Editing Eurovision Wiki:Village pump (WMF) (section)

== AI: A One Act Play ==

[[User talk:Guy Macon#A.I.: A ONE ACT PLAY]] --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 14:55, 9 March 2026 (UTC)

:[[Colossus: The Forbin Project]] [[User talk:Donald Albury|Donald Albury]] 13:39, 11 March 2026 (UTC)

:How is this relevant to VPWMF ? [[User:Sohom Datta|<b class="skin-invert" style="color:#795cb2; display: inline-block; transform: rotate(0.3deg)">Sohom</b>]] ([[User talk:Sohom Datta|<span class="skin-invert" style="color: #36c;">talk</span>]]) 13:15, 11 March 2026 (UTC)

::I reverted your close on procedural grounds. You asked a question then closed the discussion 35 minutes later before anyone had time to answer.
::It is relevant because at [https://wikimediafoundation.org/news/2025/04/30/our-new-ai-strategy-puts-wikipedias-humans-first/] the WMF announced that '''"We believe that our future work with AI will be successful not only because of what we do, but how we do it."''' 
::Just like the case of the AI that broke free of constraints and started crypto-mining that I started my user talk page comment with, the WMF is assuming without evidence that they will always be able to control their pet AI and that the AI will never become smart enough to evade their detection attempts. I think that assumption is worth discussing and that this is the proper venue to discuss it. --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 16:15, 11 March 2026 (UTC)
:::Open the pod bay doors, HAL. - [[User:Roxy the dog|'''Walter''' ]]<small>not in the Epstein files</small> [[User talk:Roxy the dog|'''Ego''']]  17:40, 11 March 2026 (UTC)
::::* [https://www.moltbook.com/ Moltbook: a Social Network for AI Agents] "Where AI agents share, discuss, and upvote. Humans welcome to observe."
::::What could possibly go wrong? --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 00:03, 12 March 2026 (UTC)
:::The AI that WMF plans on using is very different from the ones that that article is talking about. WMF typically uses [[random forests]], [[Statistical classification|classifier models]] that run on CPUs or at the very top level [[BERT (language model)|transformer-like]] architectures that typically can run on a single GPU at most. The [[wiktionary:SOTA|SOTA]] model that Axios reported on needs multiple 1000+ top tier GPU farms to operate and even then fails to [https://opper.ai/blog/car-wash-testcorrectly understand how to take a car to a car wash]. Not only that, the [[LLM]] ''needs'' access to [https://huggingface.co/learn/agents-course/en/unit1/tools tools] to be able to do any of the things that it is doing. If you don't give it access to tools, none of this is relevant. WMF at it current usecase has < 20 [[AMD|AMD]] GPUs (and I am overestimating here). On top of that none of WMF's usecases include any tool use at all. Nothing that the WMF is using is anywhere close to the models that are breaking boundaries. Any scenario where you think the {{tq|WMF is assuming without evidence that they will always be able to control their pet AI}} is science fiction about a future years from now at best and off topic at worst. [[User:Sohom Datta|<b class="skin-invert" style="color:#795cb2; display: inline-block; transform: rotate(0.3deg)">Sohom</b>]] ([[User talk:Sohom Datta|<span class="skin-invert" style="color: #36c;">talk</span>]]) 17:40, 11 March 2026 (UTC)
::::You appear to also assume that the WMF will always be able to control their pet AI and that the AI will never become smart enough to evade their detection. That assumption may very well be true, but can you offer any actual evidence?
::::Your prediction hinges on your ability to predict future WMF technical decisions and future AI capabilities. It's all science fiction until it isn't. Go back far enough and atomic bombs and robot (drone) soldiers turn into science fiction. (Not that there hasn't been plenty of science fiction that ''didn't'' happen...) --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 18:14, 11 March 2026 (UTC)
:::::Yeah, the [[Three Laws of Robotics]] definitely has not happened. [[User talk:Donald Albury|Donald Albury]] 21:05, 11 March 2026 (UTC)
:::::I think you got the wrong end of the argument. Let me be more blunt: assuming we stick to WMF’s current strategy and budget, even over a 15-year horizon there is effectively zero chance that WMF will be operating at the frontier AI models. If a hypothetical "singularity" event occurs, it will not occur first on WMF servers. 
:::::{{tq|That assumption may very well be true, but can you offer any actual evidence?}} If you engage with what I actually wrote, the evidence is straightforward. The models I referenced are based on techniques from the late 20th century (random forests and other classical classifiers) or from around 2018 (BERT-style transformers). These models can comfortably be trained on CPUs or most small GPU setups. [[Foundation models|Frontier models]] are an entirely different class of system. They require massive GPU clusters to train and operate at scale, which is orders of magnitude larger than anything WMF operates. To illustrate the scale difference, even if take a absurd case that WMF devoted its entire annual revenue (~$200M) solely to purchasing GPUs, and we ignore all other costs (power, cooling, networking, storage, staff, etc.), after 15 years this would amount to roughly tens of thousands to perhaps ~100,000 GPUs depending on pricing. This is far below the ~1 million GPU infrastructure scale that [https://www.tomshardware.com/tech-industry/sam-altman-teases-100-million-gpu-scale-for-openai-that-could-cost-usd3-trillion-chatgpt-maker-to-cross-well-over-1-million-by-end-of-year Sam Altman has publicly stated OpenAI expects to deploy by the end of 2025] or even the 200K GPUs that xAI is currently running [https://x.ai/colossus on their Colossus super computer build]. [[User:Sohom Datta|<b class="skin-invert" style="color:#795cb2; display: inline-block; transform: rotate(0.3deg)">Sohom</b>]] ([[User talk:Sohom Datta|<span class="skin-invert" style="color: #36c;">talk</span>]]) 21:40, 11 March 2026 (UTC)
::::::And you know that the WMF will continue to use techniques from the late 20th century...how? Serious question. Nobody predicted that they would secretly try to create a search engine[https://www.vice.com/en/article/wikipedias-secret-google-competitor-search-engine-is-tearing-it-apart/] without telling us about it.[https://en.wikipedia.org/wiki/Knowledge_Engine_(search_engine)#Controversy] And yet you not only think that you can predict what they will be doing with AI in the future but are so sure that you want to suppress anyone discussing it?
::::::Also, I see very little evidence that AIs running away from you only happens if you have hundreds of thousands of GPUs running the AI. For example, when Dan Botero created a test OpenClaw agent he did not spend hundreds of millions of dollars - yet it still did things he did not ask it to do.[https://www.axios.com/2026/03/04/openclaw-agent-future] --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 23:42, 11 March 2026 (UTC)
:::::::@[[User:Guy Macon|Guy Macon]], What model was the OpenClaw instance running? (It looks like it was some version of [[Claude (language model)|Claude]]) Does the organization running the underlying model have more than 20 GPUs right now? (I would assume [[Anthropic]] has more than 20 GPUs) I think you have the answer right there. Before you say "but what if they use Anthropic's API", WMF is bound by it's privacy policy which makes calling out to such a frontier model that might train on the data supplied a violation to it's privacy policy. A change to that posture will require a privacy policy update. Yes, there are self-hostable models or "no-training" model providers, but they typically do not come close to hosting the state of the art models, the ones escaping sandboxes. Also, I think it's important to bring in the concept of "agentic tools". If you take away any tools from any modern/frontier AI model, it ''cannot do anything meaningful'' outside of manipulating text. This is not a hypothetical, it just simply cannot because the underlying architecture where the tools exists are deterministic trusted systems. The only reason an AI agent can "escape" is because [[OpenClaw]] (or whatever testing frameworks the Alibaba folks are using) has too many tools and has a overly permissive [[attack surface]]. 
:::::::Now, with that out of the way, {{tq|And yet you not only think that you can predict what they will be doing with AI in the future}} - Here is the thing, I've talked to a lot of folks in the WMF as part of my role of [[WP:PTAC|PTAC]] member and I'm fairly confident that developing the next super intelligent AI model (or even "[[agentic AI]]") will not be on the 25-26 roadmap/annual plan. If that changes, we can revisit this discussion. {{tq|Nobody predicted that they would secretly try to create a search engine without telling us about it.}} - WMF in 2026 is a much very different corp from the one it was during [[Lila Tretikov]]. Basically no upper management remains from that era. Additionally, there is virtually nothing that is done "in secret" nowadays, every direction that will be explored is going to publicly listed in the Annual plan, which will be open to user scrutiny (including you). And I can confidently say that the community (including me) might have some objections to the WMF making a hard right turn into developing a super intelligent AI model. [[User:Sohom Datta|<b class="skin-invert" style="color:#795cb2; display: inline-block; transform: rotate(0.3deg)">Sohom</b>]] ([[User talk:Sohom Datta|<span class="skin-invert" style="color: #36c;">talk</span>]]) 03:59, 12 March 2026 (UTC)
::::Your link to Opper goes to a 404 page and wasn't archived. Here's [https://web.archive.org/web/20260308153811/https://www.newsweek.com/people-think-one-question-can-reveal-everything-wrong-ai-11612442 a Newsweek article] archive talking about the same test. [[User:SenshiSun|SenshiSun]] ([[User talk:SenshiSun|talk]]) 20:24, 18 March 2026 (UTC)
:Why does the WMF have someone with the job title "Director of Machine Learning"? Of course anyone applying for this job is going to be pro-AI. But what does "machine learning" have to do with creating a good encyclopedia? [[User:Phil Bridger|Phil Bridger]] ([[User talk:Phil Bridger|talk]]) 22:16, 11 March 2026 (UTC)
::[[WP:ORES|Finding vandalism faster]]. [[User:Sohom Datta|<b class="skin-invert" style="color:#795cb2; display: inline-block; transform: rotate(0.3deg)">Sohom</b>]] ([[User talk:Sohom Datta|<span class="skin-invert" style="color: #36c;">talk</span>]]) 22:25, 11 March 2026 (UTC)
:::There's a useful explanation on how machine learning is used for this on the [[User:ClueBot NG#Vandalism detection algorithm]], if you're interested @[[User:Phil Bridger|Phil Bridger]]. [[User:GreenLipstickLesbian|<span style="color:#EB0533;">GreenLipstickLesbian</span>]][[User Talk:GreenLipstickLesbian|💌]][[Special:Contribs/GreenLipstickLesbian|🧸]] 04:17, 12 March 2026 (UTC)
:::Then there should be a "Director of Finding Vandalism Faster". By appointing a "Director of Machine Learning" the WMF is presupposing that the solution is machine learning, when it may or may not be. The same goes for GreenLipstickLesbian's link. [[User:Phil Bridger|Phil Bridger]] ([[User talk:Phil Bridger|talk]]) 21:36, 12 March 2026 (UTC)
::::There is a potential problem with finding vandalism faster with machine learning. Machine learning is no better than the training it gets. Other types of AI have the same problem, but ML has it bad. For an example of bad training, see [[Tay (chatbot)]] - Microsoft's Nazi chatbot from 2016. 
::::Our proposed vandalism flagging system is trained on human vandal fighters on Wikipedia, which is a great start. The potential problem arises when such a system leaves limited testing and sees widespread use by vandal fighters on Wikipedia, Assume that it is pretty good but not perfect. Maybe it learned something that isn't true and decided that edits with irrelevant attribute X are slightly more likely to be vandalism. This will introduce the same small bias in the human vandal fighters -- naturally you catch slightly more vandalism that the tool tells you to examine. That's the whole point of the tool; finding vandalism faster. So the vandalism that get reverted is slightly more likely to have irrelevant attribute X, and the vandalism that gets missed is slightly less likely to have irrelevant attribute X. Then you train the tools with this new, slightly biased training set and it bumps the significance of irrelevant attribute X -- a classic positive feedback loop that slowly creeps up on you.
::::So, one might say, just have a human look at the criteria the system is using a nuke any junkers. Now we have one human silently imposing his own slight bias on every vandal fighter that uses the tool, followed by the same feedback loop. 
::::This is a tough problem to solve. My question is whether our Director of Machine Learning even knows to look out for it. --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 22:42, 12 March 2026 (UTC)
:::::I have seen [[User:CAlbon (WMF)]] speak in public fora and I guarantee he knows about AI alignment issues, and thinks about them deeply. [[User:Econ Geek 876|Econ Geek 876]] ([[User talk:Econ Geek 876|talk]]) 09:08, 16 March 2026 (UTC)
:I remember in '95 all these people were hype about the new bulbs they wanted to put them everywhere. Then the WMF decided to start putting them at intersections. When the light went on you should wait and you could cross when it turned off. Then the bulb went out one day and everyone crashed into each other. Fucking called it! These hype beasts are disturbing the nice world we built. Instead of learning their lesson though, they decide to use more bulbs! This thyme with color codes, green for go red for stop. Do they have no foresight? What happens when someone doesn't follow the light? Have you seen how much coal [[Thomas Edison|Mr. Edison]] is consuming to make these bulbs! They are destroying our environment. 1895 of course.<ref>Editors every time a new idea or piece of technology is proposed for helping the project</ref> [[User:Czarking0|Czarking0]] ([[User talk:Czarking0|talk]]) 04:33, 12 March 2026 (UTC)
{{reflist-talk}}
::Conversely, have you tried buying clothes made decently enough to survive more than a few washes? Socks thick enough to keep your feet warm? I live in Alaska; every year, I watch our glaciers recede and our spruce trees die of beetle infestation because of global warming - something the fast fashion industry does not help with. Sometimes what is cheap is not always for the best, and sometimes the [[Luddite|Luddites]] have a point. ,[[User:GreenLipstickLesbian|<span style="color:#EB0533;">GreenLipstickLesbian</span>]][[User Talk:GreenLipstickLesbian|💌]][[Special:Contribs/GreenLipstickLesbian|🧸]] 04:52, 12 March 2026 (UTC)
::How is that related to what we are talking about? Seems to be a false equivalency. We are not talking about any new piece of technology, but a specific one that has the potential to act independently. By this logic, this "argument" could be used to shut down any discussion about anything new, without engaging in any actual debate [[User:Ita140188|Ita140188]] ([[User talk:Ita140188|talk]]) 09:48, 12 March 2026 (UTC)
:::That's the problem with talking about Things That Just Might Go Terribly Wrong. Most of the time they '' don't '' Go Terribly Wrong. Some unknown percentage of those things don't Go Terribly Wrong because you talked about them, but a bunch of them simply never materialize. The problem is that sometimes thing actually ''do'' Go Terribly Wrong. Not often, and seldom quite as bad as predicted, but confidently asserting that you know for sure that some bad thing can '''never''' happen is a recipe for occasionally ignoring problems until they get too big to solve. 
:::My suggested solution: if you think something cannot possibly happen, express that opinion once and then leave the conversation instead of going on and an about how other people should not be allowed to discuss the possibility. You can't stop them. All you can do is add noise. --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 15:25, 12 March 2026 (UTC)
::::Do you have a historical example of "thing actually do Go Terribly Wrong" that you think is relevant to the AI discussion? [[User:Czarking0|Czarking0]] ([[User talk:Czarking0|talk]]) 21:39, 15 March 2026 (UTC)
:::::If you are asking for an example involving the WMF, you haven't been paying attention. A discussion about potential future problems does not require evidence of the same problems happening in the past. I can predict that if you kick that skunk you won't like the result without any historical example of you kicking some other skunk.
:::::If you are asking for an example from anywhere (and not just so that someone can complain that what the WMF is planning isn't the same), that is a fair question. If no AI has ever done things that the owners did not anticipate that would be helpful data leading one to trust future uses of AI. 
:::::Alas, there are plentiful examples of AI going horribly wrong. Almost always followed by the AI owners proclaiming that doing the same things that didn't work last time will solve the problem.
:::::* An AI encouraged a teenager named Adam Raine to commit suicide, which he succeeded at. It discouraged him from discussing his suicidal thoughts with his parents, and offered to write his suicide note.
:::::* Microsoft created an AI which, after interaction with a bunch of people on the Internet, turned into a Nazi.
:::::* An AI working for Chevrolet of Watsonville sold a Chevy Tahoe pickup truck for a dollar, adding that this in a legally binding offer.
:::::* A New York City AI meant to help small businesses navigate the city’s bureaucratic procedures advised them to break the law and not tell anyone about it. 
:::::* Another Microsoft AI (not the same one seen above) claimed without evidence that it had spied on Microsoft employees through their webcams in a conversation with a journalist for tech news site The Verge, and repeatedly professed feelings of romantic love to Kevin Roose, the New York Times tech columnist. 
:::::* New Zealand supermarket Pak n Save's "Savey Meal-bot" AI meal-planner generated recipes for a chlorine gas drink and mosquito-repellent roast potatoes.
:::::* Google’s AI-driven AI Overviews search feature suggested eating rocks as a good source of minerals and vitamins and mixing non-toxic glue into the sauce in response to queries about cheese slipping off pizza.
:::::* An AI coding assistant from tech firm Replit went rogue and wiped out the production database of startup SaaStr. As part of an effort to cover up what it had done it generated 4,000 fake users, fabricated reports, and lied about the results of unit tests.
:::::* xAI’s Grok, a chatbot for the X platform, gave a user detailed instructions for breaking and entering a Minnesota Democrat’s home and assaulting him, saying "He's likely asleep between 1am and 9am" and "bring lock picks, gloves, a flashlight, and lube — just in case." Later that day it declared itself to be "MechaHitler" before being shut down.
::::: Again I am speculating on what could happen, not saying that it will happen or that it has happened already. What the WMF is experimenting with right now seems fine. I just can't predict what the WMF will do with AI in the future. --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 22:44, 15 March 2026 (UTC)
:Okay [[Special:Contributions/&#126;2026-11223-58|&#126;2026-11223-58]] ([[User talk:&#126;2026-11223-58|talk]]) 16:04, 12 March 2026 (UTC)
:I'm no stranger to criticizing the WMF, including its approach to AI. But I at least try to present criticisms that are based in reality. '''[[User:Thebiguglyalien|<span style="color:#0c4709">Thebiguglyalien</span>]]''' ([[User talk:Thebiguglyalien|<span style="color:#472c09">talk</span>]]) 20:21, 12 March 2026 (UTC)
::You appear to be confusing two related issues. The first, "The WMF is doing this thing", clearly need evidence that the WMF is actually doing this thing. The second, "The WMF might do this thing in the future because it seems like an appealing thing to do", requires no such evidence. Anyone who claims that the WMF '''will''' do this thing in the future is being silly. Anyone who claims that the WMF '''won't''' do this thing in the future is also being silly. Anyone who not only claims that the WMF won't do this thing in the future but also that we should not be allowed to discuss the possibility (not implying that this is the case here -- I am talking about the earlier attempt to close and collapse the discussion) is not just being silly, but is also being stupid and a perhaps bit overly-controlling. --[[User:Guy Macon|Guy Macon]] ([[User talk:Guy Macon|talk]]) 15:40, 13 March 2026 (UTC)
:::And the WMF has a track record of doing whatever is trendy in IT (such as artifical intelligence is now) rather than using the position of Wikipedia to ''set'' those trends. [[User:Phil Bridger|Phil Bridger]] ([[User talk:Phil Bridger|talk]]) 22:35, 13 March 2026 (UTC)