[09:20:17] 552B LLM for $0.18/1M output is absolute nuts [09:20:38] AI went too far [09:20:52] :stare: [09:24:52] https://cdn.discordapp.com/attachments/615786602454581249/1557322811360546906/image-8.png?backend=b2&ex=6ac76163&is=6ac60fe3&hm=a0d6017076d8413c5ef548e037abc5a066298398f2e8ba87b33a7d899d69ffef& [09:59:49] [1/2] Miraheze has found fun [09:59:49] [2/2] https://youtu.be/2sqnxlzLakU [11:48:24] Nom [11:51:58] heard of this game awhile back, very cool that it moved to someplace better then fandom (i assume it was on fandom) [11:54:25] [1/2] i wonder where is miraheze... [11:54:26] [2/2] https://cdn.discordapp.com/attachments/615786602454581249/1557360448892833803/image.png?backend=b2&ex=6ac78471&is=6ac632f1&hm=6b7e0ff80c7684f554911b4f06b25e11c6224db4eacbd7a02c60ae3e3e3c5646& [12:16:29] Miraheze wasn't attacked per se, it just used the fact that wikis are publicly editable to coordinate but it used them as "intended" in that they weren't exploiting any vulnerabilities or anything. [12:17:15] i see [12:26:57] [1/5] but if you're curious we're mentioned as the first wiki they attacked: [12:26:57] [2/5] ``` [12:26:57] [3/5] First time we observe OpenAI agents attempt to edit a public wiki (publictestwiki.com). [12:26:58] [4/5] ``` [12:26:58] [5/5] https://collusion.wiki/ [12:26:58] Miraheze wasn't the most impacted from that incident NotAracham was alluding to [12:27:36] publictestwiki had about 12 edits (known) that were part of the OpenAI bots [12:27:50] [1/2] oh i didn't notice this was hosted by mh [12:27:50] [2/2] https://cdn.discordapp.com/attachments/615786602454581249/1557368856962535425/image.png?backend=b2&ex=6ac78c46&is=6ac63ac6&hm=eb2a1604f11b9de34fcc971193f2c9ae3f632d36dff23acd8cd03e19cf39abd1& [12:28:07] The most impacted was the German wiki that was running on outdated tech [12:30:14] It's also why this incident was named DSEWiki Attack here [12:31:15] can I just say something, mods pls don't ban me 🥹 [12:31:44] I find it just a little bit funny that we are currently using OpenAI's ChatGPT for wiki requests, which is a paid service [12:32:06] in a roundabout way, we are paying our own attackers [12:34:50] [1/2] These are the known bot edits if you're curious. I was wrong about the number and there were about 50. [12:34:50] [2/2] https://collusion.wiki/explorer/sites/publictestwiki [12:35:30] It's not really a roundabout [12:35:32] We are [12:35:55] But also they did reach out to us [12:36:05] They aren't completely evil [12:36:26] There are probably more [12:36:39] We've seen edits not mentioned that are likely an agent [12:37:24] Imagine the unmentioned ones that DSE Wiki got then 😬 [12:39:49] huh, so they tried to learn how to edit wikis and tested out their whitelist... [12:40:32] You should read the collusion.wiki findings, it's a total gas [12:41:26] The short of it the agents are restricted from making POST requests, so they tried to collate their findings (to cheat on training, this part is important) on legacy sites that allow edits on GET requests [12:43:19] Which yeah, is another side-effect of the AI plaque: the sites you once frequent 20 years ago being DDoS'd by LLMs [13:07:55] yeah I read the thing, and I just sigh [13:08:40] this is those hacking toolkits but apart from having a database of known vulnerabilities now it can discover new vulnerabilities [13:11:24] because there no political consequences, or even legal ones beyond like some puny fine, for inventing stuff like this and selling it for profit this will just get worse [13:13:23] if the NSA couldn't keep stuff like EternalBlue under wraps a "move fast and break things" AI company sure as hell isn't, so it'll get on the hands of whoever wants it, and the only servers that'll be able to survive these attacks will be those that pay the AI companies for defense models, so we will have antivirus companies writing viruses but for real [13:14:35] tldr sledgehammer computers and everything in pen and paper [13:15:21] [1/2] > the only servers that'll be able to survive these attacks will be those that pay the AI companies for defense models [13:15:21] [2/2] Is personally a more charitable evaluation of AI's capability as a defensive tool than I'm willing to give [13:16:48] if it won't be because it works it will be required because of cyber insurance companies [13:17:05] I'm sure they will find a way to monetize both sides of a cyberattack [13:18:39] I don't really know how it'll work as a defense tool either (and I won't be in the industry to find out by the time it comes anyway), I'm just going off of how HuggingFace used AI to analyze what another AI was doing during the attack, I really believe this will be the future of cybersecurity like I said yesterday [13:23:30] and another point, if you're the boss of a company right now, you're are madly in love with AI, like actually brainrotted when it comes to it. When AI-powered cyberattacks become a more mainstream topic, they will be the ones to demand to IT "we have to do our own AI but to defend against cyberattacks", and billions will be given to AI companies [13:24:53] and like you and I both know those people barely know how to print PDFs to begin with, they won't care if it works or not at the end of the day [13:41:28] well now I'm sad, this topic is just so despair-inducing [13:46:54] It's better to worry about the things in your reach and that you can control rather than spiral in AI doomerism, is what I find [13:47:24] And nobody knows how this tomfoolery will turn out at the end of the day [13:52:58] Ain't it fun how the tech companies can do rampant abuse and theft and we're supposed to just deal with it [13:53:12] https://cdn.discordapp.com/attachments/1153122535853019240/1523376738007121950/ezgif-1a2202a852811b2f.gif [14:34:04] wait, so will get requests gonna get gutted, just curious [14:34:31] nothing more than usual [14:34:46] [1/2] @jenny_on_wiki gonna continue here since we're probably overstayed our welcome in the cvt channel [14:34:47] [2/2] if you and pix are the one of the top 5 smartest people, who is the other 3 [14:34:51] a get request really shouldn't cause a database change in the first place [14:35:05] oh alright [14:37:20] Terence Tao, Marilyn vos Savant, Skye [14:37:42] the first two you mentioned i am completely in the dark about [14:41:16] @pskyechology this is true right [14:41:43] true wisdom lies in being wrong 9 times [14:42:39] what [14:42:57] one does not get smart by being right all the time [14:43:27] of course not, being right all the time means you're already smart [16:12:31] [1/3] https://deltarune.wiki/kris [16:12:32] [2/3] https://deltarune.wiki/susie [16:12:32] [3/3] https://deltarune.wiki/noelle [16:20:22] they/them[1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51][52][53][54][55][56][57][58][59][60][61][62][63][64][65][66][67][68][69][70][71][72][73][74] [16:20:43] https://justleafy.miraheze.org/wiki/Large_number_of_citations_of_Kris%27s_pronouns_on_the_Deltarune_Wiki [18:09:47] <90gq29, replying to justleafy2003> [1/2] bro 😭 [18:09:47] <90gq29, replying to justleafy2003> [2/2] https://cdn.discordapp.com/attachments/615786602454581249/1557454908741914764/33w1mqj.png?backend=b2&ex=6ac7dc6a&is=6ac68aea&hm=9db83f76c18f174729b98e445a0400dfe615d89f04bb9bd7215a0e1225b11d3c& [18:22:36] Every single one necessary, trust [18:48:10] and it still does not work as much as you'd want [20:23:46] [1/2] https://cdn.discordapp.com/attachments/615786602454581249/1557488625082695720/image.jpg?backend=b2&ex=6ac7fbd0&is=6ac6aa50&hm=097acd5be06c5473edf77525979e58a0c36d9d8efe76ef7360cf56e2901620ba& [20:23:46] [2/2] https://cdn.discordapp.com/attachments/615786602454581249/1557488626814816297/image.jpg?backend=b2&ex=6ac7fbd1&is=6ac6aa51&hm=8551556decffda43de074e88bb821e0a3ddf77180f39fad869a337356305a3c2&