

3. 🐜 The Weekend Hackathon Divergence

Just this past weekend, I joined a dev challenge. I started building an AI Ant Terrarium. The tech stack was wild: a Solana-seeded world state, Snowflake for storage, Gemini 1.5 Flash for generating real-time ant thoughts when clicked, and ElevenLabs for turning those thoughts into spoken inner monologues. It looked incredibly cool on paper, but I hit a wall: turns out generating convincing procedural foliage is significantly harder than generating procedural terrain.

So, I decided to scale back, strip away the noise, and return to the root of my passion: Dwarf Fortress.

⚔️ Enter "Dwarven Stronghold" (100% Non-IP-Infringing)

This is a mission of pure passion. I started the codebase in JavaScript, quickly realized I needed bare-metal performance, and rewrote the entire core in Rust.

Currently, the engine sits at over 200 robust unit and integration tests—passing clean and green on every build. It's still heavily a work in progress, but the vision is clear: Dwarf Fortress has a legendary ~750k lines of code benchmark. I want to match and surpass it in every single way.

📜 The Feature Manifesto

📐 No ASCII: Ditching the retro text for a sleek, modern 2.5D isometric perspective.
🌊 Unmatched Depth: High-fidelity 3D Navier-Stokes fluid velocity, NPK soil nutrient depletion, full stress psychology, and dynamic, procedural audio.
🎙️ AI Integration: Real-time voice-over and dialogue powered by ElevenLabs, translating silent text into immersive voice-overs.
👤 True First/Third-Person Mode: Eventually, you’ll be able to drop out of God-mode, choose a perspective, and play as an actual dwarf roaming your own fortress.
🥽 VR Support: Long-term, I want full virtual reality integration.

👶 The Child Within the Machine

When I look at why I'm so obsessed with these systems, it goes back to my favorite childhood game: Black & White 2.

There was something magical about that game's natural progression—the deep trainability of your creature, and the ability to seamlessly zoom from the heavens all the way down until you could see the exact paths individual ants were walking.

I think that's exactly why I am where I am today. I have an absolute fascination with procedural generation, machine learning, extreme code optimization, and the philosophy that no depth is too deep.

🎨 Overcoming the "Asset Bottleneck"

Right now, Dwarven Stronghold runs strictly in the console. If I'm being vulnerable, part of me is terrified to build the visual layer. I'm scared that looking at it will doom it to the same fate as its predecessors—halted for eternity because I can't get the graphics to match the brilliance of the underlying simulation.

As a solo dev, hiring a massive design team to scale with a game this vast is an economic impossibility. If I can add over 1,000 new features or items to the simulation logic in a single day, how could a traditional asset pipeline ever keep up?

The solution is an automated pipeline.

By choosing a 2.5D isometric aesthetic, I can simplify asset generation. I'm building template prompts and using AI asset-generation pipelines to dynamically generate sprites. It’s not perfect yet, but it’s a scalable start—and it’s the shield that will keep this project out of the graveyard.

💖 Sustained by Passion

My goal is to open-source the project the moment the core loop is functional. Who knows? Maybe the community will find it interesting, catch the vision, and want to contribute to building this massive world.

Whenever I feel overwhelmed by the scope, I think about Tarn and Zach Adams. Dwarf Fortress was started by two brothers with a shared passion, and it became their life's work. In a bittersweet twist of fate, they eventually brought it to Steam to pay for skyrocketing cancer treatments. It’s a poignant reminder of what we do this for.

Let passion drive you, and passion will sustain you—even in your darkest of hours.

💬 Let's Discuss:

How do you handle scaling asset pipelines for deep simulation games?
What is the sweet spot between simulation realism and performance in modern engines?
Have you ever scrapped a massive project because you were too intimidated by the visual layer? Let me know in the comments!

Model Kombat: The LLM Fighting Game!

UnitBuilds — Sat, 11 Jul 2026 18:59:06 +0000

Ever wondered what would happen if the world's leading Large Language Models settled their benchmark disputes in a 2D cybercity arena?

It's easy to look at model performance on standardized benchmarks (like MMLU, MATH, or HumanEval). It is much more fun to visualize their underlying architectures, parameter scales, and hardware constraints as a retro-cyber fighting game.

So, we built Model Kombat (Mixture of Experts Edition)!

🕹️ Play Directly Here

🎮 Launch Game in Full Screen

🧬 Playable ML Concepts Explained

This isn't just a basic stick-figure fighting game. Every mechanic—from rendering complexity to the speed at which characters recover—is a direct, playable representation of real-world Large Language Model engineering.

1. 📐 Parameter Scaling vs. Render Tiers

A model's representation capacity (intelligence) scales with its parameter count. In Model Kombat, a fighter's visual complexity, joint detail, and rendering fidelity directly reflect its real-world parameter size:

Tier 1 (< 5B Parameters - Gemma 2B, Llama 3.2 3B) - Primitive Capsules: Drawn as simple, single-color flat limbs with low joint segmentation. This visualizes the limited representation capacity and coarse output resolution of small edge models.
Tier 2 (7B - 14B Parameters - Mistral 7B, Claude Haiku) - Simple Vectors: Structured as thin skeletal wireframe vectors.
Tier 3 (14B - 35B Parameters - Gemini Flash, Mixtral) - Two-Tone Vectors: Rendered as dual-color, layered vector limbs.
Tier 4 (35B - 100B Parameters - Llama 8B, Claude Sonnet) - Cyborg Shading: Rendered as detailed vector cylinders with dynamic code particle streams flowing along their limbs.
Tier 5 (> 100B Parameters - o3, GPT-4o, Claude Opus) - Quantum Vectors: Rendered as glowing vector limbs with digital matrix code particles, soft drop-shadow depth buffers, and real-time afterimage motion trails.

2. ⚡ Reasoning Tokens & KV-Cache Overcharging

Instead of arbitrary "mana" or "stamina," fighters charge a Ki bar representing internal processing cycles and Reasoning Tokens (inspired by reasoning chains like OpenAI's o-series):

Charging Ki: Simulates the time-to-first-token (TTFT) phase, generating reasoning tokens.
Limit Break: Overcharging past 100% enters a golden-outlined Limit Break state, granting high-speed afterimages and super-armor.
Context Eviction & Dizzy: If a model holds its overcharged state too long, its context window overflows. This triggers Context Eviction—draining the model's HP and placing it in a Dizzy state. This represents how context window saturation degrades model coherence and leaves it vulnerable to failure.

3. 🌀 Mixture of Experts (MoE) Routing

Sparse Mixture of Experts (MoE) models do not activate all parameters on every token; instead, a gating network routes tokens to specialized experts.

Active Experts: Native MoE models in the game (like Mixtral and DeepSeek) dynamically spawn floating indicator nodes representing active Text, Math, or Vision experts.
Routing Buffs: Landing hits routes computation to these experts, granting temporary combat buffs:
- TEXT FFN: Increases walk speed (low-latency generation).
- MATH FFN: Increases damage output (logical compute).
- VISION FFN: Expands attack hitboxes (spatial awareness).

📊 Spotlight: The AI Fighter Registry

Here is a breakdown of 6 key models featured on the ladder, explaining where they excel in real life and how their unique passives translate into combat advantages:

1. 🧠 o3 (OpenAI)

In Real Life: OpenAI's state-of-the-art reasoning model. Rather than generating text instananeously, o3 employs a reinforcement learning-driven "thinking chain" to plan, verify, and correct its logic before outputting a response. This makes it a titan in mathematics, competitive programming, and complex coding.
In-Game (Tier 5): Deep Thinking Chain passive. o3 charges its Ki at double speed (representing the model's heavy pre-response thinking cycles). This lets you quickly max out your meter, activate MoE, or release powerful special attacks.

2. 📜 Claude Opus (Anthropic)

In Real Life: Anthropic’s flagship heavy model. Opus is celebrated for its high-nuance reading comprehension, literary synthesis, and strict compliance with ethical and safety guidelines (governed by Anthropic's "Constitutional AI" framework).
In-Game (Tier 5): Constitutional Blade passive. Opus has an extended melee strike range on all punches and kicks, allowing you to control the neutral game and keep opponents at a distance—visualizing the model's massive context processing reach.

3. 🎤 Gemini Ultra (Google)

In Real Life: Google's largest multimodal model. Unlike models that stitch together separate speech-to-text and vision encoders, Gemini is built natively multi-modal from day one. It processes video, audio, and text simultaneously inside a single model architecture.
In-Game (Tier 5): Ultra Stance Shift passive. Gemini can switch between TEXT, VISION (expanded hitboxes), and AUDIO (increased speed) modality stances instantly without the standard stance-transition delay, adapting to any opponent's position on the fly.

4. ⚡ DeepSeek V3 (DeepSeek)

In Real Life: The groundbreaking open-weights model from DeepSeek. V3 utilizes Multi-Head Latent Attention (MLA) to compress Key-Value caches, dramatically reducing VRAM footprint, alongside a massive Multi-head Latent Attention routing gating network.
In-Game (Tier 4): MLA Attention passive. Compress caching translates to high combat evasion: DeepSeek V3 has a 15% chance to phase-dodge incoming projectiles completely.

5. 🗼 Mistral 7B (Mistral)

In Real Life: The legendary French open-weights model that punches far above its weight class. Mistral 7B introduced Sliding Window Attention (SWA) to the open-source community, allowing the model to handle longer context streams with minimal performance decay.
In-Game (Tier 2): Sliding Window passive. The sliding window translates to faster execution: Mistral has 10% less startup frame lag on punches and kicks, letting you land strikes before your opponent's animations can finish.

6. 📱 Llama 3.2 3B (Meta)

In Real Life: Meta’s mobile-first, edge-optimized model. Llama 3.2 3B is trained specifically for local deployment on smartphones and tablets, focusing on high efficiency, quick response times, and localized fine-tuning.
In-Game (Tier 1): Fine-Tuning passive. The model is incredibly resilient to pressure: Llama 3.2 gains +5% defense every time it successfully blocks consecutive hits, adapting to the opponent's combo string.

🏁 Beat the Machine & Share Your Score

Once you defeat all 19 opponents on the ladder, you will face the reigning champion: o3.

Can you reach the top of the ladder and claim victory? Click COPY SCORE at the end of your run and paste your stats in the comments below!

💬 Let's Discuss:

Which model is your favorite main?
What was your longest combo chain on the ladder?

UnitBuilds-CC / Model-Kombat

Arcade fighting game, which LLM will come out on top?

📟 Model Kombat (SYS_08) 🧠🥊

An Interactive, Playable Visualization of Large Language Model Constraints & Architectures

Welcome, AI engineer. Model Kombat is a retro-cyber fighting game designed to teach the mechanical constraints, architectural paradigms, and hardware limits of Large Language Models (LLMs).

In this game, 20 real-world AI models face off in a 2D arena. Every visual element, movement mechanic, and combat stance directly translates core machine learning engineering concepts (such as parameter scaling, KV-caching, Mixture of Experts, multi-modality, and alignment safety) into interactive gameplay loops.

🕹️ Play Directly Here

🎮 Model Kombat Live Web App

Git Repository: https://github.com/UnitBuilds-CC/Model-Kombat

🧠 Educational Core Concepts & Game Translations

1. 📐 Parameter Scaling vs. Render Tiers

In deep learning, a model's representation capacity (intelligence) scales with its parameter count. In Model Kombat, a fighter's visual complexity, joint detail, and rendering fidelity directly reflect its real-world parameter size:

Tier 1 (< 5B Parameters -…

View on GitHub

Disclaimer: AI was used throughout this project, it is just fitting that it would co-author with me, so special thanks to the Foundry for its tireless hours toiling away and Gemini for producing the cover image.