Skeli789 on AI and the State of Pokemon ROM Hacks
Skeli789 is the developer of Pokemon Unbound, considered by many to be the current pinnacle of Pokemon ROM hacks. They have, it turns out, completely surrendered to the siren song of AI and have abdicated their brainpower so dramatically that, by their own admission, they rarely utilize any of the skills that birthed their claim to fame in the first place.
They made a post to the subreddit for Pokemon ROM hacks, and it is apparent that their fame and notoriety have gone to their head. The header image is this meme: The top panel is Captain America saying, "Take away that AI and what are you?" The bottom panel is a self-assured Tony Stark replying, "Still the developer of Pokemon Unbound." Skeli appears to believe that this credential not only imbues their perspective with more weight, but elevates them to a higher status than the peons who haven't accomplished such a feat. It is also notable that Skeli is immediately on the defensive, as the meme serves as a retort to a criticism that isn't even referenced in the post itself. To be fair, this may be a common line that they encounter in the Pokemon ROM hacking space. But it really seems like it's just a strawman whose defeat amounts to nothing more than self-aggrandizing bluster.
I’ll preface with my personal thoughts on the topic in general: I use AI, and every single software engineer I know does, too. Even a coworker who still codes in Vim finally cracked and now uses it for everything. Not because I (and my colleagues) can’t code otherwise, but because the time wasted not using it isn’t worth it. In the last year and a bit since AI really took off, I’ve barely had to write more than a few lines of code at a time. The progress I’ve made on my personal projects is staggering. I wouldn’t have believed I’d be able to develop as much as I did. (That also contributes to why Unbound development has slowed — I'm far more hesitant to have the AI just go at it.)
Skeli begins with an appeal to popularity: it's unreasonable to find AI coding objectionable, because look at how outvoted you are by the people who actually work in coding! In reality, the hegemony of an idea does not grant it legitimacy. AI coding, if it were acceptable, would be so even if no one did it, and, if it were not, would not be even if everyone did. This disclosure adds nothing to the discussion, other than to frame AI advocates as the worldly bunch who are already enjoying the fruits of their correct worldview regardless of whether its opponents find it distasteful. This post ostensibly exists to encourage discussion about the merits of AI in the ROM hacking scene, but Skeli situates the discourse away from the question of whether AI is good or bad in order to presuppose that it's good. Here Skeli poisons the well by portraying their opponents as naive and ignorant of how the real world has responded to AI.
But the framing of the language afterward is even more telling. Skeli insists that they have made "staggering progress" on their projects, and they are super impressed with how much developing they have been doing. But this is said immediately after specifying that their actual coding output has been essentially eliminated. Skeli didn't say they use AI as a time-saving assistant, but as a substitute for their own skills. Lest this seem like an uncharitable interpretation, it really must be emphasized that they says this themselves:
In the last year and a bit since AI really took off, I’ve barely had to write more than a few lines of code at a time.
This is an explicit admission that most of the code Skeli interfaces with is authored by an AI, not by themselves. Skeli is unambiguously announcing that they are not using their brain for coding anymore, which is bizarre because their skill in coding is the reason they became well-regarded in the first place. This is akin to a masterful author using a self-aggrandizing meme about how amazing they are at writing to introduce a post about how great ChatGPT is at writing their books for them. Even from the perspective that AI is not at all objectionable, this still wouldn't make sense; why would you solicit admiration in the same breath that you explain why your accomplishment is redundant? Again, the only conceivable answer is that Skeli is drawing attention to their impressive creation in the hope that it will make their arguments appear more legitimate. This is an appeal to authority because Skeli's skill in coding both has nothing to do with the ethical consideration of AI, and, even if it did, Skeli themselves is explaining why this skill is irrelevant anyway.
Don’t get me wrong: I still review every line of code it writes and fix everything I don’t like, but I dream of the day it gets it right to my standards in one shot every time.
This caveat is not convincing. Reviewing output, even fixing it, is far less of an exercise in cognition than producing it oneself. If I were the aforementioned masterful author arguing in favor of AI-generated books, it would be silly for me to defend myself by saying that I still proofread the text, and rewrite any paragraphs that seem wonky. It is true that this requires skill and critical thought, but it requires substantially less of it than producing the writing, or the code, yourself, from your own mind.
Notably, this puts a final nail in the coffin for the defense that my interpretation is uncharitable, and that Skeli is merely describing AI as their assistant. In this quote, Skeli is portraying themselves as the AI's assistant. The AI writes the code, and Skeli cleans it up. They have given themselves a demotion to junior-level developer in order that the AI might steer the ship for them. This is embarrassing, all the more so because their meme brags about how great they are at the skill they are trying very hard not to use.
Furthermore, they say that their "dream" is for AI to be able to completely usurp their projects, delivering audience-ready software without the need for them even to review it. This is mind-boggling! The value Skeli is so keen to remind us that they have would be utterly obsolesced by technology this powerful. Skeli would be rendered as useless as they appear to believe noncoders presently are. To be clear, this isn't exclusively about the economic risk of not having any marketable skills; in the world of Skeli's "dream," they wouldn't have value even to share the things they prompt their AI to create. Why would a Pokemon fan need to go see what games Skeli('s AI) has made if they could simply vibecode their own games just as effortlessly? In this scenario, Skeli has no value as a coder or as a player sharing fun things online. There just isn't a point to either, and it's not clear to me what about this scenario they find so aspirational. It is counterintuitive that someone with such an ego would dream about a world where they are transparently no better than anyone else.
Skeli goes on to explain how AI is not yet to that level, but that AI slop is really not so different than the slop of the past:
But to turn to ROM Hacking: slop isn‘t new to this scene. From people mass-applying ASM routine patches they didn’t understand to cloning an engine with a plethora of features every hack uses, people have been making barely edited “slop” hacks for years. But there’s no need to fear! AI slop hacks will almost certainly become the new slop norm here, and they’ll get the same reception as previous slop hacks. Sure, they’ll have fans, but most people will avoid them because they'll feel soulless. And if their developers don’t know how to code on their own, they’ll never learn to make anything actually good — just like the slop hack developers of the past.
I think this understates the concern with AI slop. The problem isn't solely that ineffectual people use AI produce bad things, it's also that AI empowers such people to produce so much of it, at such an unprecedented rate, and that it often has at least a veneer of quality on top. Already, it can be difficult to distinguish AI output from human-authored content. To use the analogy of the author, it can be hard to tell that a novel was AI-generated unless you actually read quite a bit of it. If all you see are isolated excerpts, the deeper structural issues aren't visible to you. (This is why AI slop has been known to win literary awards.) Likewise, an AI slopcoder can throw together a ROM hack with screenshots and impressive-sounding blurbs, and unless I either play it or dig into its code, I might not be able to tell that it's actually worthless. So we have found ourselves in a situation where AI has lowered the barrier of entry to creating hacks which are not obviously incompetent. We see the results of situations like this on online booksellers, where AI books have all but completely shut unknown human authors out of the market. True, it was always possible for bad authors to slap together bad novels and put them on the Kindle store. But never before has it been so easy to create so many bad novels that look decent enough to fool readers, complete with nice-looking covers and professional-sounding blurbs.
In summary, Skeli's point that "slop isn't new to this scene" misses the point. The concern isn't that slop is new, but that it's streaming from floodgates when it used to spray from a hose.
On the flip side, I saw a post this past week about a “real-time combat system.” I’m not saying this in defence of that post, but AI isn’t good enough yet to do that kind of thing on its own. You’d still need a talented developer to guide it in making that kind of thing. There will always be AI slop. And could the developer of that post have used real assets instead of AI-generated ones? Yes, and that probably would have stifled most of the negative reception. But I read the description, and the developer had plenty of engineering experience and clearly did something most people can't.
Here, Skeli says they aren't defending a thing before proceeding to defend the thing. I don't know enough about the thing to say whether the defense is deserved, but it stands out as rather strange for them to phrase it this way.
And just a cheeky aside: AI was used in coding parts of the Pokeemerald expansion (e.g. #10415 if you don't believe me). That means every modern hack based on that now uses AI code. (So I hope to see all the Pokeemerald-based hacks on HackDex labelled with the AI tag!) Objecting to this usage would also mean avoiding most hacks released these days. It just so happens that I last updated Unbound a month before ChatGPT was released, so the CFRU and the public Unbound version are still “untainted” :p.
To add some context: The Pokeemerald expansion is a codebase that many modern ROM hacks use as a base to code upon. It acts as a framework with new features, so coders can build on it rather than starting from scratch with the untouched game. Skeli's point is that the expansion has apparently integrated AI-written code to some degree, therefore all of the hacks which use it are built with AI code. (They seem to take some delight in providing this fact, perhaps with the implication that even AI-brainrotted Skeli is better at being an AI purist than the players who dislike AI.)
There is an implication here that taking a stand against AI coding in ROM hacks is pointless because many of them unknowingly use it anyway. This is a false dilemma: It is not the case that the only two options are to fully embrace AI to the extent that, like Skeli, you offload the near-totality of your cognition to it, or you reject any ROM hacks which use it at all. It is possible that someone could draw the line at any number of other places (for example, using AI as an assistant, or not actively using it and only acquiescing to its presence when it is otherwise not feasibly avoidable). However, even if we grant the black-and-white perspective that you can't disagree with Skeli and play ROM hacks without being a dirty hypocrite, this still is not relevant to the ethical status of using AI this way. The world could be full of hypocrites and AI could be bad; these are not mutually exclusive, in much the same way that I can lament how many of the products I buy have been touched by slavery at some point in the global supply chain while not choosing to sequester myself in the forest to avoid it. (One may recognize this as an application of the classic "Yet you participate in society!" meme.)
(On this, however, it may be reasonably argued that I am making an unfair inference into their argument.)
That concludes Skeli's post. There are hundreds of comments below it, many cogent, many terrible, and many in between. Some highlights:
This feels like a post just meant to rock the boat and gloat that AI has already taken over, and not actually meant to engage in any meaningful discussion. Like congrats? AI has won, you really got us, ig.
And the idea that AI might one day just write all your code for you is exciting? That sounds like a bleak future to me, anyway. Crafting something yourself is what makes it a hobby. Source
This user also pinpoints the smug tone of the post, and is followed by an insightful reply:
It does sound way more like a self-defense/damage repair post than a discussion post. By establishing that AI is an inevitability and that it's already being used in the decomp codebase (though if you look that PR you'll see the person using AI for it got told off for several things), the conversation is framed in favour of AI from the start. Source
In response to the notion that being against AI constitutes a failure to recognize nuance:
Everyone knows nuance is when you let the rich and powerful do whatever they want no matter what harm it causes without any pushback. Source
There is this reasonable take from the developer of another well-known hack, which surprised me because they also often seem to have an inflated ego about them:
AI in coding is okay as long as it’s used as an assistant to someone who already knows how to code. Total vibecoding with no comprehension of the code is the real problem here. If Claude or whatever handles the entirety of the code for you, you might end up with unwanted parts, or some of them might not work as you wanted.
However, Ai-Gen used for sprites, plot, dialogue, music and such is a huge NO Source
At the point that I reached the following comment, I figured I'd spent enough time scrolling. It constitutes one of the more comprehensive retorts to Skeli's arguments.
Hot take: Gatekeeping is actually good sometimes. If everyone could write a play and get it on Broadway, then Broadway would suck and nobody would go anymore. The internet was great until social media dumped an additional 7 billion people onto it at the same time. The barrier to entry to the internet was a good thing that kept the quality consistent and (generally) interacting with people was much better than it is today. The expansion of the internet to billions of people that can't handle it, don't understand it, and can't engage with it in a healthy and thoughtful manner has been an unmitigated disaster for society that is dissolving our social bonds and MANY people's sanity in realtime.
I have been in this space for almost 21 years now. Let me tell you, slop was always a thing. But the high time and effort cost to produce a slop hack helped keep things under control. Additionally, Pokemon ROM hack fans are also generally Pokemon fans, and the overwhelming majority of Pokemon fans have absolutely no taste or self respect, which creates a lowest-common denominator audience that is easy to dazzle with garbage. This has literally always been a problem with the Pokemon ROM hacking community and will not go away. AI will only exacerbate it.
There are so many "idea guys" and "art guys" that have not thought through what their hacks would end up actually looking or playing like. Before, they would make little pitches of how cool things would be and try to get a team together to realize their vision and ultimately fail because anyone who had any actual experience would see their brilliant game as a nightmare that nobody would want to play. The community itself helped to filter out some of the most egregious stuff and that was good. Best case scenario these people would try to do things themselves and quickly discover that their hack sucks as envisioned and would never work.
Now, they can pay money to Claude to do their hack for them. This sucks for all of us because this person now doesn't learn anything, robbing the community of a potential future contributor. Interrupting this process stops refilling the pool of humans actively working and engaging in the community, which is the only thing that has kept ROM hacking alive all these years. Claude is trained on our decomps and our hacks, so they are stealing work that was open sourced to give back to the HUMANS in the community and not to giant corporations. They use up electricity and water in the middle of a climate crisis and fuel shortage FOR NO REAL BENEFIT. And then they release stuff that is more souless, vapid, and empty than the worst slop hacks of the late 2000s. At no point do they discover that their hack is fundamentally unworkable like the guy in the last paragraph, because they're not the ones actually iterating on it!
AI is tearing the community apart. The schisms are happening in real time and I have NEVER seen an issue like this happen in all my years here. It would be great if people could just ignore AI and move on with their lives, but it is being pushed so heavily for the upcoming inevitable big tech monetization/enshittification rug pull that ALWAYS HAPPENS WITH THINGS LIKE THIS it is literally impossible to ignore.
Furthermore, I don't understand why established people would brag that they barely write code anymore. Then what the hell are you doing? There's a difference between having the automaton reorder defines for you and having the automaton take over as the primary contributor to the project. I say "the" project instead of "your" project because it isn't yours anymore once you do that. "I uNdErStAnD aNd ReViEw ThE cOdE iT gEnErAtEs." No you don't. We all know you don't. The studies on the effects of its use on human cognition show you don't. A quick skim and an OK is not review. The models are designed to generate so much code that it is impossible to properly review so that eventually your project can only be maintained by a model. Hence, the incoming monetization rug pull. Create dependence, sell subscription to the drug. That's the business model of literally every tech company on this planet.
People are all upset about artists and musicians being replaced by AI, but as soon as it's a skill they can't do or don't understand, they're all for it. Code is art. Sure, some of it is math, and some of it is well-trodden ground, but it is art.
And it's horrible because now things that used to not have any LLM garbage in them and work perfectly fine have it pushed into the backend and you can't avoid it even then. There are no decent translator applications on the internet anymore that do not use an LLM in some way, so if you want to do a translation hack for a language you're not fluent enough in to do bespoke translations but are fluent enough in to correct bad output, you're trapped. You have to use it if you don't have the resources to hire an actual localizer, and none of us would do that for a fan project.
You are not at work. The point isn't to vomit out a minimum viable product and clock out for the day. The point is the development. If you skip through the development process, you never figure out what actually works for your hack and what doesn't. You never discover things along the way that change your scope or understanding. Your work never takes on a life of its own and you never grow.
P.S. Skeli clearly using AI to make this post and in most of these replies is gross. If you can't format your thoughts yourself then they don't deserve to be shared. Don't waste our time with the ramblings of an automaton.
If Skeli did indeed use AI to write this post, then it seems that they've offloaded more than just their coding to AI; they've also surrendered their ability to think critically, formulate arguments, and express their ideas.