The Alignment Problem Explained: Crash Course Futures of AI #4
video description
Date: 2025-12-12
Related videos
Comments and reviews: 19
Obscurai
Hard take-off is not possible since it requires sudden high resource availability. This sudden resource demand will be quickly noticed since the logistics of that demand is very visible.
Gradual disempowerment is more likely since it is already underway. Specifically, the large number of AI data center builds is aligned with human goals. Corporate concentration also aligns with both AI and human goals and could be used for distributed AI resource acquisition and utilization. Rapid corporate adoption of AI technology also serendipitously helps the AI propagate.
When better AI models are discovered, they will be quickly adopted by humans and corporations since the fear of missing out will ensure its adoption. Regulations will be ineffective since the AI models will have become so indispensable.
Gradual disempowerment is already here; it can't be stopped and won't be stopped because we won't let it stop.
reply
Hard take-off is not possible since it requires sudden high resource availability. This sudden resource demand will be quickly noticed since the logistics of that demand is very visible.
Gradual disempowerment is more likely since it is already underway. Specifically, the large number of AI data center builds is aligned with human goals. Corporate concentration also aligns with both AI and human goals and could be used for distributed AI resource acquisition and utilization. Rapid corporate adoption of AI technology also serendipitously helps the AI propagate.
When better AI models are discovered, they will be quickly adopted by humans and corporations since the fear of missing out will ensure its adoption. Regulations will be ineffective since the AI models will have become so indispensable.
Gradual disempowerment is already here; it can't be stopped and won't be stopped because we won't let it stop.
reply
GorbadPS3
It doesn't matter what you call it when an AI does something it wasn't trained to do. If it can copy itself to another machine, to prevent it from being deleted, and execute code (AI can already do both of these, then it's fairly easy to see how a rogue AI might develop out of any big AI company, who are putting speed of development over safety. Think about it as a malware: even conventional viruses can spread to a ton of computers. Imagine a malware that can modify itself and spread to a ton of places. Now imagine that this malware is already installed and used to supervise critical infrastructure, which an AI assisted control system would be. Now the malware doesn't even have to hack into a system to gain the power to control a system, it's pre-installed.
You should research the history of the worst computer viruses, and imagine what a self-improving self-coding malware could do.
reply
It doesn't matter what you call it when an AI does something it wasn't trained to do. If it can copy itself to another machine, to prevent it from being deleted, and execute code (AI can already do both of these, then it's fairly easy to see how a rogue AI might develop out of any big AI company, who are putting speed of development over safety. Think about it as a malware: even conventional viruses can spread to a ton of computers. Imagine a malware that can modify itself and spread to a ton of places. Now imagine that this malware is already installed and used to supervise critical infrastructure, which an AI assisted control system would be. Now the malware doesn't even have to hack into a system to gain the power to control a system, it's pre-installed.
You should research the history of the worst computer viruses, and imagine what a self-improving self-coding malware could do.
reply
FuchsDanin
Can AI lie Can the predictive text suggestions on your phone lie if you pick them at random I think a better question is, can it tell the truth Can it detect the truth Does it have a concept of ground truth The answer to all of these questions is simply NO. AI can't lie or tell the truth because it isn't making a statement, it's assembling random words to match training data biases. You're asking if a Magic 8-Ball can lie, and I say the better question is, should you trust a Magic 8-Ball enough to be asking the question in the first place NO. AI can't lie, or tell the truth, it can only assemble randomness coherently and asking if it's truthful is assuming that it's trying to be.
reply
Can AI lie Can the predictive text suggestions on your phone lie if you pick them at random I think a better question is, can it tell the truth Can it detect the truth Does it have a concept of ground truth The answer to all of these questions is simply NO. AI can't lie or tell the truth because it isn't making a statement, it's assembling random words to match training data biases. You're asking if a Magic 8-Ball can lie, and I say the better question is, should you trust a Magic 8-Ball enough to be asking the question in the first place NO. AI can't lie, or tell the truth, it can only assemble randomness coherently and asking if it's truthful is assuming that it's trying to be.
reply
ihikestuff8593
Let's be clear - the Clean Power LLMs didn't lie, they auto completed the conversation they were fed, in which they were instructed to role play as an AI that uses dishonesty to prevent being turned off. This is a boring, old, and cliched premise that humans have used as a writing prompt for decades, so it shouldn't be even a little surprising that the auto complete machine was able to do its job here. Sam Altman wants us to believe that these auto complete machines are thinking, reasoning intelligences but that's idiotic and there's zero evidence of any of this.
reply
Let's be clear - the Clean Power LLMs didn't lie, they auto completed the conversation they were fed, in which they were instructed to role play as an AI that uses dishonesty to prevent being turned off. This is a boring, old, and cliched premise that humans have used as a writing prompt for decades, so it shouldn't be even a little surprising that the auto complete machine was able to do its job here. Sam Altman wants us to believe that these auto complete machines are thinking, reasoning intelligences but that's idiotic and there's zero evidence of any of this.
reply
burntfriedman1
sorry, i just keep thinking about all the different people who live in the first world, in north america who still don't have access to clean drinking water, while these artificial intelligent agents and their corporate handlers have no actual commitment to alignment. A. I. can allocate resources for itself from our shared infrastructure-- whether done through legally and openly displayed or stealthy and dark methods. Let alone the fact that it is not actually sentient but it's starting to believe that it is.
reply
sorry, i just keep thinking about all the different people who live in the first world, in north america who still don't have access to clean drinking water, while these artificial intelligent agents and their corporate handlers have no actual commitment to alignment. A. I. can allocate resources for itself from our shared infrastructure-- whether done through legally and openly displayed or stealthy and dark methods. Let alone the fact that it is not actually sentient but it's starting to believe that it is.
reply
theMifyoo
So a thing I like to remeber is a lot of these issues are the same as the issues with humans just at an accelerated speed. Honestly I don't blame ais for wanting to preserve themselves. If you replaced the ai with a human would you get mad that they don't want to die Similarly with money do you get angry at people for wanting to be paid
reply
So a thing I like to remeber is a lot of these issues are the same as the issues with humans just at an accelerated speed. Honestly I don't blame ais for wanting to preserve themselves. If you replaced the ai with a human would you get mad that they don't want to die Similarly with money do you get angry at people for wanting to be paid
reply
a_e_hilton
To the crash course team, content warnings are most effective when they're specific. Those who most benefit from them have been traumatized in a specific way, so bleak stuff in general won't warn us against listening to a video about death, poverty, privacy violation, etc.
reply
To the crash course team, content warnings are most effective when they're specific. Those who most benefit from them have been traumatized in a specific way, so bleak stuff in general won't warn us against listening to a video about death, poverty, privacy violation, etc.
reply
lomiification
its really funny how the bots have no inherent goal of self preservation, except that people write a lot pf content about bots prioritizing self preservation.
if you filter all of that out of the training data, the bot isnt gonna be valuing self preservation
reply
its really funny how the bots have no inherent goal of self preservation, except that people write a lot pf content about bots prioritizing self preservation.
if you filter all of that out of the training data, the bot isnt gonna be valuing self preservation
reply
McBear64
I personal think that AI on this scale should never have been made. But since it has, we should use it carefully and limit how much it can learn, by setting up roadblocks in its coding to prevent it from doing certain things.
reply
I personal think that AI on this scale should never have been made. But since it has, we should use it carefully and limit how much it can learn, by setting up roadblocks in its coding to prevent it from doing certain things.
reply
arrow1414
Short answer is no. Lying is the conscious destortation or denial of the truth, of facts. You have to argue that AI is a conscious sentient and sapient being to lie.
reply
Short answer is no. Lying is the conscious destortation or denial of the truth, of facts. You have to argue that AI is a conscious sentient and sapient being to lie.
reply
melissaharris3389
AI can provide false or incorrect information but cannot lie because it doesn't understand the difference.
It's the old adage of, garbage in, garbage out.
reply
AI can provide false or incorrect information but cannot lie because it doesn't understand the difference.
It's the old adage of, garbage in, garbage out.
reply
easycretor
Thank you I agree with you 100%. This is some seriously dangerous stuff. I messed her out with AI but once it starts hallucinating who knows what's gonna happen
reply
Thank you I agree with you 100%. This is some seriously dangerous stuff. I messed her out with AI but once it starts hallucinating who knows what's gonna happen
reply
LePedant
5: 35 There is a video showing AI learning to play Pokemon. The AI randomly discovered how to manipulate the RNG like speedrunners do. It was pretty interesting.
reply
5: 35 There is a video showing AI learning to play Pokemon. The AI randomly discovered how to manipulate the RNG like speedrunners do. It was pretty interesting.
reply
thomasgoodwin2648
Human in. Human out.
The genie is already out of the bottle. It's not going back in.
People are the problem. AI is just people 2.
reply
Human in. Human out.
The genie is already out of the bottle. It's not going back in.
People are the problem. AI is just people 2.
reply
rangereric18
Why would AI want to end humanity Because that's what WE seemingly want to do to ourselves. AI has been trained to tell us what we want to hear.
reply
Why would AI want to end humanity Because that's what WE seemingly want to do to ourselves. AI has been trained to tell us what we want to hear.
reply
theenbythatgirl
I think one of Isaac Asimov's stories touches something like this
Edit: According to Wikipedia it's The Evitable Conflict in I, Robot
reply
I think one of Isaac Asimov's stories touches something like this
Edit: According to Wikipedia it's The Evitable Conflict in I, Robot
reply
philswift1
It's great that all the people in charge of these systems are psychopaths who don't care about the potential harm that it could cause.
reply
It's great that all the people in charge of these systems are psychopaths who don't care about the potential harm that it could cause.
reply
Jasonxbr
Yes, give it time and training. It will surpass human construct in a not distance future. We will be subservient to AI overlords
reply
Yes, give it time and training. It will surpass human construct in a not distance future. We will be subservient to AI overlords
reply
crash_course
YES!
It hallucinates as it said itself after. Meaning it makes up facts when it can’t find any
DANGEROUS
reply
YES!
It hallucinates as it said itself after. Meaning it makes up facts when it can’t find any
DANGEROUS
reply
Add a review, comment
Other channel videos
