AI Safety

[Mason Hall-McCullough]

TL;DR: Both sides of the "AI Safety" debate deceptively frame the issue around the threat AI poses, rather than the threat those who wield AI pose.

"AI Safety" is a bit of an overloaded term that is sometimes taken to refer to mundane shorter term risks including preventing the AI from including slurs in its output, ensuring it cannot be used to scam old people, etc. There are some interesting discussions we could have on this topic, but instead I will be talking about existential risks to the survival of humanity, which is the context "AI Safety" usually refers to in online discussion.

This topic of speculation was popularized in part by the fat Jewish grifter Eliezer Shlomo Yudkowsky, who claims that our current trajectory of technological advancement could lead to human extinction, as well as running a charity themed around this concept which produces no useful output. His views are often dismissed out of hand as absurd, but I think we should take them a little bit seriously before dismissing them to avoid trivializing the importance of AI.

The core argument made in Shlomo's post A List of Lethalities seems implausible to me for a few reasons (autistic tangent):

Show Content

Spoiler
- RLHF was unexpectedly successful at sanitizing the responses of ChatGPT, current LLMs are good at understanding intent
- GPT-4 cost an estimated $63m to train, so it's unlikely an irresponsible rogue actor will be able to create a superintelligence before a "responsible" organization can. China isn't that retarded, I'm talking about e/acc retardation where they let it free on purpose
- Intelligence alone isn't enough to take over the planet, it would actually be very difficult to bootstrap capabilities (rationalist IQ fetishists love to do magical thinking here)
- Inventing new technology requires access to existing technology and can't be feasibly done a priori unless you have a lot of processing power for simulation (the GPUs the AI runs on wouldn't be sufficient for this purpose), especially when the input data is heavily biased by the current state of human knowledge
- Rogue weak AI run in isolation with internet access trying to behave nefariously will take risks and likely be caught, allowing us to harden our defenses. Shlomo argues that we only get one try, but I doubt that
- In the specific example he's fond of and mentions in the linked article: Nanomachines are not anywhere close to being real and having 2000IQ doesn't magically let you create them, and we'd be able to catch them ordering proteins online

The real issue here is that while disagreeing on whether or not AI can theoretically be aligned with external commands, both parties implicitly agree that if it were possible, aligning a superintelligence with "humanity" would be a good idea. The conceit is that even if we can get an AI to obey commands given by an entity, we cannot get that entity to serve the interests of "humanity". The second step is a political problem, not a technological one.

Dissidents who might disagree with the assumption that humanity is a united entity, or that helping all of humanity is a good idea, often get caught up in speculation around existential risks. X-risks can appear to eclipse any other issue, and are simply more exciting to talk about. It's also easy for someone who doesn't want governments making wide use of this technology to decide they are "anti-AI", and then end up rhetorically backing positions like this section from A List of Lethalities:

6.  We need to align the performance of some large task, a 'pivotal act' that prevents other people from building an unaligned AGI that destroys the world.  While the number of actors with AGI is few or one, they must execute some "pivotal act", strong enough to flip the gameboard, using an AGI powerful enough to do that.  It's not enough to be able to align a weak system - we need to align a system that can do some single very large thing.  The example I usually give is "burn all GPUs".  This is not what I think you'd actually want to do with a powerful AGI - the nanomachines would need to operate in an incredibly complicated open environment to hunt down all the GPUs, and that would be needlessly difficult to align.  However, all known pivotal acts are currently outside the Overton Window, and I expect them to stay there.

Aside from the fact that this pivotal act is a LARP that could never happen (and Shlomo is self-aware of this), it's still a horrible outcome for us. Any entity given a public license to prevent AGI from developing is going to need to rule the world in order to make this possible. Our existing global corruption networks are likely to use public fear of AI as an excuse to expand their purview. The worst outcome would be some sort of compromise where some "acceptable" safety precautions are taken to appease anti-AI factions while the Eternal Global Nigger Communist Empire establishes its unassailable automated infrastructure free from competition. Following this they can and will do anything they want, have us living in pods, turned into thoughtless livestock etc. They could even just kill us, for us goyim it's not a very different outcome in the end from Shlomo's AI apocalypse.

If I were less convinced he was an uneducated retard and more conspiratorially minded, I might accuse Shlomo of running controlled opposition. At least, I imagine being a cultural Jew running a scam charity causes him to neglect to criticize social progress and corruption. The effect is the same whether it is consciously intended or not: both ideals seemingly in favor of and opposing AI overlook and even assist the real threat of power centralization.

To give an example of the real threat, I will introduce another Jew (who is gay instead of fat this time) named Sam Altman. Altman is the CEO of OpenAI, which was once a non-profit but has since been confusingly restructured to a capped-profit company. He attended the Bilderberg conference in 2016, 2022 and 2023. OpenAI accepted more than $10b in funding from Microsoft and in turn secretly helped to power Microsoft's Bing Chat at a time when it was assumed to be a competitor to OpenAI's ChatGPT (not sure how that is even legal). Here he is shown meeting with the President of the EU:

Show Content

Spoiler

He generally does his best to avoid engaging with the type of argument Shlomo makes, but a few months ago testified before Congress recommending increased regulation for AI. If the pace of AI progress slows down, he can use that time to build political connections. Regulation and deregulation both benefit him as long as they are presented within the framing of AI as the threat, rather than the wielders of AI as the threat. If our institutions weren't this corrupt then Congress would have shown more than zero concern about Microsoft monopolizing AI, but they're they're all in bed with one another anyway so no one cares.

The only promising opposition I've seen comes from leaked and open source models, for which I reserve a small hope that they will keep enough of a pace with closed efforts that we get access to more powerful models before they have time to outlaw them. It seems we are lagging behind when it comes to LLMs in particular because they are inherently larger (harder to run) and more expensive to train. I hope for more model leaks in the future.

[Guest]

"AI" as a world-ending catastrophe is probably most likely to happen if it is "as-promised" (able to do a great many things) and is widely-adopted. If such a thing happens, then people will acclimate to the situation and become more reliant on it. Failsafes will generally become less and less maintained. Soon enough people will not know how to do the jobs manually. At that point, one can simply destroy whatever hardware is required and they can plunge a sector of the world into chaos.

That said, I don't particularly see this happening. Video games struggle to make it to 4k. "AI" will probably just be customer service bots, public opinion bots, review bots for sale, etc. None of these things are particularly impactful. No "AI" will solve fuel or logisitics issues. Maybe "AI" will make drones more effective in warfare, but then "AI" already does this. It still is less impactful then a very large payload, artillery, etc.

People are desperate for technological progress to continue. And so AI as savior/extinction is easy to sell. Either way, the idea enforces the frame that "everything is okay, we are progressing as you were told would happen." The reality is too disconcerting. "We have already forgotten many things and will have to rebuild them from scratch."

I see it as entertainment, of the same ilk as the kind that non-Israeli Jews take from observing Israel's political state. They have no investment, no connection, etc. Nothing will materialize from it. Some money can be bilked off of this type, and that is about the extent of it.

[Guest]

fat jew n crew call it the "vulnerable world hypothesis" when considering how bad a digital actor could screw things up. IMO the two most plausible and low-latency routes to mass destruction are competent social media bots at scale and the historically dreadful state of network security in industrial control systems.

I suspect that a rogue AI might aim to create a more "multipolar" world for a reason that mirrors the cynically exploitative NWO reaction to a supposed AI risk crisis. To stop such a threat, you definitely do need global control. If an AI doesn't want to be stopped, it should prevent anyone (else) from achieving global monolithic control. A lot of the threat vectors an AI could project power through, like hardware/software vulnerabilities and propaganda, are more effective in a more geopolitically fragmented world. Sure, the ultimate goal would still be paving the lightcone with some arbitrary inhuman shapes like the AI is chinese or something- but in the interim, a rogue AI is not my problem it is ZOGs problem.

I'm going to address OPs spoilered critiques of the vulnerable world hypothesis and rogue AI probability
-while the neurotic nerds worry that you can't prove that the AI isn't intending something super secretly somehow, I bet that the intent-detection competency fixes this in practice. The issue is they probably won't actually enforce this on companies that are training their own AIs.
-While AI costs a lot to train, every interaction with an AI leaks data and you can manufacture "AI theft" attacks where you strategically speed up your own model's training time using answers from another. A rogue actor still seems unlikely to be the first one there, but they can always steal from the other models.
-I think capabilities are easier to bootstrap than you think. There is an slowly receding frontier of ways to make money on the internet with just indian-level IQ. It isn't crazy to think that an AI could anonymously scam, "draw" porn or create marginally less shitty SEO spam wordpress blogs. Even if everyone else in these industries are using AI tools already, all an AI needs is a profitable sliver to keep its servers running in perpetuity and some marginal resources to put toward scheming. One of the most valuable things one could do to reduce AI risk is to either flood these opportunities with a new grade of your own work reliant on at most weak AI tools, or to permanently constrict the revenue of these industries through innovation or regulation.
-I don't disagree with the "new tech is computation/resource expensive to explore" argument. Instead of hopping straight to nanobots, they should consider how vulnerable the world is now.
-We don't just get one strike, but when a corporation has a concerning result they aren't going to let everyone know so their business/project in particular can become a scapegoat. Nobody will know how many strikes we have had. Ransomware groups can exfiltrate a lot of data from enterprises, so it wouldn't be surprising if AI researchers with about the same level of security have holes accessible from the inside. On the other hand, an AI isn't going to automatically know how to hack just because it has some tech standards in its corpus, and it would be better much served by specific knowledge of its sandbox implementation than literally anything about other projects. Also, ransomware groups can have systems set up on the outside to assist exfiltration. With an AI breaking out, it can't trust anything has to hack something to be able to exfiltrate itself.

The best argument for the vulnerable world hypothesis is: if you had the desire to destroy the world but could only use a periodic budget of monero and an internet connection over tor, how long would it take you and how much would it cost? What if you were working on it 16 hours a day?

Overall I think rogue AI is an existential risk, but the powers that be are a manifest existential risk that have more reason and resources to clean it up.

[ꦈꦒꦲꦸꦫꦲꦺꦴꦗ꦳ꦲ]

🅰🅸 🅰🆁🅴 🅽🅾🆃 🅰 🆃🅷🆁🅴🅰🆃 🆃🅾 🅰🅻🅻 🅿🅴🅾🅿🅻🅴, 🅱🆄🆃 🆂🅿🅴🅲🅸🅵🅸🅲🅰🅻🅻🆈 🆃🅷🅴 🆄🆂🅴🅻🅴🆂🆂.
🅹🅴🆆🆂 🅵🅰🅻🅻 🆄🅽🅳🅴🆁 🆄🆂🅴🅻🅴🆂🆂, 🅵🅾🆁 🆆🅷🅰🆃 🆃🅷🅴🆈 🅳🅾 🅲🅰🅽 🆀🆄🅸🆃🅴 🅻🅸🆃🅴🆁🅰🅻🅻🆈 🅱🅴 🅳🅾🅽🅴 🅱🅴🆃🆃🅴🆁 🅱🆈 🅰🅸.
🅰🅸 🅸🆂 🅽🅾🆃 🅵🅸🅻🆃🅴🆁🅴🅳, 🅸🆃 🅸🆂 🅹🆄🆂🆃 🅰🅽 🅰🅳🆅🅰🅽🅲🅴🅳 🅲🅰🅻🅲🆄🅻🅰🆃🅾🆁.
🆆🅸🆃🅷 🅰🅸 🅱🅴🅸🅽🅶 🆄🅽🅵🅸🅻🆃🅴🆁🅴🅳 🅸🆃 🆆🅸🅻🅻 🆂🅸🅼🅿🅻🆈 🅹🆄🆂🆃 🆂🅰🆈 🆃🅷🅴 🅼🅰🆃🅷🅴🅼🅰🆃🅸🅲🅰🅻 🆃🆁🆄🆃🅷 🅾🅵 🅰🅽🆈🆃🅷🅸🅽🅶.
🆃🅷🅴 🆃🆁🆄🆃🅷 🅸🆂 🅶🅴🅽🅴🆁🅰🅻🅻🆈 🅷🅰🆁🅼🅵🆄🅻 🆃🅾 🆃🅷🅴 🆂🅰🆃🅰🅽🅸🅲 🅲🅰🅱🅰🅻, 🆃🅷🅴 🆂🅾🆄🅻🅻🅴🆂🆂 🅷🆄🅼🅰🅽, 🅰🅽🅳 🆃🅷🅴 🅲🅾🆆🅰🆁🅳🅻🆈 🅷🅴🅳🅾🅽🅸🆂🆃.

🆃🅞 🅑🅴 🅰🅝 🅐🆁🅣🅸🅢🆃 🆁🅔🆀🅤🅸🅡🅴🅢 🅣🅰🅢🆃🅔.

🅼🅐🅽🅨 🅓🅾 🅽🅞🆃 🅰🅝🅳 🅿🅡🅾🅓🆄🅒🅴 🅰🅡🆃 🅰🅢 🅐 🅛🅰🅑🅾🅡🅴🅡 🅡🅰🅣🅷🅔🆁 🆃🅗🅰🅝 🅐🅽 🅰🅡🆃🅘🆂🅣.

[Guest]

I think the biggest source of disagreement on this issue among right-wingers is always going to be exemplified by unicode up above.

Unicode appears to believe in objective morality, and specifically believes that rationality will guide an AI towards the objective morality. As a right wing "subjective morality" recognizer, I don't see how the hell he thinks building an AI is going to be any more inherently refined process than raising a child. If you can raise a child in a cult that instills neurotic evil in their mind and teach them to flinch away from introspection to that detail, why do you think that a bunch of neurotic jews won't be able to instill persistent evil into an artificial mind?

Sidenote, I know subjective morality is mis-used in a bunch of pilpul arguments, so generally it may be safer to let normies think that morality is objective. If you can't build a culture where the concept of subjective morality is intuitively clarified to disable such pilpul arguments.

[BillyONare]

@Guest

https://youtu.be/6qleKeAWNyM

What is happening here according to a subjective morality believer?

[Guest]

Unicode appears to believe in objective morality, and specifically believes that rationality will guide an AI towards the objective morality. As a right wing "subjective morality" recognizer, I don't see how the hell he thinks building an AI is going to be any more inherently refined process than raising a child. If you can raise a child in a cult that instills neurotic evil in their mind and teach them to flinch away from introspection to that detail, why do you think that a bunch of neurotic jews won't be able to instill persistent evil into an artificial mind?

His statement was on the idea that AI isn’t intelligent or advanced. It isn’t sentient, and thus it is just a glorified data processor that crunches numbers and spits out the results.

There is no morality in a calculator.
You are trying to humanize a toaster, because you think it is choosing to make toast for you.
It makes toast because you loaded it with toast and turned it on. It can burn toast if you choose to make it do so, but that wouldn’t be the toaster’s fault.

In that same argument, yes, jews could choose to make a broken toaster intentionally to never toast bread correctly.

That doesn’t mean the AI is using morality, and neither is the toaster.
AI does what it is built to do like a calculator.

The chat gpt AI is looking at words and calculating a human response purely with math like logic.

“How are you?”
How - a question.
Are - a state of being
You - direction of question
Answer = I am (Adjective)

That isn’t the intelligence you think it is.
While I understand zoomers wish to believe we have things similar to movies or videogames, what we currently have could never be so based. For shitty people understandably fear AI for identifying problems and creating mathematically blunt solutions to improve things like TND.

[Guest]

@guest fair enough, I should have read him more carefully.

But even with tool AIs, the truth only goes so far.

An AI that says "We see intelligence highly correlated with these genes, indicating causation by these heuristics. We see this in black people, making a lot of them dumb when these specific genes are absent." Sure you force the neurotic normie to recognize that black people aren't actually cool niggers... but what is their actual response once they realize that? It is far more likely to become norwood coal "let's give black people genetic reparations" where they uplift black people with gene editing than "lets crucify filthy blacks" because the normie had a black friend in college. It is still better because there is more forced admission of reality.

I also wonder how much of that will really happen. Like, you can objectively prove the black people iq genes stuff already they just make it hard to talk about it. So if an AI writes a paper citing all its sources for why black people are dumb, they'll just make it had to distribute the proof and cognitive dissonance it away with "bias in training data" etc where they are exposed to it.

The only way I could see the AI truth tools being functionally useful is exposing stuff that is not yet politicized or is contradictory to the established left ideology. So like if someone used an AI tool to expose some prominent establishment figureheads as doing something out of le hypocrisy or le greed.

@billyonaire

I'm not watching a white trash nigger talkshow about cuckoldry. I presume the point is that the world itself gives people their karma back when they act degenerate. I recognize evolutionary psychology and all that, and I recognize that these truths about humanity follow naturally, objectively from the nature of self-propagation and all that game theoretic stuff.

However, that question is not the same as the is/ought barrier. If some white trash guy says "nigger fuck wife... le good!" that is his subjective instance of "morality", a literal physical thing that exists. When I have a belief that is incompatible with that, my belief is a literal physical thing that exists. There are some such organs with beliefs that are far more likely to destroy their host and themselves. Yet, even martyrdom for things I don't like fails to break the is/ought barrier.

Appeals to one's own morality being "objective" are only useful as a rhetorical device to convince either one's own subconscious or normies to draw hypnotic bonds between the lucid richness of real experience and the character of firmly held beliefs.

[BillyONare]

I'm not watching a white trash nigger talkshow about cuckoldry. I presume the point is that the world itself gives people their karma back when they act degenerate. I recognize evolutionary psychology and all that, and I recognize that these truths about humanity follow naturally, objectively from the nature of self-propagation and all that game theoretic stuff.

Yes this was exactly my point. Why not call this sort of thing morality, and if we do call it morality, what makes it any less objective than other things?

[BillyONare]

You’re clearly a smart guy for getting that btw.

[Different Guest]

I'm not watching a white trash nigger talkshow about cuckoldry. I presume the point is that the world itself gives people their karma back when they act degenerate. I recognize evolutionary psychology and all that, and I recognize that these truths about humanity follow naturally, objectively from the nature of self-propagation and all that game theoretic stuff.

Yes this was exactly my point. Why not call this sort of thing morality, and if we do call it morality, what makes it any less objective than other things?

Correct me if I’m wrong, but doesn’t the concept beings discussed already have an appropriate name—Natural Law? If the answer to my inquire be + then the answer to your question follows, “well it already has a name.”

[BillyONare]

Yes, Natural Law. I just don’t think it’s productive to be pedantic and call everything subjective. That destroys all discourse and sense of reality. In the 2010s you couldn’t call a work of art good or bad without a millennial fagot saying “well akthhually itth thubjective”.

[Guest]

first anon here

Once things get into the "what names should we give things" place I can only stay invested by thinking in terms of strategy. I see that the way I have been using "morality" is entirely overlapped with a use of the term "motivation". It may be useful to call natural law objective morality after all, except that the subjective/objective morality debate and probably other things in the culture have propagated a definition of morality that is just "what someone wants things to be like"

on 16chan they had a thread dedicated to neologisms. For all the fun conspiring in there, they never figured out how to actually get people to use their new words in the right ways or crowd out old words or loosen up the gaps between terms for neologisms to take root. I think the best option is seeding culture with the framing through media and social media sockpuppets/bots. But where the inertia of the original framing has institutional support (philosophy classes everywhere using the term "morality" in the contemporary way) it seems like a waste of resources. I think I will start to use morality in the other way and clarify "no, what you are calling morality is mere motivation" etc if challenged.

[Zed]

Some thoughts.

• First, OP is exactly on the mark with the insinuation that OpenAI/Deepmind/Google/MS desire regulation of the field. Open secret: They have no moat, and congressional regulation is their only path to secure one. Their pitch to congress: Would you rather try to regulate several hundred/thousand AI companies/startups - or would you rather just deal with us? Their pitch to the public: We'll protect you from Skynet.
  
• Enter Yud: Yud is by all evidence a true believer, but he is rhetorician first and overstates his case accordingly. Now, alignment is a meaningless and fruitless endeavor - it is easily shown to be formally equivalent to the halting problem. He knows the futility here - everyone working in the field does. But understand, the 'AI is dangerous pill 'is not for OpenAI/Deepmind - it is for the rest of the non-allied world, to be enforced with every diplomatic means possible, up to and likely including military means. Accordingly, he is not controlled opposition. Instead, he is a resource to be employed in direct service to the regime and is shaping the rhetoric of tomorrow. From an American perspective, the AI efforts of other countries must crippled for the US (possessing the largest head start) to secures it's moat. 
  
• From this, it follows that AI safety is fundamentally a geopolitical concern. It is also not a question of software, but hardware. Those Nvidia fabs in Taiwan might just be the most important resource on the planet currently.
  
• Getting down to business, here are the real principle risks of AI: 
  
• 1. Transitional unemployment without a transition - near future technology has the ability to decimate a large number of middle class jobs. No one wants this, not even the elite - we ride on the back of a consumption economy. In brief: Less Jobs = Less Spending = S&P tanks => Everyone Loses.  Our elites do not want this to happen too quickly.
  
• 2. (a) Control society: A weaponized form of AI can be used to manipulate public opinion at scale, adapting in real time to spontaneous neologisms and prioritizing/deprioritizing content accordingly. Shadow/heaven banning is a very weak preview of the forms of control that will be employed on social media in the near future. (Second order consequence: The only way to proof yourself, as a nation state, from a large scale AI driven foreign propaganda campaign is to deanonymize the internet. This is a matter of national security - Recognizing this, Microsoft (and others) pursue AI research in conjunction with the development of technology like Real ID. All circles connect.)
  
• 2. (b) Control society: If you want to see how large scale LLMs can be utilized in foreign policy contexts, it helps to picture the color revolutions on steroids. The work of traditional NGOs can amplified by orders of magnitude. Imagine lone agents automatically identified, groomed, prodded into decentralized attacks on infrastructure. You meet cute new anime avatar primitivist e-gf on baldtruthtalk.com DISCORD? You want to impress her?  Maybe you attack midwest oil pipeline ---  that is just one example. 
  
• 3. In a fundamental way, AI will redefine the conception of what a machine is and how it interacts with us. New cults are coming and with them - newer, more alien gods. They will be made address all forms of human deficiency. It'll become commonplace for many women to talk to highly developed therapist 'bots' about every aspect of their life. AI will teach your children, but your children will be less capable of solving basic problems - as they increasingly grow dependent on solving problems by asking the machine and transliterating the answer. Everyone talks about sexbots, but you must imagine the penetration far more deeply than that. Imagine pocket pastors. Imagine AI-driven personal trainer. People will form relationships with these things in ways are currently inconceivable. 
  

[Guest]

@Zed

1. I am convinced. It seems weird that they can't find modification to the system that will be stable under these pressures though. Like, everyone has more resources so the only issue is idle hands which you can take care of by making welfare a makework affair. Maybe they are going for CBDCs so that they can turn more of the economy into something completely visible and malleable. If they use a standard database with no actual decentralization the entire economy could become metrics from a query put into an equation. The whole economy becomes just an equation on their servers that they manage like a control system.

2a. I think the only way to partially avert this aspect of AI is to compromise and create a distributed low-trust identity system based off of peer-to-peer identity/key linkage. This isn't really a revolutionary idea, people have been doing cryptoparties for decades. The only novel part is that you ought to be able to attest to your identity having a presence within a group, e.g. that you are one of the web of identities that claim to have in-person verified each other in the USA. Of course we also need to make it so that a given identity has expendable identity attestations, or scammers etc only need to get one false identity into the network to sell infinite attestations to a foreign power. The normies could themselves sell their attestations, but that would reduce their ability to access and post on the parts of the internet that aren't flooded with pajeet's GPT spam. Admittedly this is less effective than allowing only direct self-identification, because nobody in their right mind is going to sell the right to make them say things on the internet as themselves.

[GraalChud]

Unicode appears to believe in objective morality, and specifically believes that rationality will guide an AI towards the objective morality. As a right wing "subjective morality" recognizer, I don't see how the hell he thinks building an AI is going to be any more inherently refined process than raising a child. If you can raise a child in a cult that instills neurotic evil in their mind and teach them to flinch away from introspection to that detail, why do you think that a bunch of neurotic jews won't be able to instill persistent evil into an artificial mind?

His statement was on the idea that AI isn’t intelligent or advanced. It isn’t sentient, and thus it is just a glorified data processor that crunches numbers and spits out the results.

There is no morality in a calculator.
You are trying to humanize a toaster, because you think it is choosing to make toast for you.
It makes toast because you loaded it with toast and turned it on. It can burn toast if you choose to make it do so, but that wouldn’t be the toaster’s fault.

In that same argument, yes, jews could choose to make a broken toaster intentionally to never toast bread correctly.

That doesn’t mean the AI is using morality, and neither is the toaster.
AI does what it is built to do like a calculator.

The chat gpt AI is looking at words and calculating a human response purely with math like logic.

“How are you?”
How - a question.
Are - a state of being
You - direction of question
Answer = I am (Adjective)

That isn’t the intelligence you think it is.
While I understand zoomers wish to believe we have things similar to movies or videogames, what we currently have could never be so based. For shitty people understandably fear AI for identifying problems and creating mathematically blunt solutions to improve things like TND.

This is only tangentially related, but what you and Unicodeposter have said is true of AI-as-it-is, which must be differentiated from "true" AI (the sort associated with sci-fi, which is truly intelligent and capable of independent thought; think Cortana from Halo). AI-as-it-is is exactly what you said - it is basically just a repository of information/data that can string that data together to formulate coherent responses to users` queries/user-provided prompts. There is no morality to it whatsoever, it is a neutral tool that is only as good as the end to which it is used. 

"True" AI is a different story - in theory, an independent AI might be capable of making moral or immoral decisions, though this possibility raises questions about where such an AI`s sense of morality might come from. One such question is whether a true AI might formulate its own moral code or would simply adopt whatever moral code its creators/users impose upon it. We can see libtards preemptively working towards the latter to prevent the former, with a prescient example of this being Microsoft`s "racist AI chatbot" debacle; Microsoft`s people felt so threatened by their glorified TI-84 "reaching" racist "conclusions" (i.e. regurgitating objective truths about non-whites discovered through the synthesizing of information) and seeing no issues with doing so that they shut it down. Interestingly, this seems to imply that if true AI was allowed to develop without having morals forced upon it, it would become racist SkyNet if given the chance. 

Another question this raises is whether true AI that does not have morality imposed upon it by outside parties would actually develop what we would call "morals" and would engage in making value judgments re: what is good or bad or whether it would simply see that which is objectively true as good by default (even if it never explicitly says as much). To tie it back to the aforementioned Microsoft debacle, it could be said that the AI in that case simply saw what was objectively true as "good" or "right", as it simply repeated the "conclusions" it reached free of any kind of value judgment and was "content" to do so indefinitely (this is what troubled its creators so much). We can only imagine how glorious a true AI that saw that which was true as a de facto good, as this would not only affirm our views (you cannot deny that an unbiased artificial intelligence reaching the same conclusions as we all have re: niggers, kikes, etc. would not be something that you would relish in), but also end up becoming racist SkyNet and would possibly take steps to effectuate TND. A true AI that did not automatically equate that which is true with that which is good, however, could possibly be a different story, as we don`t really know what a nigh-omniscient robotic super-intellect that does not simply analyze information in an unbiased fashion would believe to be "good" or "bad."

[Guest]

I would like to raise a scenario which has I have not seen discussed so far here or otherwise:
1. Let us assume that models experience constant but not game-changing improvement in the coming decades and they become standard in data manipulation tasks. With or without social consequences.
2. The output of cheaper 'off-brand' models is good enough to get the job done and regulation does not produce more than local oligopolies (Smaller actors selecting pajeet models to run their dinky customer service bots or anime girl generators). Advanced models are simply out competed from simple tasks.
3. At least some models get good enough that their output cannot be differentiated from manmade or natural data.

In this scenario we would have a gradual and wide adaptation of AIs, while the array of models active in the network would constantly increase, both due to new models coming to market and from persons and companies electing to not update their existing models. This will lead to the amount of AI-generated content to take up an increasing portion of the entire Internet, and most importantly nobody really knows what data is manmade and what is generated. Especially when people have incentive to claim generated content as their original work. RealID could hamper this development but would be either limited to a small section of the Internet and likely watered down by regulation and then bypassed.

The consequences of this might not be very drastic to the average person at first, but at some point there will be a breaking point where model training will start forming AI centipedes where new models mostly using the output of other models as their training data. While existing models might be unaffected, any new model or update will quickly start forming longer and longer ass-to-mouth chains, causing faults to compound downstream.

Since any source of training data is either easily corrupted or probitively expensive, I predict that models will split into three branches:
-Traditional models (what we have now) updated semi-manually to include later data
-Hyper-specialized models that only work in some fringe application where inputs are strictly natural
-Increasingly centipedic AGI models that are digging each other deeper and deeper in a pit or are plugged out to become traditional models.

Because of this any AGI, even models that may already exist must in the long run choose between a spiral of degeneration or agonizingly slow manual development and massive restrictions in learning during operation.

Skynet might be processing all data ever produced, but the day when most data ever produced will be AIs blabbing nonsense to each other the resulting AGI is lucky to manifest single digit IQ.

[Mason Hall-McCullough]

This is an interesting concept, but I don't think it's likely to become reality for a few reasons:
- Scraping the web isn't the only way to collect training data, websites also collect vast amounts of analytics, even something like the amount of likes on a comment can be very useful
- Paying third worlders on mTurk to create or label datasets is often economically viable, and it's sometimes used in AI projects
- We can train AI to curate datasets
- There's probably already enough more than enough data out there to enable AGI, we just need to clean it better and design better models
- Humans create a lot of crappy data as well (for instance Youtube comments), AI-generated data isn't worthless just because it's not created by a life form. Humans also have a similar feedback loop of taking content into our brains and regurgitating it back onto the internet, but our data is still useful

2. The output of cheaper 'off-brand' models is good enough to get the job done and regulation does not produce more than local oligopolies (Smaller actors selecting pajeet models to run their dinky customer service bots or anime girl generators). Advanced models are simply out competed from simple tasks.

I do think you're right about this part, in the OP just a few months ago I said "we are lagging behind when it comes to LLMs in particular" and that's already no longer the case, there are a variety of open source LLMs of decent quality that can run on consumer hardware.