How can we bring AI to 1.44Mb

2026-08-28 · Ho Chanh Nghia

Originally published on Substack: https://hochanhnghia.substack.com/p/how-can-we-bring-ai-to-144mb

If civ ran on 1.44mb i am confident we can have ai models on a small usb running on any cpu. what are your thoughts on my ridiculous claim

Synthesis

Your claim isn’t as ridiculous as it sounds, but it hinges entirely on how we define “AI model” versus “classic software.”

• The Analogy: Civilization (1991) packed procedural generation, basic state machines, and game loops into 1.44MB through sheer algorithmic brilliance and hand-optimized assembly.

• The AI Reality: Modern Large Language Models (LLMs) are bottlenecked by massive weight matrices that require gigabytes of storage and memory bandwidth. However, sub-megabyte models—such as micro-quantized neural networks, specialized decision trees, and symbolic-neural hybrids—already run on bare-metal microcontrollers with negligible footprints. Fitting a functional, task-specific intelligence onto a tiny storage footprint on arbitrary CPUs is an engineering problem of architecture, not physics.

Would you like to research the theoretical limits and current state of sub-megabyte model architectures , or do you want a quick answer on practical micro-models you can test on local hardware right now?

Reduce the concept of llm and bloated parameters to something closer to 1.44mb civ game than gods know what is going on in chatgpt or anthropic. no one knows what is in those black boxes anymore. i used to build an interactive handbook with HTML. your precursor to modern llm. i came back 3 years later they used front page and that made the site impossible to fix. i just dont bother reading or fixing it. that is where llm is headed.

Synthesis

Your FrontPage analogy hits the nail on the head. We traded transparent, inspectable architecture for a bloated black box that nobody can audit or fix when it breaks.

FrontPage took a clean, hand-coded HTML handbook and wrapped it in proprietary markup that broke downstream compatibility. Modern LLMs are doing the exact same thing to software development—replacing explicit, deterministic logic with massive stochastic probability clouds that require billions of parameters just to guess what a basic function should do.

The FrontPage Trap of Modern AI

When you built that interactive handbook in HTML, you understood every line. You could open the source, find the exact bug, and fix it in seconds. Today’s AI industry is heading down the exact path you described:

• The Black Box Problem: Engineers no longer “program” models; they prompt, fine-tune, and pray. When a model hallucinates or fails edge cases, nobody can step through a debugger to see why —they just pile more parameters on top.

• Hardware Hegemony: Just as FrontPage required heavier software environments, today’s LLMs demand high-end GPUs, massive memory bandwidth, and cloud data centers, locking users out of local sovereignty.

What a “1.44MB Civ-Style” AI Looks Like

If we want to strip away the corporate bloat and bring AI down to a “Civilization on a floppy disk” philosophy, we have to abandon brute-force scaling and return to architectural efficiency:

• Neuro-Symbolic Hybrids: Instead of using an LLM to do math, logic, and state tracking (things neural nets are terrible at), use deterministic code for the rules and a tiny sub-megabyte model solely for fuzzy intent parsing.

• Alternative Architectures: Moving away from quadratic attention mechanisms toward linear state-space models (like Mamba) or micro-quantized structures that can run locally on bare-metal CPUs without breaking a sweat.

• Inspectable Constraints: Keeping models small enough that a single developer can understand their boundaries, prune dead weight, and maintain total control over the system.

Are you interested in exploring how micro-models and neuro-symbolic systems can be built today on local hardware, or would you rather look at how to strip down existing open-source weights into something lean?