What reversing, modernising old games tells us about the economic impact of AI
this.os.isfine.org
this.os.isfine.org
Game development, especially this kind of porting, is an extremely technically demanding task. But it is far more constrained and precise in its requirements than the sort of software development that happens outside of the game world. Most developers deal with vague and (importantly for AI) unverifiable requirements that change continuously as a project develops. AI is not as good of a tool for this type of programming because unlike the tasks demonstrated in the article, there is no verification function an AI can cheaply use to validate if it is approaching completion of the task.
In the example in the article, the bots can look at the entire memory state of the game, interact with it to see how the state changes, and then repeat the process on their working copy of the port. So long as the overall behavior of the port is getting closer to the original, the AI can validate its getting closer to its goals. Extremely impressive stuff. But this verification loop exists in a closed world. No equivalent exists for most programming disciplines. No equivalent can exist because shifting the requirements to fulfill the vague business goals is an essential job of a software developer.
So while I am extremely impressed about what AI can accomplish every day, I also have a loud voice in the back of my mind skeptical of the actual economic impact. Because the market keeps demonstrating that these technically simpler but vague requirement tasks have a lot more value than something like game dev, at least monetarily.
Game industry generates hundreds of billions in revenue. There is plenty of value in it. The value happens to not be in decompiling 80s/90s games you have no legal right to sell, of course.
You can do N-version programming with coding agents and you get tests for free. It is easier to implement something than to test the same thing. This approach turns implementing into automated oracle for testing, takes less time (can implement in parallel), and every divergence between versions is either a bug fix or a requirement clarification. In order to make the N-versions more diverse we can use different model providers, programming language or libraries.
But none of this "implementing stuff" is making progress to solving the issue of unclear / changing / vague / missing requirements.
It took claude an hour or two to reverse the gcode dialect and camera controls. Now I just tell it what to cut, load the material, and claude sorts out where to place the work, power, speed, variations for corners and stuff, and cuts. If there’s difficulty I pop the work on a scanner and claude diagnoses and adjusts.
It is going to be very hard to control anything that’s in consumer hands, whether for good or bad reasons.
It is a faithful port of Jordan Mechner's 1989 original game and the guiding idea was just one sentence: port the game exactly, rewrite the engine completely
I didn't write a single line of code or test, not even the readme. I don't know if it delivered swift, assembly or brainfvck as directed but the result is spectacular and I would never had done it in ten lifetimes
Coding is solved, human programming is dead, apps will cost close to zero and those who don't see the writing on the wall are hopeless
Do you have an example of what you think AI is not capable of in programming? I can give it a try
i do not have an answer for your question, I may only extend from it tangentially. can you think of a way that generative ai could be offered such that end users ultimately experience a greater (rather than lesser) capacity to understanding their programming output? the loss of this understanding is one of many reasons I don't like to use generative ai.
The work has already been done. Recreate this to be identical but just in a different programming language.
I'm going to say never.
I would decompile the Google browser extension to extract the hash algorithm they used to sign page rank lookup requests, rewrite it in PHP and ship it to SEOs on weird forums.
They updated the hash every few months and I made enough money to pay for some summer activities as a teen.
One of the points the article makes is that copying is now essentially free. So if you spend money building something, someone else can copy it and use all the money they didn't spend marketing their version. And why license software anyway, when you can just clone it and even tweak it better towards your business cases? Will you be paying your Adobe subscription or buying the knockoff version of Photoshop?
That's not really what AI Boosters mean by "solving coding".
This is absolutely mind-blowing. And at the same time definitely overblown.
This stuff is impressive, but reversing is specifically what these models seems best at and porting new games to weaker hardware is way tougher than this. Not to mention making new games.
The real challenge with porting is, as I mentioned, porting modern games to constrained hardware. Specifically, porting demanding new PC games to current-gen consoles.
Trust me, operated by a skilled individual AI will have zero problem with that. I never make the case for amateurs doing these things.
Catalog is under “shop”.
The sane ones recognize that sales, service, and support have value. Is that teenager going to ensure 24/7 operation for a multinational?
The less sane ones think “if one teenager can build MS office, our thousands of SWEs can make thousands of, uh, help me out here…”
But instead of cloning their assets and rewriting their code, a dynamic x86 to WASM JIT recompiler with some web platform support code proved fast enough to run the original binaries. It is even playable on an old Android smartphone in Firefox.
The final web player is ~600KB and only needs the game installer as input, so people still have to buy/possess the game. I think this is a lot fairer for the original devs...
But the approach should work with all of those ~2000s simple 2D/light 3D Windows games, like Anno, C&C etc. that have less than 1GB of assets and no serious multithreading. Newer ones probably won't fit into a browser tab anymore or would take too long to load.
The biggest problem for the smartphone browser version is extensive use of left+right mouse input of that era which is naturally not a good fit for touchscreens.
Like most things the majority of these attempts will be garbage but some will certainly pan out into great new or extended experiences.
having that precise a grounding is literally what llms thrive in, esp across compactions.
https://github.com/s-macke/starflight-reverse
this has now become a piece of cake for the agent. The hard part has already been done. The question now is: how familiar is the agent with Forth?
...The unauthorised reproduction, translation, adaptation or transformation of the form of
the code in which a copy of a computer program has been made available constitutes an
infringement of the exclusive rights of the author. Nevertheless, circumstances may exist
when such a reproduction of the code and translation of its form are indispensable to obtain
the necessary information to achieve the interoperability of an independently created program
with other programs. It has therefore to be considered that, in these limited circumstances
only, performance of the acts of reproduction and translation by or on behalf of a person
having a right to use a copy of the program is legitimate and compatible with fair practice
and must therefore be deemed not to require the authorisation of the rightholder. An objective
of this exception is to make it possible to connect all components of a computer system,
including those of different manufacturers, so that they can work together. Such an exception
to the author's exclusive rights may not be used in a way which prejudices the legitimate
interests of the rightholder or which conflicts with a normal exploitation of the program.
...Protection of computer programs under copyright laws should be without prejudice to the
application, in appropriate cases, of other forms of protection. However, any contractual
provisions contrary to the provisions of this Directive laid down in respect of decompilation
or to the exceptions provided for by this Directive with regard to the making of a back-up
copy or to observation, study or testing of the functioning of a program should be null and void.
...The person having a right to use a copy of a computer program shall be entitled, without
the authorisation of the rightholder, to observe, study or test the functioning of the
program in order to determine the ideas and principles which underlie any element of the
program if he does so while performing any of the acts of loading, displaying, running,
transmitting or storing the program which he is entitled to do.
...In the absence of specific contractual provisions, the acts referred to in points (a) and (b)
of Article 4(1) [reproduction and alteration of the computer program — JvD] shall not require
authorisation by the rightholder where they are necessary for the use of the computer program by
the lawful acquirer in accordance with its intended purpose, including for error correction.
...The authorisation of the rightholder shall not be required where reproduction of the code and
translation of its form within the meaning of points (a) and (b) of Article 4(1) are indispensable
to obtain the information necessary to achieve the interoperability of an independently created
computer program with other programs, provided that the following conditions are met...
In case you mistake him for an “AI booster”, you should see this guy’s other posts, he is extremely critical of the AI frontier labs and Tech Bros and this whole AI bubble.
But he is also a very experienced game developer, and under his control the latest models have pulled off these crazy feats. What TFA is the analysis of a domain expert and what it means for the industry.
Does anybody have a link to it? I can't find it on the blog.
https://news.ycombinator.com/from?site=linkedin.com
They're all [dead] (you'll need "showdead" on in your profile).
It works for me in an Incognito window, but YMMV.
Let's take this claim:
> the state of the art on 3d asset generation already changed dramatically. GPT-6 Astra, GPT-6 Sol and finally Claude Opus 5.5 all moved closer into "Senior 3d artist territory"
I opened Codex, gave Astra Max access to Blender and a model, shaders, materials from Persona 5 and confirmed that it could work with it. Then I gave it reference artwork of an unrelated character and prompted it to use the Persona 5 model as a reference and model the new character in that style. The result: https://d.uguu.se/BOuGsmyR.png
Yes, it is nifty a language model can do anything at all with that prompt. But come now. Senior 3D artist. Give me a fucking break. Lie after lie after lie. Do you enjoy the feeling of clout you get from lying and people sharing your fabrications in a panic? Is that it?
I wish we could talk rationally about the capabilities of tools. Public discourse has degraded beyond reason. As one of the apparently extremely tiny minority of people who are grounded in reality, who objectively evaluates claims with their own two eyes, it is not very enjoyable to be bombarded with conversations from and between liars day in and day out. What is the purpose of this liars' society we have built?
> Public discourse has degraded beyond reason
I rest your case.
I’m telling you Astra+ is massively improved under proper guidance and it’s going to be good enough for 80% of games. Maybe not your GTA, Kojima or Call of Duty but that claim was never made. We have the pipelines of tools to deal with mass produced assets. Especially if you pass reference assets not just pictures which all game studios have aplenty.
There’s no way to assume the technology, given training data, will not solve this completely either.
Have a look at myroom.isfine.org - an 1 day experiment with one of said pipelines.
"I work for AAA studios that are routinely mocked for how low effort their productions are, and I know how cheap we can go" is a very different claim from "senior 3D artist". I don't deny you'll be able to cheapen out more than ever before and produce even sloppier games.
> Electronic Arts and Ubisoft and led game industry aligned teams at Facebook/Meta
It's amazing you can boast about your resumé, it's a who's who of deplorables known for shameless moneygrubbing. No wonder you're so enthusiastic at the prospect of replacing quality art with the cheapest slop imaginable.
> I’m telling you Astra+ is massively improved under proper guidance and it’s going to be good enough for 80% of games.
You think that model I generated is good enough for 80% of games (note also this is walking back your claim of senior 3D artist)? Let's not be liars. If even a junior artist produced that model, they would be let go immediately. I did also pass a reference asset, you know. If I'm wrong, show me. Not a room full of textured boxes that might as well be a mixup of existing asset packs - show me an actual novel human character model generated from concept artwork references. Your words have zero credibility without proof.
> There’s no way to assume the technology, given training data, will not solve this completely either.
This is a religious belief, nothing more. I have been hearing this every day for the past several years. "Trust me bro, just three more months and then it'll be AGI!"
I literally did and posted a link to the model. Honestly, fuck off.
This is exactly what I mean. A society of liars. Backed into a corner so all you can do is bring out another lie, pretending I didn't already do this in the conversation. If I try to bring reality into anything, you will deny what's literally right in front of you and continue living in a fantasy world of your own making, detached from any concept of proof or evidence. People like you have completely destroyed the value of public discourse. Congratulations.
Given your work described on this great article, how do you see a path where the LLM produces a game replication that a human could directly understand and manually improve - even if arguably unnecessary?
Take a game like Wasteland, which has a terrible modern port. I'd be interested in creating a modern port as a hobby while manually making improvements myself. But that would require code that is reasonably modular which a human can understand. Is this possible using your method and current models?
They are also usually heavier on content than other games so you have file format shenanigans like custom compression schemes or crazy memory management (see Ultima 7) to make them work / get to their internals.
If you’re lucky they have a scripting VM that drops much of the gameplay in place (Adventures of Robin Hood for example) because you can just translate it, but many older games are a mess of data files, data embedded in the EXE and procedural rules.
Coming from BioWare games and having contributed to making that massive complexity mess testable in the 2010:, RPGs are the hardest to test games overall because branch converge is very hard to achieve.
If you take for example SWTOR - designers store quest variables on random animals in the world. Thoughts and prayers even for the LLM ;)
While he's never been dismissive of the technology itself, it is fascinating to see this kind of work from him even as he rails at all the hype around AI.
Hmm,
> Remember the relentless “AI is beating everyone in Math Olympiads” and “it’s all over for humans, no point in learning STEM? Well…
> The 2025 Math Olympiad problems are out and naturally people tested the leading LLMs on those… and… spoiler .. they all fail, badly.
>That’s it LLMs are all about storage in training data which enables recombinatorial retrieval. Don’t expect them to solve or respond to new things like a human.
> They fail with novel and unknown tasks because their ability to generalize is vastly overstated by the people invested in the technology.
> The value of LLMs is in their data. No data, no answer. You have a novel piece of work that’s never been done before? OpenAI will insist it’s worth nothing because it’s a spec in the Ocean, yet that’s not true at all.
https://www.linkedin.com/posts/georgzoeller_proof-or-bluff-e...
Today you can solve IMO problems with ~$5.
Its up to you to verify that he has dismissed the technology and that he has been wrong about its performance in mathematics.
Edit:
Even without lean I can solve IMO problems using Sol/Fable today.
You made a categorical statement about LLMs
> That’s it LLMs are all about storage in training data which enables recombinatorial retrieval. Don’t expect them to solve or respond to new things like a human
> They fail with novel and unknown tasks because their ability to generalize is vastly overstated by the people invested in the technology.
It was not contextual but categorical. And it is false today.
We hadn’t broadly moved into lean proof harnesses and brute force compute spend at the time yet and the statement itself continues to be true for LLMs on their own.
The concern over what will happen to people who can no longer earn economic resources from their occupation misses the point. Post-economic abundance is already the ultimate goal of an economic system. And in so far as AI eliminates jobs, it does that by moving us to that post-economics paradigm. People are going to be orders of magnitude better off.
Imagine we lived in a world where oxygen was scarce. There were oxygen merchants that sold you oxygen so that you could breathe. And then suddenly the planet transformed and oxygen was plentiful. It became a free resource instead of a scarce resource.
Everyone would be better off. Even the oxygen merchants, while being harmed in one narrow sense, would be vastly better off in general.