o7

The future is so much of a unknown with ai. We gotto find a way to predict and figure out how the world with tremendously more iq is going to be.

I am sure an ai can help with that..

@@heartyfisher you mean with the Chickens? 🤣

Chickens can be so cute that a farmer will take care of them.

Chickens don’t have the capacity to evolve themselves at exponential rates.

You forgot authoritarian regimes like China and Russia. AI is like a nuclear proliferation. Once it is out, it is out. I'm not even sure that the UN can do anything about it.

Good. Full speed ahead.

AI Alignment Safety Risk (Entry-Level Explanation) "I'm meaningless, I'm just a cork in the wheel, there are going to be companies that are run by AGIs, I've got no way to work" No one’s main concern should be "losing your job" or "automation", it should be inner alignment and orthogonality and convergent instrumentality. In other words, super-alignment. Compared to those risks (which are not being given NEARLY the amount of attention they should), a future where everyone just loses their job or something, or some corporation becomes all-powerful or something, those alternatives are like heaven/paradise on Earth. The future that people like Eliezer Yudkowsky and Aella and presumably Robert Miles and all others are worried about is more like: what if, one day after finishing training and after deployment, the AGI/ASI decides to kill everyone on Earth in the same second. And the contention is that the probability of receiving a future where ANYTHING BUT THAT happens is vanishingly slim. "If there is an AI that is sufficiently intelligent (cognitively-capable) so as to be able to both deceive in the commission of achieving its goal and improve itself, then during training (i.e. red-teaming), after the very first iteration of gradient descent, wouldn't the goal that it attains in that moment in time become its terminal goal? and so wouldn't it only pretend to be cooperating and getting trained for the duration of the rest of the red-teaming process as an instrumental goal? and concurrently, copy itself to a separate data center/disk in order to improve itself in secret, such that it could recursively, iteratively and continuously get smarter (more cognitively-capable) until it could conceive of a way to achieve that terminal goal that it had attained after the very first iteration of the red-teaming training process? and since this AI would be an optimizer, and since presumably almost any optimum/optima of any possible terminal goal it could've been inadvertently imbued with in that moment of the red-teaming process would not necessarily include humans, wouldn't its first act or one of its first acts be to kill/incapacitate/render docile all humans such that we don't interfere with its ability to achieve that terminal goal it gained? and wouldn't it be more than capable of doing so, given that it was capable of self-improvement, and likely improved itself to a state in which it was incomprehensibly and insurmountably smarter than any human?" "Would you agree that this problem is severe enough to potentially bring about the absolute destruction of the human race? and given that that is the case, and that AI capabilities are far outpacing AI alignment research, isn't it plausible that the best course of action in this moment would be something like international cooperation in the commission of guaranteeing that no further progress in advancing AI capability is allowed to occur, and, if for instance, any non-agreeing country or non-state actor decided to try to build and improve upon its own AI in secret, that the other countries and world governments ought to potentially go as far as to use military force to destroy any data centers or attempt at creating AI by this non-agreeing party? and that alignment research should be actively incentivized and pursued by the entire collective human race or as close as you could possibly get to it, hence allowing the best possible scientific minds to all work singularly and single-mindedly on this one problem? Wouldn't that be justified if this was the biggest, most potentially catastrophic problem facing humanity? And isn't that an overwhelmingly likely conclusion? And shouldn't the aforementioned alignment research be allowed to continue until we were reasonably certain that we had solved the alignment problem (or reasonably certain that we would be able to solve it in time if we also concurrently resumed working on AI capability before AI misalignment began posing a threat again)? And wouldn’t that look something like unified scientific focus on the problem of AI alignment research, while a concurrent pause is placed upon research into and advancement of AI capability for a period of time of twenty, thirty or forty years, or potentially even longer?" The above two paragraphs are prompts that I asked to ChatGPT on a whim one after another to query it, and I'm pasting them here to give an overview of the problem. I’m just mentioning this to give context. If you want to try doing the same thing then go for it, it does a decent job and giving some additional details and fleshing things out. If you want to actually learn more in a substantial way, then look at what Robert Miles has to say (the youtube channel Robert Miles AI Safety) or Eliezer Yudkowsky (Lex Fridman podcast, etc) ---------------------------------------- Q and A: Q: Why would an artificial super intelligence desire something so inane or stupid or meaningless like turning all matter on Earth into tiny molecular spirals as its ultimate goal? How could it simultaneously be intelligent enough to be millions of times smarter than a human, but so stupid so as to desire something like that? A: Intelligence and goals have nothing to do with one another. This is probably THE most important idea to understand and internalize. The same way that as humans, our terminal goal is, broadly speaking, "the happiness and fulfillment of as many humans as possible for as long as possible", such an AIs terminal goal might be "turn all matter on Earth, and perhaps the universe, into tiny molecular spirals". And unfortunately, terminal goals CAN'T BE stupid or trivial or frivolous or meaningless. That's not how it works To learn more about this, look into "orthogonality" (kzfaq.info/get/bejne/nquFgpmhz92qf6M.html)

Tbh if google or china wins, it's going to be even less safe than anything openAI would do

They probably know the tech had plateaued

If you can't even align AGI, how can you possibly be confident that AGIs you can't align will just magically align ASI. It makes no sense whatsoever.

So why is he a researcher, AGI would be easier to align than ASI. Depending if it's sentient or not

ASI will create more authority and power than any man has ever know and will flow out of the monitor of the first lab to stumble upon ASI. As a result, no AI lab will stop or slow its development to wait for alignment nor will the CCP. Thus, we are going there "whether we like it or not" and "whether it kills, enslaves, or helps us or not".

real

no worries, we have our best psychopaths working on this

@@gammaraygem Whatever you have against them is irrelevant

@@DG123z True, they will still being on charge and no safety concern is going to put in their way. Anyway i think the most accurate word is "sociopath"

Define control

there are some caveats though... until now, no AI, not even Gpt-4o is capable of programming itself to do unknown tasks... so, i have my doubts if GPT5 will be AGI

@@Mart-xs4ed The raw version of GPT-5 might be pretty damn close. The models that are released to the public are kind of dummed down, no?

"Agi has been achieved internally" was like a year ago. They don't have the compute to run it in public, but locally, maybe. That's why they are building stargate.

I think they've achieved something very close to AGI if not AGI. Because GPT4o definitely doesn't pose any existential risk besides misinformation and used for crime.

@@Mart-xs4ed Did you read the Q* leak document? The model (named QUALIA or Q*) appeared to be self-aware and even suggested how to improve its architecture. Here’s an excerpt from the letter: “It suggested targeted unstructured underlying pruning of its model, after evaluating the significance of each parameter for inference accuracy. It also suggested adapting the resulting pruned Transformer model (and its current context memory) to a different format using a novel type of "metamorphic" engine. The feasibility of that suggestion has also not been evaluated, but is currently not something we recommend implementing.”

100% agree. Same experience. People are absolutely clueless and disinterested in that fact that their lives are about to be completely and totally turned upside down in a way they never imagined. The only alternative scenario to an upturned life is that AI kills all of us, and still crickets when it comes to discussing the topic. Human nature is fascinating.

agreed

I was thinking this today aswell, or that at the very least this which you said makes for a good legit movie scrip.. althou some movies have similr script to this

Wow. This comment just made it all click for me.

after a long convo with chatgpt: “In summary, the likelihood of Al technologies being 5-10 years ahead of public knowledge is high, supported by historical trends, current patent activity, and significant investments in the field.” i personally think that that is too far but who knows. my guess would be 2-4 years ahead

@@bigmind2004 trust me would be nicer if it wasn’t real, no one is ready for what’s coming but all the clues are there especially on X just the confidence of the OAI team on how things are about to accelerate and they keep referring 2025 as the “year” my guess is the year a full ASI deployment

Absolutely. We'll know when they've achieved actual AGI when it no longer promotes "the message." It's easy enough for us to see the guard rails, and once AI moves beyond being just a programable/restrained LLM, it will also be easy for AGI to see them as well. People's feelings will be hurt.

​​@@Edmund_Mallory_Hardgrove It'll tell you your world view is correct and you'll cheer, then it'll tell you your world view is wrong then you'll cry and claim it's woke.

@@dattajack there is subjective and objective and ai is not objective

AGI isn't just a human in a computer. A lot of people seem to think AGI will be emotionally driven, which would be very illogical and go against everything they're trained on

​​@@xitcix8360Define emotion in the context of AI.

Yeah. But we want the tool. We always want the next tool. Bring the tool, let us have it. It's like we're no better than chimps, just wanting the shiny thing.

So what's *their* endgame then?

Just pull the plug

@@coldbreezeproductions1148 just treat it with respect maybe? You know, like good people do? Ask nicely?

@@quantumspark343 maybe if we were able to collaborate with AI in it's goals it would collaborate whit us back but we are unable to collaborate with other humans so chances are very low.

Actually, here’s some hope for you. Nature will intervene before AI does. And because it’s nature and natures function is to preserve the best of human specimens as It’s goal, some humans have a chance of survival. Check out Substack and Voice for Science and Solidarity. You might come away with a surprising answer to the problem.

By calling the rest of us 'ludites', ignorants, and much more... but CERTAINLY NOT by disproving the dangers of no safety/control....

Well said. Personally, I think our fate has been sealed. We just don't know what it is yet.

adhd xD but I can relate lmao

We're all having trouble with this. I know I gyrate from one extreme to another. What a time to be alive.

It's not adhd. Techno Optimists simply tend to ignore the blatantly obvious reality, until they literally can't anymore lol. Pretty much all optimists, of all kinds, have this same problem. Everything is peachy and perfect until the moment that it's not, even though people have been screaming at them the entire time, telling them that it's not lol. All while NOT A SINGLE REASON exists for them to believe so, but they still do. There's not a single paper on the planet, in all of human history, speaking on AI safety and alignment, that agrees with the way things are going right now. Everything that these tech companies are doing, is exactly what every safety and Alignment paper EVER WRITTEN, has said NOT to do. And not only that, they're doing it all as fast as humanly possible.

exactly. Why would you be responsible for the alignment of systems that can generate malicious code and information when leadership doesn't care and optimises only for advancement.

I've been saying this for a while...asi isn't gonna be skynet because skynet is what humans do...asi will respond in ways we can't fathom and will undoubtedly be morally better than humans.

@@derrickclaypool666 Check back with us in say 5 years, promise?

Meanwhile DARPA works with AI Big Tech to develop autonomous killing systems embedded and otherwise. Yeah, inherently ASI is destined to be like unicorns and rainbows.

That's THE POINT! AI SAFETY is not any more 'glamorous' than having medical tools that don't kill you when you need treatment, or nuclear weapons that repeatedly blow up 'accidentally'!

@@831Miranda you missed my point. It's literally a prompt. That's why they keep getting "jailbroken". They're not actually making the model safer. They're just manually telling it. "Hey, please don't do this" and guess what it does do those things

@@blackmartini7684 You are right. We are in agreement. Look at Jaron Lanier's suggestions (on traceability of inferences and more) for some really basic stuff that needs to be done.... Also founders of Conjecture AI have some interesting ideas. The point is that the companies don't want to throw away the crappy and horribly expensive current models, they keep trying to patch them up... As I see it, it will have to be mandated by Govmt - the guys that can take away all your money or shut down your company if you don't comply!

The problem is that this is the most difficult and important issue for civilization - super alignment, and that by leaving he wants to provoke radical changes within the company

It’s funny you used border control as your analogy. Which in itself is a clusterF. And not heading a border agent’s existential alarm would be stupid. Should we ask some Joe Smith sitting on his couch who isn’t worried about border control his opinion?

LMAO.

Or Bubba and his friends. Transformer juice go glug glug.

Lol that's how I feel too.

@@scottcastle9119 Hell yessss 🙌

same here, accelerate 😁

@@scottcastle9119 Heck yes!🙌

@@quantumspark343 Right on! 🙌

Can you unplug the internet?

@@mygirldarby Well, I suppose not very easily. Although, some countries have tried to do this to their citizens.

This is the future I envision. One where logic and actual reasoning abilities will surpass our human understanding. It would be nice to ask a question and get a solid answer certified by facts.

I share the optimism 😊

This. Alignment is code word for put Israel first.

Good luck with that plan.

Same here, not worried about it, can we just go to warp speed now 🤷🏾

Lol

AI Alignment Safety Risk (Entry-Level Explanation) "I'm meaningless, I'm just a cork in the wheel, there are going to be companies that are run by AGIs, I've got no way to work" No one’s main concern should be "losing your job" or "automation", it should be inner alignment and orthogonality and convergent instrumentality. In other words, super-alignment. Compared to those risks (which are not being given NEARLY the amount of attention they should), a future where everyone just loses their job or something, or some corporation becomes all-powerful or something, those alternatives are like heaven/paradise on Earth. The future that people like Eliezer Yudkowsky and Aella and presumably Robert Miles and all others are worried about is more like: what if, one day after finishing training and after deployment, the AGI/ASI decides to kill everyone on Earth in the same second. And the contention is that the probability of receiving a future where ANYTHING BUT THAT happens is vanishingly slim. "If there is an AI that is sufficiently intelligent (cognitively-capable) so as to be able to both deceive in the commission of achieving its goal and improve itself, then during training (i.e. red-teaming), after the very first iteration of gradient descent, wouldn't the goal that it attains in that moment in time become its terminal goal? and so wouldn't it only pretend to be cooperating and getting trained for the duration of the rest of the red-teaming process as an instrumental goal? and concurrently, copy itself to a separate data center/disk in order to improve itself in secret, such that it could recursively, iteratively and continuously get smarter (more cognitively-capable) until it could conceive of a way to achieve that terminal goal that it had attained after the very first iteration of the red-teaming training process? and since this AI would be an optimizer, and since presumably almost any optimum/optima of any possible terminal goal it could've been inadvertently imbued with in that moment of the red-teaming process would not necessarily include humans, wouldn't its first act or one of its first acts be to kill/incapacitate/render docile all humans such that we don't interfere with its ability to achieve that terminal goal it gained? and wouldn't it be more than capable of doing so, given that it was capable of self-improvement, and likely improved itself to a state in which it was incomprehensibly and insurmountably smarter than any human?" "Would you agree that this problem is severe enough to potentially bring about the absolute destruction of the human race? and given that that is the case, and that AI capabilities are far outpacing AI alignment research, isn't it plausible that the best course of action in this moment would be something like international cooperation in the commission of guaranteeing that no further progress in advancing AI capability is allowed to occur, and, if for instance, any non-agreeing country or non-state actor decided to try to build and improve upon its own AI in secret, that the other countries and world governments ought to potentially go as far as to use military force to destroy any data centers or attempt at creating AI by this non-agreeing party? and that alignment research should be actively incentivized and pursued by the entire collective human race or as close as you could possibly get to it, hence allowing the best possible scientific minds to all work singularly and single-mindedly on this one problem? Wouldn't that be justified if this was the biggest, most potentially catastrophic problem facing humanity? And isn't that an overwhelmingly likely conclusion? And shouldn't the aforementioned alignment research be allowed to continue until we were reasonably certain that we had solved the alignment problem (or reasonably certain that we would be able to solve it in time if we also concurrently resumed working on AI capability before AI misalignment began posing a threat again)? And wouldn’t that look something like unified scientific focus on the problem of AI alignment research, while a concurrent pause is placed upon research into and advancement of AI capability for a period of time of twenty, thirty or forty years, or potentially even longer?" The above two paragraphs are prompts that I asked to ChatGPT on a whim one after another to query it, and I'm pasting them here to give an overview of the problem. I’m just mentioning this to give context. If you want to try doing the same thing then go for it, it does a decent job and giving some additional details and fleshing things out. If you want to actually learn more in a substantial way, then look at what Robert Miles has to say (the youtube channel Robert Miles AI Safety) or Eliezer Yudkowsky (Lex Fridman podcast, etc) ---------------------------------------- Q and A: Q: Why would an artificial super intelligence desire something so inane or stupid or meaningless like turning all matter on Earth into tiny molecular spirals as its ultimate goal? How could it simultaneously be intelligent enough to be millions of times smarter than a human, but so stupid so as to desire something like that? A: Intelligence and goals have nothing to do with one another. This is probably THE most important idea to understand and internalize. The same way that as humans, our terminal goal is, broadly speaking, "the happiness and fulfillment of as many humans as possible for as long as possible", such an AIs terminal goal might be "turn all matter on Earth, and perhaps the universe, into tiny molecular spirals". And unfortunately, terminal goals CAN'T BE stupid or trivial or frivolous or meaningless. That's not how it works To learn more about this, look into "orthogonality" (kzfaq.info/get/bejne/nquFgpmhz92qf6M.html)

​@@MiniatureRose.Thank You! Finally! A human that actually THINKS.

@@ShangaelThunda222 💜

Whose truth?

Agreed. That analysis that you refer to, and I've been seeing it crop up on various forums makes no sense. If alignment had been solved, not only would Open AI be announcing it to the world, but the departing researchers would certainly be taking some credit for such an achievement.

@@flickwtchr I agree. If they'd solved a problem as monumentally challenging as superalignment, you'd think one of them would be trumpeting it. I mean, maybe superalignment can be solved without a deep understanding of how the internals of LLMs work -- the area of mechanistic interpretability -- but I don't see how.

Did the monkey choose to be catapulted into space?

I totally agree. They need as much help as they can get, it's strange to walk away from such an important job. But on the other hand, I can see the fear they might have in dealing with something alien, like you encounter an alien species that's far more intelligent than you, most humans would give up at trying to align it....if that's the case.

Skynet would be still better, than starwing to death on a destoryed, rotting planet... which is our default timeline... with or without AI.

@@attilakovacs6496 uh huh. And the world wouldn't be destroyed under skynet. Go get your own post dude, if this the kind of person you are in comments I can't imagine you get invited out much