I won't beat around the bush this edition. Want to know something neat that happened this week? I taught an AI agent how to (preliminarily) play Final Fantasy Tactics Advance (FFTA) for the GBA:
Raw memory addresses for Final Fantasy Tactics Advance, which help give eyes to an AI - a memory address is a place you store more raw information
If you've noticed my pixel portrait on public media (as of present), it's actually a derivative style of a character portrait from FFTA; you put 300 hours into a childhood game and tell me you don't share a cosmic bond with everybody that played the exact same one. Looking at you, Legacy of Goku II players. Also, I named my character a button mashed ZWZVXWYVP because as a kid, I expected I'd just reset the name at some point. Never did, and here we are!
From an AI perspective, this is pretty nuts precisely because of the following reasons:
FFTA is not an open-source or decompiled (aka: fully reverse-engineered) game.
It's a complex turn-based story with elements of random chance.
It has a lot of different states, including...
A world map
A battle map
An entire party/job system
An in-battle random law system where certain actions may be forbidden or encouraged
And much more.
In other words, from a traditional standpoint, this would be a nightmare to automate with an AI agent.
It's not a clean API (doesn’t exist), not well documented - the only best next sources are cheat code sources (that take very precise values like HP, money, etc. and do simple value toggles on them). Nothing as complex as "where is the character on the map right now, and how do I move them?"
That being said, this is where AI-enabled engineering takes good root. mGBA, which is an emulator engine for the GBA, exposes scripting under the hood. With that, I can inspect the raw application / loaded ROM state at any given point! Second problem. All of it is obfuscated: nothing is clean or neat like "here's the enemy group you're fighting and their positions". It's closer to the screen below, where we're actively reverse engineering what-exists-where at a given point in memory.
If you were doing this manually, you'd frankly go fucking nuts (pardon my vulgarity). But with an AI coding agent that can see the outputs of our scripted bridge and send commands over to the game by manipulating memory directly? Easier.
AI automatically checking / examining memory address values
AI finds the memory addresses needed to build heuristics / reproduce behaviour!
All of this is a perfect example of how tasks previously impenetrable by AI and AI agents are now suddenly surmountable. In one of my past articles, I wrote about how I hung out / co-worked with a company in New York while I was out there: they're working on financial agents that can trawl the web and make payments on behalf of users/companies. Stripe, Circle, and other fintech companies are doing similar.
Every application that can be interfaced-with, or inspected, can be modified and reverse-engineered by an AI agent now. It does not matter how complex or obfuscated it is (in the case of FFTA); it is only a matter of time until behaviour can be precisely engineered regardless of the original source. In software, there's a term called "clean room" implementations, where source code is copied in functionality (but not expression) to avoid copyright (or copyleft). Good example here.
This has massive implications for interactions on the internet, security, agentic automation, and so on - for practically every category of software and human interaction. It's not unlike the significance of the Human Genome project, to sequence and understand human DNA for the progress of medicine.
When software can communicated to as easily as "move this unit here", or "finish the story using only X units or job classes", and an agent reverse-engineers a path to it, we have eventual computational analogies to "create a medicine that cures human thyroid cancer without affecting the host healthy cells". Both are structured systems with hidden states that respond to precise inputs.
It just so happens that some of the best representations of poorly understood (but structured) systems are compiled obfuscated video games from the early 2000s. Speedrunners spend their entire careers trying to get faster at a single game (or category of games) - ZFG being a famous one for the Legend of Zelda: Ocarina of Time. He discovered a famous glitch known as "arbitrary code execution", or in other words: forcing the game to do anything the player wanted.
I don't think we'll be far off from AI agents being able to do that with, well, anything. You take Google's Gemma model, toss it at something you've never seen before, and watch it break down walls you never knew existed. You are not conscious of your own limits (especially with AI), even if you think you are.
As always, be well.
Michael Kirsanov