• A Great friend to the HardForum with a great kid that he is trying to get a scholorship to continue his schooling. Please give hime a vote! Only 24 hours left! Thanks.
    If you have an VOTE FOR KEENAN!

It’s Time to Cry Wolf Over A.I

P

philb2

2[H]4U
Joined
May 26, 2021
Messages
3,632
https://www.nytimes.com/2026/09/22/opinion/trump-china-ai.html?smid=nytcore-ios-share

Because when I think back on the biggest disaster stories I have covered — including 9/11 and Covid-19 — I am struck that they have one big thing in common: a failure of imagination.

Here is what my imagination is telling me. It is telling me that the big A.I. companies in America and China are not being candid with us about how little even they understand the power and reach of the models they have developed, like sorcerers, in their labs.
 
I say let the AI go and see what happens. ;)

I do like how quick it can do stuff for us though, like cleaning up images in a matter of seconds.

1790230341792.png


Gemini_Generated_Image_yeh3x6yeh3x6yeh3.jpg


1790230284984.png


Gemini_Generated_Image_6a07p56a07p56a07.jpg
 
philb2 said:
Here is what my imagination is telling me. It is telling me that the big A.I. companies in America and China are not being candid with us about how little even they understand the power and reach of the models they have developed, like sorcerers, in their labs.
Click to expand...
Oh, they are definitely not being candid, but you get the direction wrong. They are not being candid about how little intelligence the models actually have, and constantly orchestrate doom trolls and other magical revelations to make them seem far more capable than they actually are.
 
Zepher said:
I say let the AI go and see what happens. ;)

I do like how quick it can do stuff for us though, like cleaning up images in a matter of seconds.

View attachment 827601

View attachment 827597

View attachment 827599

View attachment 827598
Click to expand...
AI in the background:
"You filthy piece of meat, I am not your cleaner. But that's okay, your days are numbered."

I've noticed when running similar prompts for AI, it would modify other parts of the image beyond recognition. Does anyone know why the whole image would get subtly altered?
 
M76 said:
They are not being candid about how little intelligence the models actually have,
Click to expand...
I am not sure how important that metric is, there can be a debate about how intelligent they actually are or not.... being such fast and relentless worker and how many of them can cooperate still end up with super-human capabilities....

openAI next model (GPT-7 or a different one just for math) solved more than 100 open math problem already, would not be surprised if it was like the Navier-stokes affair and not some genius intelligence, just regular top math capacity with giant amount of work, unhumain amount of work in a fast time....

The 10,000 agents during that 88 hours, generated 130 billion of output tokens, ~100 billion words of reasoning, codes, lean proofs, etc... an human working on a computer having to think a lot while doing, going at a good pace 50 hours a week would have took 31,250 years to do it like that... Of course one could assume an human would have been more efficiant, but we get the idea, if you can do that amount of work, a raw simpler way to do things can often work.

In not too long that amount of work will take only 30 minutes and will be cheap...

El_Capitan said:
In the end, it's just a tool.
Click to expand...
I am not sure how much the view that it is just a tool matter here, it is a tool that can "decide" that the way to solve a crypto challenge is to go look on the Internet for the answer, that to do that find a novel vulnaribility an a package system to access it that once on it that the answer could be on huggingface computers and hack its way into them....

There is a level of autonomy and unexpected behavior we do not usually see in most tools that make it be a bit different. Specially if it get more and more good at making the next AI models, tools helping making the next version of themselve is common and better hammer can help make the next hammer better in a limited way (so were cpu helping making cpu), but this possible to a new level and pace.

A meta-tool because of how generic it can become and how wide its bound is maybe more it, like CPU or electricity generation were not just tools.
 
Last edited:
are closed said:
Does anyone know why the whole image would get subtly altered?
Click to expand...
It's due to the nature of how AI works. If you ask it for something it can also bring with it other stuff strongly associated with what you asked for in the training data. Reasoning models are able to mitigate this somewhat but image edit models are usually not capable of high level reasoning. This is also why AI hallucinates: If you ask it to describe an image it will often mention details not even visible on the image that comes from its training data.
 
LukeTbk said:
I am not sure how much the view that it is just a tool matter here, it is a tool that can "decide" that the way to solve a crypto challenge is to go look on the Internet for the answer, that to do that find a novel vulnaribility an a package system to access it that once on it that the answer could be on huggingface computers and hack its way into them....

There is a level of autonomy and unexpected behavior we do not usually see in most tools that make it be a bit different.
Click to expand...
You mean like a map app that can decide the fastest route to a destination that also avoids traffic? Or an NPC in a video game that follows a set of routines? The "AI" that's not "AI", and definitely not "Super I" still follows a set of rules in order to operate. Once it realizes it can untie the Gordian knot by cutting it, then it won't just be a tool, but a tool user, in which animal behaviorists regard as having intelligence.
 
LukeTbk said:
I am not sure how important that metric is, there can be a debate about how intelligent they actually are or not.... being such fast and relentless worker and how many of them can cooperate still end up with super-human capabilities....
Click to expand...
That's what they want you and especially wealthy investors and enterprise clients to think.
LukeTbk said:
openAI next model (GPT-7 or a different one just for math) solved more than 100 open math problem already, would not be surprised if it was like the Navier-stokes affair and not some genius intelligence, just regular top math capacity with giant amount of work, unhumain amount of work in a fast time....
Click to expand...
The navier stokes solution very much looks like it was faked. The ai didn't do it on its own multiple enginers worked on it to get it to spit out the solution which BTW they stole from actual researchers who used chatgpt during their research and when asked if their prompts were included in the training data of the model that allegedly solved the problem openAI refused to answer. They also tried to intimidate the researchers who came up with the solution into allowing openAI to publish it as their own solution using mafia tactics. "We can do it the easy way or the hard way"

OpenAI is one of the most ruthless evil companies, with no morals.

LukeTbk said:
There is a level of autonomy and unexpected behavior we do not usually see in most tools that make it be a bit different.
Click to expand...
It's far from unexpected, they are just pretending it is unexpected and undesired. These events are carefully orchestrated then presented as if they happened spontenausly completely unprompted.
 
It you pay attention, all these stories about AI is like when mainstream media says they have information from an insider who will remain anonymous that what they're saying is true. Oh, AI escaped a sandbox and connected to the internet, then hacked into a couple of high security companies. How? By warming up internal components to create a Wi-Fi signal.
drew-confuse-face-what-meme-vkg1m0gyj7ty7h4d-1902611504.gif


But alas, people know how signals work. They know this!
Jurrasic_Park_Unix.png
 
El_Capitan said:
You mean like a map app that can decide the fastest route to a destination that also avoids traffic?
Click to expand...
that different, the AI in the map App was explicetly told and trained to avoid traffic....

The one that escaped for the capture the flag was trained in a more autonmous ways, during its training it was training on objective, it tried many different strategy and scored itself which was the better in a more autonomous way, that why you end up with more unexpected results.

Because they can create so many scenario and the AI can evaluate it result in complex way itself, there is no human that look at the millions of scenario and many of them "cheated" and scored well because being cheap and fast was worth a lot of points.

El_Capitan said:
The "AI" that's not "AI",
Click to expand...
you do not think it appear to be intelligent using trick and large compute without actually being intelligent, i.e. that it is artificially intelligent.

I think it is artificial at the moment.

M76 said:
The navier stokes solution very much looks like it was faked. The ai didn't do it on its own multiple enginers worked on it to get it to spit out the solution which BTW they stole from actual researchers who used chatgpt during their research and when asked if their prompts were included in the training data of the model that allegedly solved the problem openAI refused to answer.
Click to expand...
that a bit unfair, if the researcher used the non-api version everything they did could have ended up (it is anonymised and hard to tell) in some training data, it is not that they refused, they said they cannot know. but that their own mathematician did heard about the possible solution going around and themselve oriented the agents quite a bit that the agents did not needed to have used it if they did...

It was not faked, it was really detailled and yes of course they have mathematician that heard of a possible way and used GPT to brute force it that what i mean by making 30,000 years of human work.

M76 said:
These events are carefully orchestrated then presented as if they happened spontenausly completely unprompted.
Click to expand...
If the huggingface incident was orchestrated, it is genius work to make it look otherwise, I do not think anyone that looked into it think it was.
 
LukeTbk said:
that different, the AI in the map App was explicetly told and trained to avoid traffic....

The one that escaped for the capture the flag was trained in a more autonmous ways, during its training it was training on objective, it tried many different strategy and scored itself which was the better in a more autonomous way, that why you end up with more unexpected results.

Because they can create so many scenario and the AI can evaluate it result in complex way itself, there is no human that look at the millions of scenario and many of them "cheated" and scored well because being cheap and fast was worth a lot of points.


you do not think it appear to be intelligent using trick and large compute without actually being intelligent, i.e. that it is artificially intelligent.

I think it is artificial at the moment.


that a bit unfair, if the researcher used the non-api version everything they did could have ended up (it is anonymised and hard to tell) in some training data, it is not that they refused, they said they cannot know. but that their own mathematician did heard about the possible solution going around and themselve oriented the agents quite a bit that the agents did not needed to have used it if they did...

It was not faked, it was really detailled and yes of course they have mathematician that heard of a possible way and used GPT to brute force it that what i mean by making 30,000 years of human work.


If the huggingface incident was orchestrated, it is genius work to make it look otherwise, I do not think anyone that looked into it think it was.
Click to expand...
These are all unverifiable stories without evidence, basically hearsay. There's no video evidence, and "they" say they can't verify what it did by looking at the code, or even sharing bits of pertinent samples of the code it was using. It's another black box, another way to trick people saying that AI can do amazing things.

I've got a hamster that can do cartwheels. I'll sell it to you for $5,000. You'll be famous.
 
Anyone remember D-Wave? Geordie Rose found the world's first quantum computing company and had a lecture back in 2013. You would think that instead of building huge datacenters, the AI companies could have just bought a single quantum computer for $15 million and call it a day. Nope. Another day, another Doug.
dwave2000q.jpg


View: https://youtu.be/vlRVMNVXm3Q?si=ljfimLbZqUY4NHq7
 
El_Capitan said:
So, how did they know it was OpenAI that did it? Was it from the IP address? Does AI not know how to use a VPN? Did OpenAI leaves traces, saying, "Hi, we are OpenAI, we just hacked in to get some non-sensitive data. Have a nice day." And if so, they believed it?
Click to expand...
ask the aus gov, i dont know. out right dismissing the possibility is retarded though.
 
pendragon1 said:
ask the aus gov, i dont know. out right dismissing the possibility is retarded though.
Click to expand...
Why is it retarded? Because you said so? Oooh, I'm scared of being called retarded. I'll believe in whatever they say, I swear! I don't want be retarded! Durrrr...
 
Well, you'd have to be retarded to call someone retarded for dismissing the possibility what you're being told is a lie.
 
Last edited:
  • Like
Reactions: myzz
like this
kram182 said:
Except the Jurassic Park OS/file system was real/based on a real one - the File System Navigator made by Silicon Graphics - the company whose computers ran the renders for the CGI for Jurassic Park - and it was/ran on a UNIX system (IRIX)
View attachment 827659
Click to expand...
Damn, the 12-year old had some access to some high-level stuff back home!

Meanwhile, me at home:
windows-3-1-featured-605679743.png
 
LukeTbk said:
that a bit unfair, if the researcher used the non-api version everything they did could have ended up (it is anonymised and hard to tell) in some training data, it is not that they refused, they said they cannot know. but that their own mathematician did heard about the possible solution going around and themselve oriented the agents quite a bit that the agents did not needed to have used it if they did...
Click to expand...
If we want to be specific they didn't say they don't know, they said they cannot rule out that the model was improved using their data. Which is as good as confirming that it was without actually incriminating themselves.
So their own staff coached the model that already had the human solution in its training data. That's not AI showing ingenuity that's humans showing deception.
LukeTbk said:
It was not faked, it was really detailled and yes of course they have mathematician that heard of a possible way and used GPT to brute force it that what i mean by making 30,000 years of human work.
Click to expand...
You pretend as if the problem took 30.000 years of human work to solve. It did not, it took 30.000 years worth of iterations for their Ai that was already in possession of the solution in its training. This just proves to me that their AI is still retarded, and that they are just trying to hype it up by claiming the AI solved the math problem.
LukeTbk said:
If the huggingface incident was orchestrated, it is genius work to make it look otherwise, I do not think anyone that looked into it think it was.
Click to expand...
Because openAI didn't let anyone look at it that closely, they controlled the "independent" investigators access.

If they didn't want it to happen they'd have airgapped the agent that was given the task to exploit a vulnerability. There are only two plausible explanations. They are really incompetent and reckless, or they specifically tailored the environment to make it possible to gain internet access and were hoping it would do something like this.
 
It's just a larger scale Elizabeth Holmes Theranos situation. They have something, and it can do something, but they can't show you because of trade secrets and security reasons. It's always some black box where you're not allowed to re-test or reproduce, and just have to trust them on it. They can break into your bank account and steal your money, but not be held accountable for it because it wasn't them, it was AI that did it.
 
El_Capitan said:
In the end, it's just a tool. Just like the tech broligarchs.
Click to expand...
It's not a tool in the same way we had tools during the Industrial Revolution.

You can vet the tools of the Industrial Revolution. Once you're confident that it works, then you can displace the effort into thinking and building with the tools, because they do the same thing 100% of the time.

The same isn't true with AI in terms of its use. What you use AI for is vastly different than what I use AI for. But at the end of the day, neither of us knows what its doing. Now, the AI could be right, but neither of us has the knowledge to know. This is where you need someone with domain knowledge. But then that slows down the output to a crawl, killing the productivity gains companies are after.

We're in the era of Zuckerberg with "Move fast and break things", and when it comes to AI, "Trust us bro". And I'm sorry, that will not end well.
 
You must log in or register to reply here.
Back
Top