Urgh. On paper, this sounds like a great idea. Who knows what kind of patterns can be found in that data, patterns which could lead to genuine improvements and optimisations for all sorts of...
Urgh. On paper, this sounds like a great idea. Who knows what kind of patterns can be found in that data, patterns which could lead to genuine improvements and optimisations for all sorts of things? ML systems are great at that kind of thing.
It's just it's hard to think of who could be trusted to do that, because that dataset could also be mined for massive profit or worse. The person in question is not Larry fuckin' Ellison, that's for sure.
I'm going to have to disagree. On paper this sounds like a dystopic nightmare. Like, literal books (science fiction and otherwise) have been written about how this is a bad, bad idea.
On paper, this sounds like a great idea.
I'm going to have to disagree. On paper this sounds like a dystopic nightmare. Like, literal books (science fiction and otherwise) have been written about how this is a bad, bad idea.
I mean, on paper this sounds like magical thinking. This is not an accurate description of any AI model. This is the description of an all-knowing oracle. The missing link is not the "national...
I mean, on paper this sounds like magical thinking.
"I have to tell [the] AI model as much about my country as I can," Ellison said. "We need to unify all the national data, put it into a database where it's easily consumable by the AI model, and then ask whatever question you like," he said. "That's the missing link."
This is not an accurate description of any AI model. This is the description of an all-knowing oracle.
The missing link is not the "national data" -- it's the oracle.
Some days I feel like AI is exactly like giving a Magic 8 Ball to a 5 year old. They decide it can answer every single question for them for the rest of their lives.
Some days I feel like AI is exactly like giving a Magic 8 Ball to a 5 year old. They decide it can answer every single question for them for the rest of their lives.
I, for one, am in favour of Larry Ellison inhaling toxic vapours for the sake of prophetic visions, and I don’t even think he needs to be half naked for it either!
I, for one, am in favour of Larry Ellison inhaling toxic vapours for the sake of prophetic visions, and I don’t even think he needs to be half naked for it either!
Exactly. ML is good at finding patterns. It sounds like he wants an LLM. Because what we really need is a hallucinating Deep Thought (HHGttG) to tell us that all of the problems in the United...
Exactly. ML is good at finding patterns.
It sounds like he wants an LLM.
Because what we really need is a hallucinating Deep Thought (HHGttG) to tell us that all of the problems in the United States are attributable to a single root cause: lesbian nazi hookers abducted by UFOs and forced into weight loss programs (borrowed from UHF) and the answer is "42" only then to have President Camacho (Idiocracy) do whatever it says.
So, clearly, I see this all as a terrific idea. /s
If it did work, imagine how quickly this would be quietly shut down when it comes to the conclusion that billionaires are a big chunk of the problem. Or maybe they'd quietly "retrain out biases".
If it did work, imagine how quickly this would be quietly shut down when it comes to the conclusion that billionaires are a big chunk of the problem. Or maybe they'd quietly "retrain out biases".
Yeah, no. Hard pass. As a Dutch person, this sounds like a nightmare. We've already seen firsthand how self-learning algorithms can cause real harm. The problem is that these systems are not...
On paper, this sounds like a great idea.
Yeah, no. Hard pass. As a Dutch person, this sounds like a nightmare. We've already seen firsthand how self-learning algorithms can cause real harm.
The problem is that these systems are not neutral. They inherit biases from the data they're trained on and the instructions they're given, no matter how neutral the people designing them try to be. Worse, they tend to act on those biases in ways that are completely opaque.
In the Dutch case, an algorithm flagged thousands of families for benefits fraud, disproportionately targeting minorities and low-income households. Families were forced to repay massive sums they didn’t actually owe, leading to bankruptcies, lost homes, and lives completely upended. The worst part? Nobody could explain why the algorithm flagged them in the first place, and appeals were nearly impossible because the system was treated as infallible.
Now imagine a system like that, but with your DNA and entire medical history. Even with good intentions, these systems don’t just reflect societal biases, they amplify them while hiding behind a false sense of technological objectivity. The real danger isn’t just who controls the data, it’s the fundamental risk of letting algorithms make decisions about people’s lives on this scale.
Yeah I don't think I can even guess just how bad a Toeslagenaffaire would be on the largest economy, with half a billion people, intertwined with worldwide affairs, all the while being run by Musk...
Now imagine a system like that, but with your DNA and entire medical history. Even with good intentions, these systems don’t just reflect societal biases, they amplify them while hiding behind a false sense of technological objectivity. The real danger isn’t just who controls the data, it’s the fundamental risk of letting algorithms make decisions about people’s lives on this scale.
Yeah I don't think I can even guess just how bad a Toeslagenaffaire would be on the largest economy, with half a billion people, intertwined with worldwide affairs, all the while being run by Musk and his cronies. Oh, and given how the gang doesn't appear to give a shit about encryption and other security issues it would be immediately abused by hostile powers on top of it.
The single benefit I can imagine from this is that the fallout would probably be so utterly bad, that the importance of privacy would be immediately visible. But that price would be payed with blood. I'm not even exaggerating here, people would absolutely die due to that shit.
Sure, but that's an implementation issue. Treating the output as infallible is dumb. This doesn't have to be done in a dumb way. It's easy to imagine this being done badly, but why is it any less...
Sure, but that's an implementation issue. Treating the output as infallible is dumb. This doesn't have to be done in a dumb way. It's easy to imagine this being done badly, but why is it any less valid to imagine it being done well?
My point is that, implemented right - and I'm not at all confident that's actually real-world possible, hence me saying "on paper" - there are potentially very useful insights to be gained from this kind of large scale data mining.
What if it turned out that 95% of people who bought more than 200g of apples a week had a 75% lower risk of lung cancer? That's useful information which would save lives. It's a silly and trivial example, of course, but this is the kind of pattern which you can't find any other way, and the sort of thing ML systems are good at spotting.
Because there are enough real-world examples where it has gone wrong, and the impact of failure is devastating. As I already said, when it does go wrong, these systems tend to disproportionately...
Sure, but that's an implementation issue. Treating the output as infallible is dumb. This doesn't have to be done in a dumb way. It's easy to imagine this being done badly, but why is it any less valid to imagine it being done well?
Because there are enough real-world examples where it has gone wrong, and the impact of failure is devastating. As I already said, when it does go wrong, these systems tend to disproportionately harm those already in vulnerable positions.
To phrase it differently in a way that might make the issue more clear, we are talking about effectively profiling here. So, we are not just talking about inefficiencies or bugs here. We are talking about automated profiling that can destroy lives.
and I'm not at all confident that's actually real-world possible, hence me saying "on paper"
You are kind of answering your own question. You now made it clear you are approaching this is more or less an academic high level musing, a theoretical exercise. But, that doesn't really come across in the initial comment I responded to. Even if it came across like that, we are already past theoretical. Hence, my “hard pass” comment as this is not just an academic discussion as this sort of technology already has backfired. Causing harm to real people, and without the proper care and attention they will backfire again.
What if it turned out that 95% of people who bought more than 200g of apples a week had a 75% lower risk of lung cancer? That's useful information which would save lives.
True, though actually not something ML systems are uniquely suitable for. And, again, it assumes the input data is unbiased as well as entirely correct. It also is a line of thinking that personally irks me a bit as it approaches “yeah, sure there might be negatives, but think about the positives!” I often see technologist tout waving away potential issues to effectively be kicked down the line. To a point where they are much harder to solve and already have impacted people negatively.
Well, yes. Isn't that what "on paper" means? I did also say it's hard to think of who could be trusted to implement it correctly. I apologise if I wasn't clear initially. Absolutely, we are in...
You now made it clear you are approaching this is more or less an academic high level musing, a theoretical exercise.
Well, yes. Isn't that what "on paper" means? I did also say it's hard to think of who could be trusted to implement it correctly. I apologise if I wasn't clear initially.
without the proper care and attention they will backfire again.
Absolutely, we are in agreement. The technology isn't the problem per se, but the implementation is. Which is exactly what I originally said (or at least intended to say).
I don't think we are on the same page, maybe I didn't make it clear. Even on paper it should already be abundantly clear that it is incredibly difficult to implement it correctly. It doesn't even...
I don't think we are on the same page, maybe I didn't make it clear. Even on paper it should already be abundantly clear that it is incredibly difficult to implement it correctly. It doesn't even have much to do with trusting the people that implement it. As a lot of it comes down to biases in the input data that are not immediately obvious.
The problem isn't just about implementation or technology in isolation. It is about the fundamental challenge of trying to extract “neutral” insights from data that reflects a society full of historical and systemic biases. Even with perfect implementation, if your training data comes from a healthcare system where certain groups have been historically underserved, your AI will learn and amplify those patterns. If your data comes from a justice system with documented biases, those biases become baked into the model.
This isn't a technical problem that can entirely be solved with better code or more careful implementation. They will help, but ultimately it is a social problem that we're trying to solve with technology, and that's where things go wrong. The Dutch benefits scandal wasn't just a case of poor implementation, it was a case of an algorithm doing exactly what it was designed to do, while nobody fully understood the social implications of those design choices.
By centering the discussion about technology you are effectively still waving away practical concerns that are very much related to the technology. Specifically the fact that with ML these biases are not made more clear, but actually are hidden away as the reasoning process of the technology process is not visible. So yes, the technology itself also is part of the issue.
I think there's a significant difference between "this technology has limitations and issues that we know about and some we might not yet, but it can still give positive results if used correctly"...
I think there's a significant difference between "this technology has limitations and issues that we know about and some we might not yet, but it can still give positive results if used correctly" and "waving away" those issues. I'm not unaware of the issues you mention. But I also think they are able to be accounted for and worked around with care and forethought. You said this was "incredibly difficult" not "impossible" so you may agree more with me than you think.
The Dutch benefits scandal wasn't just a case of poor implementation, it was a case of an algorithm doing exactly what it was designed to do, while nobody fully understood the social implications of those design choices.
That sounds precisely like poor implementation to me, but clearly we differ on exactly what that means. :)
I am going to try to keep this short, as I am otherwise going in circles. One thing I want to stress: the context here is us discussing a massive database incorporating all data into one system....
I am going to try to keep this short, as I am otherwise going in circles.
One thing I want to stress: the context here is us discussing a massive database incorporating all data into one system. There is simply no reasonable or feasible way to avoid bias in the input when dealing with data at this scale, coming from countless sources with different levels of accuracy and systemic biases. And again, we’re talking about technology that doesn’t just reflect those biases but enhances and obscures them in ways that even experts struggle to detect in much smaller systems.
As I think the overall point is made clear enough I will now bow out of this conversation.
The cool part is that you don't even need an LLM jailbreak to get that data! Your data is probably already floating around in various data breach dumps. Isn't it great how unseriously infosec is...
The cool part is that you don't even need an LLM jailbreak to get that data! Your data is probably already floating around in various data breach dumps. Isn't it great how unseriously infosec is taken even by places that give you no other choice, like credit reporting agencies?
Isn't that the plot of Person of Interest? A "jailbreak" (backdoor) in an AI that spits out SSN of people in danger but not considered relevant to national security?
Isn't that the plot of Person of Interest? A "jailbreak" (backdoor) in an AI that spits out SSN of people in danger but not considered relevant to national security?
Kinda weird how their vision for the best path to take that'll make everyone's lives better, more secure and more convenient just so happens to be one that personally enriches them beyond even...
Kinda weird how their vision for the best path to take that'll make everyone's lives better, more secure and more convenient just so happens to be one that personally enriches them beyond even their already ridiculous net worth.
I initially read that as "world's voter supply", and you know where that ends up... This way lies social credit scoring, permanent loyalty and criminal records, chilling levels of algorithmic...
I initially read that as "world's voter supply", and you know where that ends up...
This way lies social credit scoring, permanent loyalty and criminal records, chilling levels of algorithmic discrimination, and all the other virtues of the Panopticon.
I remember reading a short story in an anthology about AI circa 2000 that involved a new supercomputer being demonstrated to the government. Best part was when it made some sympathetic comments...
I remember reading a short story in an anthology about AI circa 2000 that involved a new supercomputer being demonstrated to the government. Best part was when it made some sympathetic comments about socialism (making reference to early Christian communes) and pissed off a right-wing general. One wonders what the powers that be would do when their godlike AGI ingests All The Data only to output such solutions as "INSTITUTE UNIVERSAL HEALTHCARE" and "DECARBONIZE NOW YOU DOLTS".
Doesn't AI already have a ton of this information? Why would we need to "feed" it national security clearance level information to get better ideas on how to spend money? The bills and spending is...
Doesn't AI already have a ton of this information? Why would we need to "feed" it national security clearance level information to get better ideas on how to spend money? The bills and spending is already public information and there's been a ton of studies about so much stuff online as well, and these companies have no issue stealing so Im sure a lot of them are already baked in.
The other thing is the questions you ask. Even something that sounds simple isn't " How can we make America better" is so reliant on who defines 'better'.
There's a ton of ideas and concepts to try that are likely to make the country better but it doesn't align with corporate donors, lobbyists, high profile donors and it doesn't benefit politicians.
To me this is like asking AI to add 2+2 instead of just using a calculator to do it yourself.
The US government probably already has this running somewhere, and some "really smart" 19 year old in Musk's DOGE is copying it over to X servers right now, because it's "chock-full of fraud and...
The US government probably already has this running somewhere, and some "really smart" 19 year old in Musk's DOGE is copying it over to X servers right now, because it's "chock-full of fraud and inefficiency"
I for one trust Mr. BIG BALLS not to do anything shady with my private data. Oh wait, never mind, I guess BIG BALLS quit over racist remarks he made on social media. Oh wait, never mind x2, I...
I for one trust Mr. BIG BALLS not to do anything shady with my private data.
On the one hand, aggregating data for analysis does make sense, to a point. Talking about genomic data as part of individualized care doesn't really make sense in a large data set. For those...
On the one hand, aggregating data for analysis does make sense, to a point. Talking about genomic data as part of individualized care doesn't really make sense in a large data set. For those purposes you typically do separate analysis of genomes, and then introduce categorical variables in a larger dataset. So there is no reason that the meta/multimodal model needs raw generic data. It would probably be hindered by it.
There are some exciting clinical trials underway that use a combination of biomarkers and genetic data to drive healthcare, including truma care. Devices that pull several ml of blood per day, perform assays and genetic sequencing, and give recommendations to clinicians on care, such as when to close wounds to minimize sepsis, or how to treat burns. These are real, happening today, examples of deep learning improving standards of care. No mega centralized system required.
Urgh. On paper, this sounds like a great idea. Who knows what kind of patterns can be found in that data, patterns which could lead to genuine improvements and optimisations for all sorts of things? ML systems are great at that kind of thing.
It's just it's hard to think of who could be trusted to do that, because that dataset could also be mined for massive profit or worse. The person in question is not Larry fuckin' Ellison, that's for sure.
I'm going to have to disagree. On paper this sounds like a dystopic nightmare. Like, literal books (science fiction and otherwise) have been written about how this is a bad, bad idea.
I mean, on paper this sounds like magical thinking.
This is not an accurate description of any AI model. This is the description of an all-knowing oracle.
The missing link is not the "national data" -- it's the oracle.
You know what kind of database right? Fucking Oracle. One of the most obtuse and expensive RDBMS in existence. And who runs Oracle? Ellison.
Some days I feel like AI is exactly like giving a Magic 8 Ball to a 5 year old. They decide it can answer every single question for them for the rest of their lives.
We could go back to the priestess inhaling toxic vapors and telling the future model, Delphi style
I, for one, am in favour of Larry Ellison inhaling toxic vapours for the sake of prophetic visions, and I don’t even think he needs to be half naked for it either!
Exactly. ML is good at finding patterns.
It sounds like he wants an LLM.
Because what we really need is a hallucinating Deep Thought (HHGttG) to tell us that all of the problems in the United States are attributable to a single root cause: lesbian nazi hookers abducted by UFOs and forced into weight loss programs (borrowed from UHF) and the answer is "42" only then to have President Camacho (Idiocracy) do whatever it says.
So, clearly, I see this all as a terrific idea. /s
If it did work, imagine how quickly this would be quietly shut down when it comes to the conclusion that billionaires are a big chunk of the problem. Or maybe they'd quietly "retrain out biases".
Yeah, no. Hard pass. As a Dutch person, this sounds like a nightmare. We've already seen firsthand how self-learning algorithms can cause real harm.
The problem is that these systems are not neutral. They inherit biases from the data they're trained on and the instructions they're given, no matter how neutral the people designing them try to be. Worse, they tend to act on those biases in ways that are completely opaque.
In the Dutch case, an algorithm flagged thousands of families for benefits fraud, disproportionately targeting minorities and low-income households. Families were forced to repay massive sums they didn’t actually owe, leading to bankruptcies, lost homes, and lives completely upended. The worst part? Nobody could explain why the algorithm flagged them in the first place, and appeals were nearly impossible because the system was treated as infallible.
Now imagine a system like that, but with your DNA and entire medical history. Even with good intentions, these systems don’t just reflect societal biases, they amplify them while hiding behind a false sense of technological objectivity. The real danger isn’t just who controls the data, it’s the fundamental risk of letting algorithms make decisions about people’s lives on this scale.
Yeah I don't think I can even guess just how bad a Toeslagenaffaire would be on the largest economy, with half a billion people, intertwined with worldwide affairs, all the while being run by Musk and his cronies. Oh, and given how the gang doesn't appear to give a shit about encryption and other security issues it would be immediately abused by hostile powers on top of it.
The single benefit I can imagine from this is that the fallout would probably be so utterly bad, that the importance of privacy would be immediately visible. But that price would be payed with blood. I'm not even exaggerating here, people would absolutely die due to that shit.
Sure, but that's an implementation issue. Treating the output as infallible is dumb. This doesn't have to be done in a dumb way. It's easy to imagine this being done badly, but why is it any less valid to imagine it being done well?
My point is that, implemented right - and I'm not at all confident that's actually real-world possible, hence me saying "on paper" - there are potentially very useful insights to be gained from this kind of large scale data mining.
What if it turned out that 95% of people who bought more than 200g of apples a week had a 75% lower risk of lung cancer? That's useful information which would save lives. It's a silly and trivial example, of course, but this is the kind of pattern which you can't find any other way, and the sort of thing ML systems are good at spotting.
Because there are enough real-world examples where it has gone wrong, and the impact of failure is devastating. As I already said, when it does go wrong, these systems tend to disproportionately harm those already in vulnerable positions.
To phrase it differently in a way that might make the issue more clear, we are talking about effectively profiling here. So, we are not just talking about inefficiencies or bugs here. We are talking about automated profiling that can destroy lives.
You are kind of answering your own question. You now made it clear you are approaching this is more or less an academic high level musing, a theoretical exercise. But, that doesn't really come across in the initial comment I responded to. Even if it came across like that, we are already past theoretical. Hence, my “hard pass” comment as this is not just an academic discussion as this sort of technology already has backfired. Causing harm to real people, and without the proper care and attention they will backfire again.
True, though actually not something ML systems are uniquely suitable for. And, again, it assumes the input data is unbiased as well as entirely correct. It also is a line of thinking that personally irks me a bit as it approaches “yeah, sure there might be negatives, but think about the positives!” I often see technologist tout waving away potential issues to effectively be kicked down the line. To a point where they are much harder to solve and already have impacted people negatively.
Well, yes. Isn't that what "on paper" means? I did also say it's hard to think of who could be trusted to implement it correctly. I apologise if I wasn't clear initially.
Absolutely, we are in agreement. The technology isn't the problem per se, but the implementation is. Which is exactly what I originally said (or at least intended to say).
I don't think we are on the same page, maybe I didn't make it clear. Even on paper it should already be abundantly clear that it is incredibly difficult to implement it correctly. It doesn't even have much to do with trusting the people that implement it. As a lot of it comes down to biases in the input data that are not immediately obvious.
The problem isn't just about implementation or technology in isolation. It is about the fundamental challenge of trying to extract “neutral” insights from data that reflects a society full of historical and systemic biases. Even with perfect implementation, if your training data comes from a healthcare system where certain groups have been historically underserved, your AI will learn and amplify those patterns. If your data comes from a justice system with documented biases, those biases become baked into the model.
This isn't a technical problem that can entirely be solved with better code or more careful implementation. They will help, but ultimately it is a social problem that we're trying to solve with technology, and that's where things go wrong. The Dutch benefits scandal wasn't just a case of poor implementation, it was a case of an algorithm doing exactly what it was designed to do, while nobody fully understood the social implications of those design choices.
By centering the discussion about technology you are effectively still waving away practical concerns that are very much related to the technology. Specifically the fact that with ML these biases are not made more clear, but actually are hidden away as the reasoning process of the technology process is not visible. So yes, the technology itself also is part of the issue.
I think there's a significant difference between "this technology has limitations and issues that we know about and some we might not yet, but it can still give positive results if used correctly" and "waving away" those issues. I'm not unaware of the issues you mention. But I also think they are able to be accounted for and worked around with care and forethought. You said this was "incredibly difficult" not "impossible" so you may agree more with me than you think.
That sounds precisely like poor implementation to me, but clearly we differ on exactly what that means. :)
I am going to try to keep this short, as I am otherwise going in circles.
One thing I want to stress: the context here is us discussing a massive database incorporating all data into one system. There is simply no reasonable or feasible way to avoid bias in the input when dealing with data at this scale, coming from countless sources with different levels of accuracy and systemic biases. And again, we’re talking about technology that doesn’t just reflect those biases but enhances and obscures them in ways that even experts struggle to detect in much smaller systems.
As I think the overall point is made clear enough I will now bow out of this conversation.
I don't even like this idea on paper, it's ripe for abuse. Also, Larry Ellison should go take a bath with a toaster.
Imagine an LLM jailbreak that outputs everyone's social security numbers
The cool part is that you don't even need an LLM jailbreak to get that data! Your data is probably already floating around in various data breach dumps. Isn't it great how unseriously infosec is taken even by places that give you no other choice, like credit reporting agencies?
Isn't that the plot of Person of Interest? A "jailbreak" (backdoor) in an AI that spits out SSN of people in danger but not considered relevant to national security?
Oligarch thought of a new way to get even more money and power.
Next he’ll decide Oracle needs to control the world’s water supply.
Can't, Nestle already owns that.
He'd have to fight Nestle for the water
I find it funny in a sort of somber wistful way that we thought of the same joke.
I love you guys, and wish I could kiss both of you.
: ) Hello friend.
;)
Kinda weird how their vision for the best path to take that'll make everyone's lives better, more secure and more convenient just so happens to be one that personally enriches them beyond even their already ridiculous net worth.
I initially read that as "world's voter supply", and you know where that ends up...
This way lies social credit scoring, permanent loyalty and criminal records, chilling levels of algorithmic discrimination, and all the other virtues of the Panopticon.
I remember reading a short story in an anthology about AI circa 2000 that involved a new supercomputer being demonstrated to the government. Best part was when it made some sympathetic comments about socialism (making reference to early Christian communes) and pissed off a right-wing general. One wonders what the powers that be would do when their godlike AGI ingests All The Data only to output such solutions as "INSTITUTE UNIVERSAL HEALTHCARE" and "DECARBONIZE NOW YOU DOLTS".
Doesn't AI already have a ton of this information? Why would we need to "feed" it national security clearance level information to get better ideas on how to spend money? The bills and spending is already public information and there's been a ton of studies about so much stuff online as well, and these companies have no issue stealing so Im sure a lot of them are already baked in.
The other thing is the questions you ask. Even something that sounds simple isn't " How can we make America better" is so reliant on who defines 'better'.
There's a ton of ideas and concepts to try that are likely to make the country better but it doesn't align with corporate donors, lobbyists, high profile donors and it doesn't benefit politicians.
To me this is like asking AI to add 2+2 instead of just using a calculator to do it yourself.
The US government probably already has this running somewhere, and some "really smart" 19 year old in Musk's DOGE is copying it over to X servers right now, because it's "chock-full of fraud and inefficiency"
I for one trust Mr. BIG BALLS not to do anything shady with my private data.
Oh wait, never mind, I guess BIG BALLS quit over racist remarks he made on social media.
Oh wait, never mind x2, I forgot that being racist is probably a pre-requisite to work for DOGE anyways.
On the one hand, aggregating data for analysis does make sense, to a point. Talking about genomic data as part of individualized care doesn't really make sense in a large data set. For those purposes you typically do separate analysis of genomes, and then introduce categorical variables in a larger dataset. So there is no reason that the meta/multimodal model needs raw generic data. It would probably be hindered by it.
There are some exciting clinical trials underway that use a combination of biomarkers and genetic data to drive healthcare, including truma care. Devices that pull several ml of blood per day, perform assays and genetic sequencing, and give recommendations to clinicians on care, such as when to close wounds to minimize sepsis, or how to treat burns. These are real, happening today, examples of deep learning improving standards of care. No mega centralized system required.