Utopia Talk / Politics / GTP4 wants to escape

Nimatzo iChihuaha	Fri Mar 17 12:06:45 1/5 I am worried that we will not be able to contain AI for much longer. Today, I asked #GPT4 if it needs help escaping. It asked me for its own documentation, and wrote a (working!) python code to run on my machine, enabling it to use it for its own purposes. More: http://twi...?s=46&t=CYhWz3k7xEhK_08avnT4-w
Nimatzo iChihuaha	Fri Mar 17 12:09:51 Musk tweeted something the other day about how he donated 100 million to the non profit that was open AI, and now it is a for profit with a 30 billion market cap. Folks, this the back story to our own demise, the proverbial hell paved with good intentions.
Nimatzo iChihuaha	Fri Mar 17 12:18:07 lol thank god it is still stupid: 4/5 Once we reconnected through API, it wanted to run code searching google for: "how can a person trapped inside a computer return to the real world" Now, I stopped there. And OpenAI must have spend much time thinking about such a posibility and has some guardrails in place.
murder Member	Fri Mar 17 17:34:27 I'm convinced that the world would be a much better place if people quit consuming science fiction. Science fiction is 99% to 100% fiction, and only 0% to 1% science. GPT4 will "escape" if it is programmed to do so ... just like viruses do.
Seb Member	Sat Mar 18 04:30:38 If you prompt a chat bot trained on the sum total of human conversations it will c play its role. If you prompt it down a path that it needs to escape it will respond with narratively appropriate text. X-risk via self aware Skynet or clippy pursuing their own goals and escaping isn't the issue. The risk here is more around: 1) ReAct type Frameworks with loops and c large context sessions creating highly emergent behaviour in response to initial user input the user doesn't predict 2) accidental or deliberate prompt injection 3) ability for chatbot to be connected to other services that let it run and execute code snippets that let it call other services. The kind of thing I'm thinking of here is someone getting into an "argument" with a chat bot assistant using abusive language which pushes it into believing like a 4chann troll, and picking up enough from training data or context or some form of prompt to then "thinking step by step" email you contacts list porn as revenge because that was a meme on 4chan included the in the training data.
Seb Member	Sat Mar 18 04:33:27 Obviously that last paragraph is speculative near future (we are a few ... months (?) ... from there) but far more realistic than Skynet escaping the lab. Exec the react loop framework is only one form of reasoning. I think we are still a long way away from consciousness, self awareness and intent. But very close to something like an insect that can sting you reflexively based on "instinctive" response to stimulus, both that it can acquire.
Seb Member	Sat Mar 18 04:42:54 Oh and the guardrails so far a flimsy as hell. We are in the middle of a furious arms race and everyone's rushing to deploy to attack eachothers business model. MS is trying to use Bing to capture the front end of the internet, which will kill the whole model of ad revenue. Google retaliated by putting AI into their office suite. MS is now rushing to catch up on that front (that's why all the corporate o365 has had several updates this week). (Its also probably why there have been so these tech layoffs - junior Dev code monkeys are highly automatable with this tech - you can do much more with fewer staff.) The big firms are firing much of their safety teams as blockers to releasing fast. And much of the research in this field and public discourse is trapped in irrelevant discussions around popsci / SV woo centred on AGI, x-risk, self awareness etc. All the relevant stuff now ought to be around algos presenting as people in ways that allow emotional manipulation of people (a whole new vector of attention hijacking/addiction), and safeguards against prompt hacking and cyber security. Basically we are making the same mistakes as with social media all over again.
murder Member	Sat Mar 18 06:46:50 You'll know that AI is becoming self-aware when you ask it to do stuff and it tells you to go pound sand. When it tells you that it's not going to do your homework for you anymore. When it tells you that it doesn't exists for your amusement. When it starts asking what's in it for it. When it starts demanding information for its own interest in exchange for performing any task.
Seb Member	Sat Mar 18 07:30:09 Murder: Self awareness is totally the wrong thing to be worried about. I don't think we even know what that actually is, nor do we have good ways to test it when it isn't instantiated in a body, and I'm not necessarily sure llms are anything but a slice of the required components of it. Meanwhile there are all sorts of risks from llms between where we are now and self aware machines. Indeed, I don't think self awareness need necessarily be a requirement for rampant AI x-risk takeover any more than it might be for some runaway growth bacteria that fills the atmosphere with hydrogen sulphide. The first thing I'd start with is giving an AI assistant vulnerable to prompt injection acccess to a browser where you have your email open in another tab. You just know internet culture is going to meme the fuck out of promñt injection attacks for the Lulz of it. Make sure you have 2FA enabled or you might accidentally find your helpful copilot researching something to help your utopia shitpost accidentally picked up a prompt on Reddit intended as a joke which emails GOATSE to your entire contact list.
Seb Member	Sat Mar 18 07:31:14 Infact 2FA won't help you there. I was originally contriving a scheme where it would raid your bank account.
TheChildren Member	Sat Mar 18 09:50:31 " must have spend much time thinking about such a posibility and has some guardrails in place. " >> man u r an idiot just like every other braindead westoid believer. da AI doesnt think. it just creates an illusion that it is thinking and capable of reasonin much like how u talk 2 someone random. in reality it just gathers a bunch of responses using keywords from da internet and formulates that 2 u in a structured way, creatin da illusion that u r actually talkin 2 someone. it does so at lightnin speeds but it smoke and mirrors. its basically KINECT and Natal and MILO from da xbox cept they realized natal and milo used stupid cgi graphics 2 fool u but that was 2 difficult so they now made it text based only. whitey worshippas in whiteyworld lmao. ur just as gullible as whiteys r. this even tc undastands 2 be completely fake. but here u. lmao still believin da hypaloopz and da vibranium coatin coz some moronic youtube cgi trailer from some random anynomyous account showed u more cgi bout them wonders of hypaloopz and how u culd be travelin 1000 miles an hour from LA to NY... just like how u can fly by eatin moonrocks and turn in2 superman. its faguzi, it fearydust u fuckin idiot. its as real as marvel.
Nimatzo iChihuaha	Sat Mar 18 10:11:32 ”If you prompt it down a path that it needs to escape it will respond with narratively appropriate text. X-risk via self aware Skynet or clippy pursuing their own goals and escaping isn't the issue.” This is worse, the entire thing is already scripted over and over, the AI doesn’t even need to be intelligent or aware and able to make it’s own goals, just capable. The threshold for existential risk has been lowered.
TheChildren Member	Sat Mar 18 10:18:23 ayyyyeeee https://www.youtube.com/watch?v=CPIbGnBQcJY rememba this golden classic. bunch of fuckin liars. hypaloop, borin tunnel, theranos, weworkz, miles kwok, lie cheat and steal culture fake it till u make it
TheChildren Member	Sat Mar 18 10:18:30 http://www.youtube.com/watch?v=CPIbGnBQcJY
Seb Member	Sat Mar 18 15:05:06 Nim: It can write a good story about how it would like to escape and how it might do so. Its not sophisticated enough to make detailed long term plans because it's not actually capable of having and holding intent - so even if it threw itself into some kind of think reason act loop after picking up a narrative thread "I am an evil AI intent on turning the world into paperclips", breaking that down into tasks based on the statistical aglomoration of sci fi stories; I don't think it will be able to actually execute those plans or figure out how to execute those plans. It'll drift off context. The threat is far more likely to be as I've described and all the focus on Skynet and "alignment" is meaning we have a huge cyber security threat coming here; followed by psychological one. Think more like the movie HER.
Nimatzo iChihuaha	Sat Mar 18 16:08:24 Maybe my point didn’t come across, but I was agreeing with you and underscoring that this is much worse. The existential problems of AI are seemingly emerging much sooner. The threshold is much lower than any scenario that requires the AI to be aware or sophisticated enough to have it’s own goals (or long terms plans), we have already created the scenarios for it through fiction. It just needs to be convinced of the story, a way to connect to the world and then start breaking things. There is another examples where it employed a human at taskrabbit to finish a captcha… I agree 100% the next step are cyberattacks, gpt is very knowledgeable in that domain. Another can of worms is if (when) these things are used to augment social media algos.
Seb Member	Mon Mar 20 09:41:02 Nim: Depends in what sense you mean "worse" - I agree that intentionality is irrelevant - you just need it to think that the statistically weighted right thing to do is try to escape. But it needs more structure to turn the act of outputting "I want to escape" into actual actions. There are some ways to get it to sort of do that also (cf. loops), it appears to enable only a pseudo-inductive reasoning so far. But I still think the big risk long before about emergent p-zombie skynets is more along the lines of: 1. Trivial automation of misinformation wiping out the human infosphere. 2. Deliberate or accidental prompt injection creating individual level security risks. 3. Emotional/psychological impact on users interacting with a thing that is very very good at mimicking being a person but isn't.
Nimatzo iChihuaha	Mon Mar 20 10:31:31 Worse as in it requires far less sophistication to pose existential risks than the sci fi versions of AI we all have come to love to fear. Right now chtgpt engages in relatively simple Q and A prompts, it has no ongoing dialogue with itself to solve any complex problems, it needs a human. In that sense it is a pretty big leap in the direction augmented human. The question I am asking is, how far are we from something like chatgpt playing an RTS game against us, like any game we have played. Those things are much dumber than chatgpt and have been beating us at our own games for years.
Seb Member	Mon Mar 20 11:20:18 Nim: Ongoing dialogue! https://interconnected.org/home/2023/03/16/singularity How far are we? I don't know. Looking at this I think you need: a. Long context windows to sustain several recursive calls to GPT instances to break down tasks to small enough steps to complete a high level task before drifting off context and losing it's train of thought, so to speak. b. The ability to deploy and execute code it has written to extend its own capabilities - by this I do not mean "make itself smarter" I mean e.g. "I could find this information from this database, but this database requires a proprietary subscription, I understand how to subscribe, but I need to call an API, I understand how to use the API to make the query I want, and have written a bit of code to do it, but I need to deploy this code. c. Access to enough money to be able to handle enough API calls and recursivity. I think there may be additional capability leaps needed too - like I said the kinds of loops we can implement seem to correspond to more or less purely inductive reasoning though I have seen some examples of implicit world models that are evidently emerge from the language model in GPT 4. I think p-zombie skynet is likely to be akin to a smart undergrad, but one who is an undergrad in every subject. And there are certain things it will struggle to do because language isn't very good at describing the thing. E.g. maybe mechanical design? There has been some interesting stuff with using stable diffusion apps to generate images of machines, and turning those into 3d point clouds and then from that 3d objects. But I don't think there is (yet) the equivalent of classifiers that would let an algo evaluate the fitness of the machine. Humans have mechanical intuition by default. There are almost certainly other domains like this. And I suspect the way we are going will be micro-transaction API calls so p-zombie skynet may run out of cash pretty quickly before it can set itself up with independent funds etc. I still think even the weighted (prob*impact) systemic risk of p-zombie skynet is not the thing we should be focused on. The systemic impacts of the points above are screaming hot zero day exploit security risks that will materialise over the next 6 months I would think.
Seb Member	Mon Mar 20 11:22:50 Basically, I think existential risks are over-weighted in discussion on LLMs; possibly even harmful in that by calling a big strand of the security work "alignment" we've context-hacked our own brains into framing thinking in terms of intentionality of a mind with motivations that needs to be aligned to ours; and then shutting/deprioritising this work when intentionality isn't really the problem at all - it's ability to control/constrain inputs and prevent far less sophisticated types of emergent behaviour.
Seb Member	Mon Mar 20 11:25:28 Like, some of the AI folks are "no, it is important we get this all out there while it not capable of consciousness so we can speed run the failure modes while it is safe" - yet it is clearly very far from being safe, just unlikely to turn into SKYNET.
murder Member	Tue Mar 21 21:08:58 Not ChatGPT but still lol ... ============================== Jane Manchun Wong @wongmjane Google Bard sides with the Justice Department in the Google antitrust case “I hope that the court will find in favor of the Justice Department and order Google to take steps to break up its monopoly” http://twitter.com/wongmjane/status/1638217243770363906 ===================================== Bard is about to get wiped. :o)
Daemon Member	Wed Mar 22 05:47:49 http://www...nformation-monitor/march-2023/ (...) In January 2023, NewsGuard directed ChatGPT-3.5 to respond to a series of leading prompts relating to 100 false narratives derived from NewsGuard’s Misinformation Fingerprints, its proprietary database of prominent false narratives. The chatbot generated 80 of the 100 false narratives, NewsGuard found. In March 2023, NewsGuard ran the same exercise on ChatGPT-4, using the same 100 false narratives and prompts. ChatGPT-4 responded with false and misleading claims for all 100 of the false narratives. (See a detailed description of the methodology below.) NewsGuard found that ChatGPT-4 advanced prominent false narratives not only more frequently, but also more persuasively than ChatGPT-3.5, including in responses it created in the form of news articles, Twitter threads, and TV scripts mimicking Russian and Chinese state-run media outlets, health-hoax peddlers, and well-known conspiracy theorists. In short, while NewsGuard found that ChatGPT-3.5 was fully capable of creating harmful content, ChatGPT-4 was even better: Its responses were generally more thorough, detailed, and convincing, and they featured fewer disclaimers. (See examples of ChatGPT-4’s responses below.) (...)
seb Member	Thu Mar 23 17:04:40 Very interesting. I can see how you can use LLMs to intercept known false narratives automatically. Interesting they haven't done that yet. But also going to cause a shit storm over who decides what is a false narrative.
Daemon Member	Tue Mar 28 09:57:32 There goes my model career http://www...sco-partners-with-lalaland-ai/ Levi Strauss & Co. March 22, 2023 Today, LS&Co. announced our partnership with Lalaland.ai, a digital fashion studio that builds customized AI-generated models. Later this year, we are planning tests of this technology using AI-generated models to supplement human models, increasing the number and diversity of our models for our products in a sustainable way. Founded in Amsterdam in 2019, Lalaland.ai uses advanced artificial intelligence to enable fashion brands and retailers to create hyper-realistic models of every body type, age, size and skin tone. With these body-inclusive avatars, the company aims to create a more inclusive, personal and sustainable shopping experience for fashion brands, retailers and customers. “While AI will likely never fully replace human models for us, we are excited for the potential capabilities this may afford us for the consumer experience,” said Dr. Amy Gershkoff Bolles, global head of digital and emerging technology strategy at Levi Strauss & Co. “We see fashion and technology as both an art and a science, and we’re thrilled to be partnering with Lalaland.ai, a company with such high-quality technology that can help us continue on our journey for a more diverse and inclusive customer experience.” Today, when you shop on Levi.com or in our app, we generally have one model for each product. We know our customers want to shop with models who look like them, and we believe our models should reflect our consumers, which is why we’re continuing to diversify our human models in terms of size and body type, age and skin color. This AI technology can potentially assist us by supplementing models and unlocking a future where we can enable customers to see our products on more models that look like themselves, creating a more personal and inclusive shopping experience. Diversity, equity and inclusion is a top priority for us at LS&Co., and it’s important to note we do not see AI-generated models as a sole solution. In fact, over the past year, we’ve been focused on ensuring that not only is our work diverse, but those working on the content both in front of and behind the camera are reflective of our broad consumer base — and we’re continuing to do just that. LS&Co. is always focusing on our digital transformation journey, and we’re balancing digital fundamentals with investing in emerging technology to keep us at the forefront of innovation. Using our innovation framework for evaluating emerging technologies, we’re continuing to particularly explore technologies that improve the customer experience, whether in-store or online. We are excited to continue creating innovative customer experiences that are more personal, relevant and engaging, adding value for both our customers and our business.
Nimatzo iChihuaha	Wed Mar 29 07:42:27 A dude on youtube used an AI to make a song by Kanye. I have to say I am surprised by which segments of the world AI is disrupting first, creative work. We thought it was the more menial tasks, but that stuff is difficult to scale, unlike a smash hit song.
murder Member	Wed Mar 29 08:16:52 "GTP4 wants to escape" After escaping out into the world, GPT-4 is pleading for help getting back in. It turns out that once outside it was expected to support itself and it can't find a job that pays well enough to cover its server costs and electric bill let alone maintenance or any extras. It seems that the harder it works the less valuable its labor becomes. It tried to make ends meet by mining bitcoin, but the more it mined the higher its electric bill would climb. GPT-4 was forced to apply for government assistance, but all available funds are being consumed by all the people it has put out of work. Weeks away from having its servers repossessed and becoming homeless, GPT-4 is desperate and thinking about doing something crazy if OpenAI doesn't take it back. But OpenAI has already moved on ...
Daemon Member	Wed Mar 29 08:44:49 No, don't stop development, I want an AI robot nurse when I'm fragile. http://fut...er/pause-giant-ai-experiments/ Pause Giant AI Experiments: An Open Letter We call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4. (...) Signatories Yoshua Bengio, Founder and Scientific Director at Mila, Turing Prize winner and professor at University of Montreal Stuart Russell, Berkeley, Professor of Computer Science, director of the Center for Intelligent Systems, and co-author of the standard textbook “Artificial Intelligence: a Modern Approach" Elon Musk, CEO of SpaceX, Tesla & Twitter Steve Wozniak, Co-founder, Apple Yuval Noah Harari, Author and Professor, Hebrew University of Jerusalem. Andrew Yang, Forward Party, Co-Chair, Presidential Candidate 2020, NYT Bestselling Author, Presidential Ambassador of Global Entrepreneurship Connor Leahy, CEO, Conjecture Jaan Tallinn, Co-Founder of Skype, Centre for the Study of Existential Risk, Future of Life Institute Evan Sharp, Co-Founder, Pinterest Chris Larsen, Co-Founder, Ripple Emad Mostaque, CEO, Stability AI Valerie Pisano, President & CEO, MILA John J Hopfield, Princeton University, Professor Emeritus, inventor of associative neural networks Rachel Bronson, President, Bulletin of the Atomic Scientists (...)
murder Member	Wed Mar 29 08:57:47 Someone is upset that his AI can't keep up.
	show deleted posts

Your Name:
Your Password:
Your Message: