Why don't you boycott Jaffar-Plus? It's a brute-forcer. It consumes electricity, like LLM datacenters, and you'll burn Gigawatts of electricity before actually achieving something useful in your TASes.
GenAI and brute forcers like Jaffar are not directly comparable.
TASVideos does not support LLM usage in its current state. The environmental impact caused by data centers, the emotional and mental damage done to people manipulated to rely on it, and the overall ethical concerns of, among many other things, harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it, are far too much to overlook individually, let alone all together.
If the message is to prevent environmental impact caused by TASVideos members, then we may as well put our TAS-related prompts to chats on one scale, and Jaffar-Plus bruteforces performed by eien86 on another scale. Who consumes less energy? xD
TASing is like making a film: only the best takes are shown in the premier.
https://xkcd.com/3246/
Joined: 1/7/2023
Posts: 121
Location: Somewhere on planet earth probably
Dimon12321 wrote:
LLM processions is basically electricity consumption. Why don't you boycott Jaffar-Plus? It's a brute-forcer. It consumes electricity, like LLM datacenters, and you'll burn Gigawatts of electricity before actually achieving something useful in your TASes.
This is a classic case of "False Equivalence". You are telling us that two very fundamentally different things are the same and trying to justify your point by doing that.
Also, you are typing this on a computer, which uses electricity. Why don't we boycott YOU for using electricity? Why don't we boycott every person in the world that has ever used an electric device? Why don't we eradicate eels, seeing as they also handle electricity? Oh wait, the brain produces electric signals. Why don't we boycott every creature that has a brain?
Your point is so stupid that, if it were a person, it wouldn't have to worry about those brain boycotts I suggested earlier
"He who is tired of TASing is tired of life" Something Homer Simpson definitely said
I think the proposed policy is mostly good, but I think it is missing some details with regards to attribution. I remember a couple years back seeing discussions of the first AI videos and a common phrase I saw was 'get ready because this will only get better, faster.' If you compare AI videos now to back then that has certainly turned out to be true. The same was true in math and programming, with major advancements in math happening just last week with basically just a few AI prompts. my point is that eventually, probably faster than we think, someone will submit a TAS with a submission text like 'I fed this to Claude and it found an improvement in this run, I basically did nothing lol.' How will authorship be attributed in this case? Will people be able to add 'Claude' as coauthor or will it be treated similar to ordinary botting despite the submitter doing next to no work?
If the message is to prevent environmental impact caused by TASVideos members, then we may as well put our TAS-related prompts to chats on one scale, and Jaffar-Plus bruteforces performed by eien86 on another scale. Who consumes less energy? xD
That is not the point of the policy. The policy is being created in order to distance the site from all the terrible things that current LLM technology involves, and also to contribute to the global push-back against all those terrible things.
Warning: When making decisions, I try to collect as much data as possible before actually deciding. I try to abstract away and see the principles behind real world events and people's opinions. I try to generalize them and turn into something clear and reusable. I hate depending on unpredictable and having to make lottery guesses. Any problem can be solved by systems thinking and acting.
Joined: 7/11/2016
Posts: 50
Location: 🇰🇷 South Korea
KusogeMan wrote:
Tas is a technical achievement that can be better enabled by AI and I'd like to see new TASers that start TASing because of it and old TASers as well be enhanced through the new tools AI allows for. I personally can't think of the benefits so far but it's pretty clear smarter people found ways to use it and i'm happy their work is being done easier and faster. This is a hobby which is often underappreacited, not very valued, takes a load of time and i don't wanna see anybody being discouraged of TAsing for using AI an inteligent way.
...
What would be a healthy limit for AI usage in TASing? seems like a more reasomable question if i read correctly
Note that the draft recognizes and allows the usage of genAI in TASing to some extent:
it is still possible to use LLMs for assistance in TASing through helping with analysis and code writing. It is, in theory, also possible for AI-generated assets to be used in an arbitrary code execution run. Due to these possibilities, we are treating these usages the same as any other contextual usage: We strongly discourage, but will still hesitantly allow, text generation in research and analysis, while media generation will still be outright banned ...
The forbidden ones are AI-generated media, which you wouldn't usually need anyway. Even for ACEs or drawing-based games, I think it's a reasonable restriction, especially considering that accepted TASes are published to YouTube for everyone to watch, and it's a polarizing topic among the audience as well.
Also I think this thread was created exactly to discuss a healthy limit for AI usage in TASing.
EDIT: how did I end up saying "publicated"
A possibility I am not certain about from the draft though, is that it might be possible to hook AI agents with TASing softwares to directly interact with the game and create inputs. (maybe... I'm also against genAIs in the current state and not gonna pay hundreds of dollars just to test this lol) Would this fall under text generation under research and analysis, or something else?
If the message is to prevent environmental impact caused by TASVideos members, then we may as well put our TAS-related prompts to chats on one scale, and Jaffar-Plus bruteforces performed by eien86 on another scale. Who consumes less energy? xD
That is not the point of the policy. The policy is being created in order to distance the site from all the terrible things that current LLM technology involves, and also to contribute to the global push-back against all those terrible things.
Then why do we need the second sentence in this paragraph, at least, at the beginning of the policy? If we read it through, it has almost no relation to the concern points (actual to-the-point statements) described further. What is located at the top tends to have the highest meaning.
TASVideos does not support LLM usage in its current state. The environmental impact caused by data centers, the emotional and mental damage done to people manipulated to rely on it, and the overall ethical concerns of, among many other things, harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it, are far too much to overlook individually, let alone all together.
TASing is like making a film: only the best takes are shown in the premier.
https://xkcd.com/3246/
Joined: 8/30/2020
Posts: 213
Location: 🇦🇺 Sydney, Australia
I echo Dimon12321's sentiments: The criticism of AI for its energy consumption applies to only a handful of (mostly American) tech companies, and will likely cease to be relevant after the bubble bursts.
I have zero respect for anyone who relies on AI but only uses it by sending their prompts off to the USA via a webpage or Electron app, and even paying for the privilege of doing so, because it's possible to run "open" models which are about a year behind the SotA on your own machine. An LLM can write a new interface/harness for itself, and may be able to self-improve generally. Funding (for example) Anthropic's next PR campaign in exchange for a timeshare in the latest model is dumber than buying new corn seeds every year or shipping in fresh air from overseas.
TASes can be art, but not always. The characteristic shared by almost all TASes is that they are inherently objective, which is rare outside of mathematics and computer science. (See recent news stories re: AI and mathematical conjectures.) Objectivity is the bane of an LLM trying to produce a working run token-by-token, but it's also the thing that enables someone to build a system for improving a run without human intervention.
There's a history of TASers employing computational methods to optimise movement or the routing of the run itself, and I see new ML techniques as just that, new techniques/tools with which to solve problems. It happens that not everyone has the background to understand and implement those techniques for botting, but now there's the option of having an LLM explain them to you or even writing the code itself (see above).
The new rule(s) should be a precise and narrow prohibition on the behaviour(s) you want to dissuade people from. Don't want vibe-summarised notes with submissions? Write that.
Split the sections re: use of AI by users away from the sections re: use of AI by/for the site. If you have "ethical concerns of [...] harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it", you should ditch reCAPTCHA and YouTube, since those feed Google's panopticon.
I contribute to BizHawk as Linux/cross-platform lead, testing and automation lead, and UI designer. This year, I'm experimenting with streaming BizHawk development on Twitch. nope
Links to find me elsewhere and to some of my side projects are on my personal site. I will respond on Discord faster than to PMs on this site.
Hey look buddy, I'm an engineer. That means I solve problems. Not problems like "What is software," because that would fall within the purview of your conundrums of philosophy. I solve practical problems. For instance, how am I gonna stop some high-wattage thread-ripping monster of a CPU dead in its tracks? The answer: use code. And if that don't work? Use more code.
Then why do we need the second sentence in this paragraph, at least, at the beginning of the policy? If we read it through, it has almost no relation to the concern points (actual to-the-point statements) described further. What is located at the top tends to have the highest meaning.
TASVideos does not support LLM usage in its current state. The environmental impact caused by data centers, the emotional and mental damage done to people manipulated to rely on it, and the overall ethical concerns of, among many other things, harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it, are far too much to overlook individually, let alone all together.
I don't understand which part of my answer doesn't apply to what you're asking.
Warning: When making decisions, I try to collect as much data as possible before actually deciding. I try to abstract away and see the principles behind real world events and people's opinions. I try to generalize them and turn into something clear and reusable. I hate depending on unpredictable and having to make lottery guesses. Any problem can be solved by systems thinking and acting.
I find the draft at EDIT 4 mostly alright, with the exception of one moment:
it is still possible to use LLMs for assistance in TASing through helping with analysis and code writing. It is, in theory, also possible for AI-generated assets to be used in an arbitrary code execution run. Due to these possibilities, we are treating these usages the same as any other contextual usage: We strongly discourage, but will still hesitantly allow, text generation in research and analysis
What does "still" mean in the bold phrase? You may prohibit it in the future?
I'll tell you what. If any policy hinders me with producing TASes in whichever way I want, I will leave this community. Or have a bunch of skilled Lua programmers monitoring #scripting channel on the Discord server, so I get my Lua scripts up and running, instead of pointlessly asking for help. I have plenty of stuff to learn and do. Lua skill is a low-priority for me, because it has no other use in my life.
If you want to play in eco-activists, go out of the streets and protest against OpenAI, Claude and other hyper-corporations. People use whatever they are offered. That's the corporations who can pull the plug off.
LLM processions is basically electricity consumption. Why don't you boycott Jaffar-Plus? It's a brute-forcer. It consumes electricity, like LLM datacenters, and you'll burn Gigawatts of electricity before actually achieving something useful in your TASes.
P.S: I'm happy NFTs failed. The first humanity victory in this 21st-century money-ruled world. Too bad cryptocurrency and mining still exist.
feos wrote:
Dimon12321 wrote:
If the message is to prevent environmental impact caused by TASVideos members, then we may as well put our TAS-related prompts to chats on one scale, and Jaffar-Plus bruteforces performed by eien86 on another scale. Who consumes less energy? xD
That is not the point of the policy. The policy is being created in order to distance the site from all the terrible things that current LLM technology involves, and also to contribute to the global push-back against all those terrible things.
Dimon12321 wrote:
feos wrote:
Dimon12321 wrote:
If the message is to prevent environmental impact caused by TASVideos members, then we may as well put our TAS-related prompts to chats on one scale, and Jaffar-Plus bruteforces performed by eien86 on another scale. Who consumes less energy? xD
That is not the point of the policy. The policy is being created in order to distance the site from all the terrible things that current LLM technology involves, and also to contribute to the global push-back against all those terrible things.
Then why do we need the second sentence in this paragraph, at least, at the beginning of the policy? If we read it through, it has almost no relation to the concern points (actual to-the-point statements) described further. What is located at the top tends to have the highest meaning.
TASVideos does not support LLM usage in its current state. The environmental impact caused by data centers, the emotional and mental damage done to people manipulated to rely on it, and the overall ethical concerns of, among many other things, harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it, are far too much to overlook individually, let alone all together.
YoshiRulz wrote:
I echo Dimon12321's sentiments: The criticism of AI for its energy consumption applies to only a handful of (mostly American) tech companies, and will likely cease to be relevant after the bubble bursts.
I have zero respect for anyone who relies on AI but only uses it by sending their prompts off to the USA via a webpage or Electron app, and even paying for the privilege of doing so, because it's possible to run "open" models which are about a year behind the SotA on your own machine. An LLM can write a new interface/harness for itself, and may be able to self-improve generally. Funding (for example) Anthropic's next PR campaign in exchange for a timeshare in the latest model is dumber than buying new corn seeds every year or shipping in fresh air from overseas.
TASes can be art, but not always. The characteristic shared by almost all TASes is that they are inherently objective, which is rare outside of mathematics and computer science. (See recent news stories re: AI and mathematical conjectures.) Objectivity is the bane of an LLM trying to produce a working run token-by-token, but it's also the thing that enables someone to build a system for improving a run without human intervention.
There's a history of TASers employing computational methods to optimise movement or the routing of the run itself, and I see new ML techniques as just that, new techniques/tools with which to solve problems. It happens that not everyone has the background to understand and implement those techniques for botting, but now there's the option of having an LLM explain them to you or even writing the code itself (see above).
The new rule(s) should be a precise and narrow prohibition on the behaviour(s) you want to dissuade people from. Don't want vibe-summarised notes with submissions? Write that.
Split the sections re: use of AI by users away from the sections re: use of AI by/for the site. If you have "ethical concerns of [...] harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it", you should ditch reCAPTCHA and YouTube, since those feed Google's panopticon.
I agree by all four points. AI is bad, and I really like that everybody is pushing back against AI and the world's global AI pushback as a whole. But the policy in its current state needs fixing. And these points are just the perfect solution. TASing must not be restricted to only specific approved techniques. TASers should be able to TAS however way they want.
I've read through the AI Policy draft and, while I agree with the main idea, I think the wording in the TASVideos AI Policy section is too opinionated.
The first section, Site Code AI Disclosure, is pretty much perfect as it is. It clearly lays out the promise that TASVideos staff will not use agentic code going forward, with a disclosure that some code was committed before the policy was drafted. As a user, that's pretty much all I need.
The section after that, TASVideos AI Policy, has the right idea, but it's more opinion than actual policy. I'd take out most of this and leave the disclosure that while TASVideos does not support LLM usage in its current state, we cannot fully guarantee that code generated by LLMs will never slip into the site or supported software through PRs from outside users. Statements like "Unfortunately, we are forced to live in a reality where LLMs are rapidly proliferating and forcibly being implemented into every facet of modern technology" is not a policy, it's just a way of telling the reader that it's being allowed tolerated in some situations despite the writer's own opinion, because there's no practical way to ban it completely. Again, I think you could cut out most of that section and still make it clear that site staff will not use Generative AI going forward and they prioritize No-LLM alternatives to software they support.
As for the rest of the policy, I feel the two sections under User AI Policy are fine. I also appreciate the clarifications at the bottom. Other than that, I don't have much else to say about it.
I echo Dimon12321's sentiments: The criticism of AI for its energy consumption applies to only a handful of (mostly American) tech companies, and will likely cease to be relevant after the bubble bursts.
I have zero respect for anyone who relies on AI but only uses it by sending their prompts off to the USA via a webpage or Electron app, and even paying for the privilege of doing so, because it's possible to run "open" models which are about a year behind the SotA on your own machine. An LLM can write a new interface/harness for itself, and may be able to self-improve generally. Funding (for example) Anthropic's next PR campaign in exchange for a timeshare in the latest model is dumber than buying new corn seeds every year or shipping in fresh air from overseas.
TASes can be art, but not always. The characteristic shared by almost all TASes is that they are inherently objective, which is rare outside of mathematics and computer science. (See recent news stories re: AI and mathematical conjectures.) Objectivity is the bane of an LLM trying to produce a working run token-by-token, but it's also the thing that enables someone to build a system for improving a run without human intervention.
There's a history of TASers employing computational methods to optimise movement or the routing of the run itself, and I see new ML techniques as just that, new techniques/tools with which to solve problems. It happens that not everyone has the background to understand and implement those techniques for botting, but now there's the option of having an LLM explain them to you or even writing the code itself (see above).
The new rule(s) should be a precise and narrow prohibition on the behaviour(s) you want to dissuade people from. Don't want vibe-summarised notes with submissions? Write that.
Split the sections re: use of AI by users away from the sections re: use of AI by/for the site. If you have "ethical concerns of [...] harvesting data from non-consenting users and not only selling it back to them, but allowing others to effortlessly profit from it", you should ditch reCAPTCHA and YouTube, since those feed Google's panopticon.
FitterSpace wrote:
I've read through the AI Policy draft and, while I agree with the main idea, I think the wording in the TASVideos AI Policy section is too opinionated.
The first section, Site Code AI Disclosure, is pretty much perfect as it is. It clearly lays out the promise that TASVideos staff will not use agentic code going forward, with a disclosure that some code was committed before the policy was drafted. As a user, that's pretty much all I need.
The section after that, TASVideos AI Policy, has the right idea, but it's more opinion than actual policy. I'd take out most of this and leave the disclosure that while TASVideos does not support LLM usage in its current state, we cannot fully guarantee that code generated by LLMs will never slip into the site or supported software through PRs from outside users. Statements like "Unfortunately, we are forced to live in a reality where LLMs are rapidly proliferating and forcibly being implemented into every facet of modern technology" is not a policy, it's just a way of telling the reader that it's being allowed tolerated in some situations despite the writer's own opinion, because there's no practical way to ban it completely. Again, I think you could cut out most of that section and still make it clear that site staff will not use Generative AI going forward and they prioritize No-LLM alternatives to software they support.
As for the rest of the policy, I feel the two sections under User AI Policy are fine. I also appreciate the clarifications at the bottom. Other than that, I don't have much else to say about it.
I agree fully with both of these posts (even if they don't necessarily agree with each other!) and just want to add my worthless take.
A lot of the concerns are overblown (especially with regards things like energy - I can generate a thousand HD images locally on my PC in about an hour and watch the energy usage match that of about the same amount of time gaming, or as mentioned before about the same amount of time bruteforcing a solution in a TAS); a lot of the concerns are genuine (especially with regards to slop - mass production of poorly fact-checked content is a major problem, especially when it then gets used as poorly curated input back into a generative model to perpetuate the cycle).
I think it's perfectly understandable to take a stance on LLMs and similar generative algorithms, but I think the arguments are most effective when made from a quality standpoint, as that's a completely uncontroversial concern and already exists currently as a rejection reason. I don't think anyone wants to see submission texts that start "This isn't just a tool-assisted speedrun of Hybrid Front—it's the fastest, the most optimized, and the most technically refined run to date." (DISCLAIMER: I asked ChatGPT to write that sentence only.)
At the same time, I've seen live someone pasting a memory dump from a Spectrum game into Claude and asking it how the RNG function works, and getting back a correct (I checked it myself to make sure) answer almost immediately, complete with memory locations and Z80 assembly code of the function. While I wouldn't do this myself (I find it fun to figure these things out the "old-fashioned" way) these things make creating a TAS more accessible to those without the deep assembly knowledge (or free time to learn it), and could result in high-quality TASes that otherwise wouldn't have been made.
On the other hand (yes I'm alternating stance by paragraph) it's important that such use is minimal, human-verified, and generally discouraged to avoid hallucinations and slop. I'll repeat that I think quality assurance is the best argument to make here.
I actually agree with YoshiRulz more than it sounds, but I also understand that the staff have already pledged not to use Generative AI, so I'm just trying to help make sure the written policy is as objective as possible.
If you want my personal opinion (which should have no bearing on any decisions for this site), I personally don't care if people use LLMs for coding, translations, etc. I think they're very useful tools that aren't going away any time soon, even after the bubble pops. I haven't messed with them much, but I can see how it would be preferable to combing through documentation or bothering someone (or interrupting someone else asking for help) in #scripting to write a LUA script.
That Super Mario World TAS that got submitted recently got a bunch of nasty replies in the submission thread because they disclosed the fact that they used ChatGPT to help write a LUA script (among other things I'd rather not get into right now). I just find that whole situation really silly, because I highly doubt it's the first TAS that used AI somewhere in the process, and it certainly won't be the last. The mean comments won't stop people from using LLMs, it'll just stop them from disclosing it.
I was going to write a paragraph about the downsides of AI being so easily accessible, and that it paves the way for a ton of low-effort garbage that wouldn't otherwise exist, but I'm sure everybody reading this is already aware of that. It just means you'll have to spend more time moderating forums and GitHub PRs to filter out that garbage. However, there is still a lot of good content TASers are putting a lot of effort into, and we shouldn't discourage that just because they used AI to help with a small part of it. It's only going to get more common as the tools get better.
Thanks for this. I think the Igor thread was sad. Putting the religious stuff aside, I don’t think the onus should be on an author to morally justify their use of AI (replace AI with LLM/Generative AI/whatever as required). An author also shouldn't be shamed or embarrassed in their submission. Either the site allows AI or it doesn’t. I don’t think the discussion needs to go much beyond that.
I read the policy draft and wanted to share a few thoughts.
Generally, I think the content could be streamlined. I appreciate that many people are interested in the site’s values, but, as an author, I mainly want to know what is and isn’t allowed. Can I use AI to write my submission text? What if I only use it to fix my spelling and grammar? Can I use AI to write Lua scripts or other supporting tools for me? (admittedly I have no idea how you’d ever check this). I’m not saying these questions haven’t been answered, but it would be nice if I could find things like this at a glance in a more clear-cut manner. I think whatever rules are finalized for text in particular should also be presented to authors at the moment of submission (just so there are no excuses!).
I agree with FitterSpace that the draft feels too opinionated. This is most obvious in the TASVideos AI Policy section, which feels a bit bloated and may invite unnecessary challenges. Another example is the paragraph beginning with ‘Just remember, you have an infinite amount of time to make a TAS.’ While I appreciate the sentiment, I think statements like this are unnecessary and potentially provocative. For example, maybe I don’t have as much time as I’d like to make a TAS, or to learn how to write Lua scripts, or to learn Assembly to understand a complicated glitch (I was actually faced with this problem recently). These are cases where AI could be useful to people.
In general, I don’t think so many opinions/value judgements are needed. You could outline at the beginning that LLMs are not supported due to various ethical concerns (perhaps with some links/references to sources that describe the issues in more detail) and have a brief statement of the site’s values. The rest of the policy could then focus more on practical questions about what is and isn't accepted, accessibility considerations etc.
This may or may not matter if the policy is streamlined in the way I’m imagining, but it seems to me that there is not a sufficiently clear distinction between AI and 'the AI industry'. I think this is important because most of the moral issues seem to concern the latter. The sentence ‘TASVideos does not support LLM usage in its current state.’ seems clear enough (although what is meant by state? state of the technology or the industry?), but later we have ‘This policy will not change until the AI industry does. For as long as it remains unchallenged, unregulated, and unethical, TASVideos staff and developers will remain opposed to using it.‘ Here it seems ‘it’ refers to the industry. Admittedly I don’t know much about this, so what I’m saying might just be rubbish, but, does this for example mean that the site is not in principle opposed to self-hosted, open-source, AI models?
Joined: 8/30/2020
Posts: 213
Location: 🇦🇺 Sydney, Australia
Oh yeah, I forgot to bring this up:
our developers have pledged to not use agentic code on the site at all going forward.
No pledge or documentation has been checked in to the repo yet.
I contribute to BizHawk as Linux/cross-platform lead, testing and automation lead, and UI designer. This year, I'm experimenting with streaming BizHawk development on Twitch. nope
Links to find me elsewhere and to some of my side projects are on my personal site. I will respond on Discord faster than to PMs on this site.
Hey look buddy, I'm an engineer. That means I solve problems. Not problems like "What is software," because that would fall within the purview of your conundrums of philosophy. I solve practical problems. For instance, how am I gonna stop some high-wattage thread-ripping monster of a CPU dead in its tracks? The answer: use code. And if that don't work? Use more code.
I think they're very useful tools that aren't going away any time soon, even after the bubble pops. I haven't messed with them much, but I can see how it would be preferable to combing through documentation or bothering someone (or interrupting someone else asking for help) in #scripting to write a LUA script.
This point is brilliantly (IMO) disputed over at SMWCentral:
BD_PhDX wrote:
How Generative AI Erodes Community
It's no mystery that getting into SMW hacking has a high learning curve. Assembly is difficult to learn, it isn't always clear what tools do what task, and the information resources available can sometimes be tough to follow and understand. The library of knowledge it takes to feel free to create what you wish can feel hard to parse at best, and outright unwelcoming at worst.
Still, the library exists. More that that, the people who have contributed to that library are still here, and willing to share their knowledge with you. Those who did the work to understand Super Mario World, build the tools, or write the information or resources available are still around. Additionally, so are all the people who continue to contribute to or build on it to keep us moving forward.
Generative AI takes the work those people have done and removes the personal contributions entirely. By inserting itself into a space where a personal conversation would otherwise happen, generative AI erodes the connections between people, and undermines that very community we've built to help others learn and get into this hobby. How does anyone have a chance to feel part of this community if a chatbot is their go-to instead of real people?
Additionally, because we do not have line of sight into the answers given by generative AI, we aren't able to grow, learn, and adapt what we do to better address people's questions and needs. How does the community continue to grow when people opt for a convenience option over engaging with a human to understand imperfectly available information? How can we can improve things if people won't share with us their struggles?
A community only grows and thrives if it continues to have people willing to engage with others in it. In one that revolves around a craft like ROM hacking, that means talking to others interested or knowledgeable in the craft. Every time someone engages with generative AI to answer a ROM hacking question, a connection to the community lost. The illusion of convenience that a chatbot gives is incomparable to actually talking with people who you should be confident in being helpful.
There's no shame in not knowing things. We all start out that way, and we gain knowledge and grow as creators by engaging with other, more experienced creators. We then have the opportunity to pass that on. That's what being part of a community is about. We have a human responsibility to carry that forward that generosity, be helpful to others, and to leave the community better than we found it. Generative AI not only does none those things, but actively fosters harm to the community through removing people from it.
The only case I'd understand is if someone did ask community members several times but got no answer. Then using LLM is the lesser evil/last resort.
Warning: When making decisions, I try to collect as much data as possible before actually deciding. I try to abstract away and see the principles behind real world events and people's opinions. I try to generalize them and turn into something clear and reusable. I hate depending on unpredictable and having to make lottery guesses. Any problem can be solved by systems thinking and acting.
AI ban is still not checked in to the repo.
Would you consider allowing AI-assisted decomp/disasm in userfiles (Lua scripts) and wiki pages? Also, are homepages included in wiki pages?
edit:
Non-agentic forms of generative AI, such as neural networks and other machine learning algorithms, are still allowed.
In deep learning, the transformer is a family of artificial neural network architectures based on the multi-head attention mechanism, in which text is converted to numerical representations called tokens, and each token is converted into a vector via lookup from a word embedding table.
I contribute to BizHawk as Linux/cross-platform lead, testing and automation lead, and UI designer. This year, I'm experimenting with streaming BizHawk development on Twitch. nope
Links to find me elsewhere and to some of my side projects are on my personal site. I will respond on Discord faster than to PMs on this site.
Hey look buddy, I'm an engineer. That means I solve problems. Not problems like "What is software," because that would fall within the purview of your conundrums of philosophy. I solve practical problems. For instance, how am I gonna stop some high-wattage thread-ripping monster of a CPU dead in its tracks? The answer: use code. And if that don't work? Use more code.
Joined: 11/13/2006
Posts: 2934
Location: Northern California
YoshiRulz wrote:
AI ban is still not checked in to the repo.
Presumably it will be when this becomes official.
Would you consider allowing AI-assisted decomp/disasm in userfiles (Lua scripts) and wiki pages? Also, are homepages included in wiki pages?
As the rest of the policy should have made clear, anything directly containing AI-generated content is not allowed. If it doesn't directly contain anything AI-generated, it's fine. Yes, homepages are included, because they are wiki pages.
??? LLMs are neural networks.
"Non-agentic" is meant to imply anything outside of the current prompt-based plagiarism agents. No ChatGPT, no Claude, etc. If it's that unclear, then at least give me a suggestion on how to word it properly instead of trying to educate me on something I have no interest in learning further.
I'd like to point out that local, on-device AI options are available, such as desktop apps like Jan and Ollama, along with open-weight and/or open-source models like GLM-5 and Qwen3.5.
Well-written draft, I see no major issues.
Despite its occasional shortcomings, what AI can do nowadays is astonishing and scarily powerful; it can generate images and videos, transcribe audio inputs, summarize and simplify text, and even perform tasks on the browser. Useful as is, the need for controlling AI usage is often highlighted, yet there are still bad actors and tech companies wanting to leverage this power for their benefits.
AI Policy Draft 2 wrote:
Usage of LLM-assisted machine translation tools for site communication and accessibility is allowed as long as the original untranslated text is provided.
I don't suppose linking back to the original source would be an option here, would it?
Also, there exists voice-typing/speech-to-text tools for outputting text with capitalization and punctuation through your voice, such as Handy and OpenWhispr, in case manual typing happens to become a burden. I'd assume usage of those tools must be disclosed or perhaps be restricted to a minimum as well.
Myself, I don't use AI in excess (currently; only AI-assistedtranslations through DeepL), but if I happen to have AI by my side in my projects, I know better than to claim all the credit; some won't exactly know better. Perhaps, at most, warnings may be needed if other users aren't being truthful about AI usage too.
At least let us remember that sometimes there's this handy little disclaimer at the bottom that says something like, "AI can make mistakes. Be careful/double-check responses, important info, etc./and so on."
I try to check in whenever possible, at least when I'm feeling it.
Look out for yourself and each other; there's quite a bit going on.
Usage of LLM-assisted machine translation tools for site communication and accessibility is allowed as long as the original untranslated text is provided.
I don't suppose linking back to the original source would be an option here, would it?
Linking to the original text, while perhaps convenient, isn't a good option. The linked content might change over time (e.g. a wiki article), or might disappear altogether. Having the original on the site ensures that it won't become lost. Plus, it may help deter those who would abuse translation as an excuse/avenue for using generative AI to create forum posts.
Asumeh wrote:
Also, there exists voice-typing/speech-to-text tools for outputting text with capitalization and punctuation through your voice, such as Handy and OpenWhispr, in case manual typing happens to become a burden. I'd assume usage of those tools must be disclosed or perhaps be restricted to a minimum as well.
Voice recognition has existed for decades, even Windows XP had it built in (it wasn't very good, but that was 25 years ago.) Most tools like this will have used some form of neural network. The difference nowadays is that it's not just trying to match fragments of a voice sample to a word, but rather a more complex system of analysis and "interpretation" indicative of LLMs. Bottom line though is that it's an AI that is generating text. Without a decisive way of determining that a tool addresses the concerns of this policy, I don't think it's possible to allow any of the text generated by those tools.
As an alternative, people could presumably upload their voice recording somewhere and link to that for people to listen to. It's not ideal considering this is a text-based site, but if someone is physically incapable of writing text themselves, that'd be an option. Human transcription is also an option; they could ask a friend or someone in this community to transcribe their voice to text.
Joined: 8/30/2020
Posts: 213
Location: 🇦🇺 Sydney, Australia
Bigbass wrote:
Asumeh wrote:
[something about STT]
Voice recognition has existed for decades, even Windows XP had it built in (it wasn't very good, but that was 25 years ago.) Most tools like this will have used some form of neural network. The difference nowadays is that it's not just trying to match fragments of a voice sample to a word, but rather a more complex system of analysis and "interpretation" indicative of LLMs. Bottom line though is that it's an AI that is generating text. Without a decisive way of determining that a tool addresses the concerns of this policy, I don't think it's possible to allow any of the text generated by those tools.
As an alternative, people could presumably upload their voice recording somewhere and link to that for people to listen to. It's not ideal considering this is a text-based site, but if someone is physically incapable of writing text themselves, that'd be an option. Human transcription is also an option; they could ask a friend or someone in this community to transcribe their voice to text.
So someone who struggles to use a keyboard can be conveying their own thoughts with their own words in the way that's most comfortable to them, and if you somehow catch wind of them using certain STT software, they get banned? And your best suggestion is essentially "hire a friend as an unpaid typist". That doesn't sound right to me.
Samsara wrote:
Would you consider allowing AI-assisted decomp/disasm in userfiles (Lua scripts) and wiki pages?
As the rest of the policy should have made clear, anything directly containing AI-generated content is not allowed. If it doesn't directly contain anything AI-generated, it's fine.
Alright. I think that is a myopic decision, since current AI decomp is almost on par with humans (anecdata: [1][2][3], quantitative comparison: [4]). There's a huge gap between how much annotated game code currently exists and how much would be useful to TASers, and I imagined that TASVideos would be a good place to keep that centralised and public.
Samsara wrote:
YoshiRulz wrote:
Non-agentic forms of generative AI, such as neural networks and other machine learning algorithms, are still allowed.
??? LLMs are neural networks.
[snippet from enwp]
"Non-agentic" is meant to imply anything outside of the current prompt-based plagiarism agents. No ChatGPT, no Claude, etc. If it's that unclear, then at least give me a suggestion on how to word it properly [...]
It's become muddied by overuse as a buzzword, but I understand 'agentic' to mean "having the ability to make and execute a multi-step plan for completing the given task with little-to-no human oversight, such as by spawning and overseeing separate worker instances", or in short, "having agency". I would contrast that term with 'chatbot', though in both cases your instructions would be called a 'prompt' (same with diffusion image generation). My wording suggestion is "AI systems tailored to specific tasks, not general text/image generation, are still allowed."
I don't believe it's possible to distinguish the kind of AI usage I don't like from the rest, except by subjectively estimating how much effortcare the human author put into the program/article/whatever. Were I to try, I would resort to capping the scale (parameter count) of the model, or the cost per token including amortised training costs (I'm not sure if that's available or even measurable). Your examples gave me one other idea: limit usage to open-weight models run locally (my thoughts on that are explained above).
Samsara wrote:
[...] instead of trying to educate me on something I have no interest in learning further.
Being ignorant is normal, being reluctant to learn about something which you have contempt for is understandable, but I'd hope you in your role as legislator would want to be aware of at least the names of the things you're trying to regulate and the broad strokes of how they work.
I contribute to BizHawk as Linux/cross-platform lead, testing and automation lead, and UI designer. This year, I'm experimenting with streaming BizHawk development on Twitch. nope
Links to find me elsewhere and to some of my side projects are on my personal site. I will respond on Discord faster than to PMs on this site.
Hey look buddy, I'm an engineer. That means I solve problems. Not problems like "What is software," because that would fall within the purview of your conundrums of philosophy. I solve practical problems. For instance, how am I gonna stop some high-wattage thread-ripping monster of a CPU dead in its tracks? The answer: use code. And if that don't work? Use more code.
There's a huge gap between how much annotated game code currently exists and how much would be useful to TASers, and I imagined that TASVideos would be a good place to keep that centralised and public.
What can possibly make tasvideos a good place to host full decomp of a game?
Warning: When making decisions, I try to collect as much data as possible before actually deciding. I try to abstract away and see the principles behind real world events and people's opinions. I try to generalize them and turn into something clear and reusable. I hate depending on unpredictable and having to make lottery guesses. Any problem can be solved by systems thinking and acting.
There's a huge gap between how much annotated game code currently exists and how much would be useful to TASers, and I imagined that TASVideos would be a good place to keep that centralised and public.
What can possibly make tasvideos a good place to host full decomp of a game?
Echoing this. I'm not clear on the legal ramifications that could exist for us doing so in the first place, but it's especially risky with AI decomps, since they may well include illegally obtained source code (Nintendo gigaleak, Game Freak teraleak etc.) as training data.
Echoing this. I'm not clear on the legal ramifications that could exist for us doing so in the first place, but it's especially risky with AI decomps, since they may well include illegally obtained source code (Nintendo gigaleak, Game Freak teraleak etc.) as training data.
To be quite frank when it comes to legal ramifications, decomps are almost always infringing regardless of how you look at them. Decomps are more only available when the copyright holder simply doesn't care enough to bother doing a takedown (or think it's better for them to not do a takedown), but they would very well be in their legal rights to send a takedown. "Don't include the assets" is cargo cult legal fantasy here, that doesn't magically make a decomp of all of a game's code fair use. It's basically only useful in the sense of maybe making the copyright holder less inclined to bother with a takedown (but said copyright holder can proceed to do a takedown regardless).
To remain in fair use territory, you would only be able to release very small portions of decompiled code in practice, with high level explanations of what the code does. Something that'd probably in a submission text and would easily fall under fair use. That, and/or, never release your entire decomp.
I wouldn't worry about leaked source code as training data here anyways, unless you are outright looking at the games with leaked source code existing (which case you probably should be more worried about humans using leaked source code, absent of AI concerns). They aren't particularly useful for some generic AI doing a random game's decomp.
