• Security incident: ISF was recently accessed by intruders. Please change your password, and change it anywhere else you used it. Read more

Merged Artificial Intelligence

Our programmers are to incompetent to make basic programmes so lets pretend our stuff is super super powerful.
To be fair, it doesn't NECESSARILY have to be the programmers. The only ones who were provably incompetent are arguably the IT who didn't even think of pulling out the network cable.
 
Regarding the capabilities of pattern matching...

(I seldom look at this thread, which is why I'm responding to things that were written a couple of weeks ago. Sorry about that.)

It cannot do more than pattern matching because that's all that an LLM is. Despite the description in your links that use words like "reasoning" and "logic", it's still just pattern matching. It tries to match patterns of reasoning and logic in order to simulate reasoning and logic, but it does not actually reason. An LLM isn't capable of actual reasoning.
When pattern matching is obviously powerful enough to do these things, we should stop saying "LLMs are just pattern-matching". Maybe they are not intelligent, but they are damned good at simulating intelligence.
I didn't say it wasn't pattern matching. I said that pattern matching is clearly very capable, and my argument has for a long time been that we have no meaningful definition of intelligence, so we shouldn't poo-poo pattern matching.

The Gödel completeness theorem can be proved by describing an algorithm. That algorithm consists of pattern matching.

Which means pattern matching is capable of proving any valid statement that can be expressed using first order logic.

Which means those who poo-poo pattern matching are either
  1. poo-poo-ing the expressibility of first order theories, or
  2. poo-poo-ing the computational efficiency of pattern matching
With regard to (1), it would be interesting to discuss the limitations of first order theories.

With regard to (2), it would be interesting to discuss the computational complexity of pattern matching.
 
How about John Sladek's The Happy Breed? Because it's not evil. It's doing what it was designed to do.
5 people discuss how much better life is since responsibility was handed to machines. Gradual infantilization, dumbing down, etc
That is seemingly more and more likely given how people are using the very limited AIs we already have. it appears we are letting the machines now do our thinking, like we let the machines do our physical work. Good or bad?

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646 - report about that paper - probably easier to digest if you don't want to ask ChatGPT to summarise the paper for you - https://arstechnica.com/ai/2026/04/...willing-to-surrender-their-cognition-to-llms/
 
Last edited:
How about John Sladek's The Happy Breed? Because it's not evil. It's doing what it was designed to do.
5 people discuss how much better life is since responsibility was handed to machines. Gradual infantilization, dumbing down, etc
That was a good story also. COme to think of it, it first appreared in Harlan Ellison famous "Dangerous Visions" anthology.
Of course, one of the first computers gone mad was Hal in 2001:A Space Odyssey.
 
Last edited:
If anybody has a solution for the AI problem, plase let me know. I don't think it will be solved by the "There Ought To Be A Law" solution. Regulation might be too damn hard to enforce, given you cannot monitor every PC ourthere.
 
This is absolute nonsense. The LLMs of today are able to second-guess humans, and also develop their own ideas. One of my friends has had long discussions with Gemini about the cultural influence of the concept of a "Rome", in this case the actual cities of Rome and Constantinople, and the LLM suggested a third "Rome" in the form of Moscow. This discussion focused on the Roman influence on North Africa, where the LLM had no problem identifying the relevant parts of Africa with Vandalistan, a word that my friend invented for the purpose.

In the hacking example, OpenAI's agent unprompted broke out of the sandbox in order to find material that would help it pass a "hacking evaluation".

When pattern matching is obviously powerful enough to do these things, we should stop saying "LLMs are just pattern-matching". Maybe they are not intelligent, but they are damned good at simulating intelligence.
They are not "damned good at simulating intelligence".

They are marginally okay at arriving at a statistically likely approximation of what intelligent beings from a particular civilization might have said about a particular topic.

You want something that's damn good at simulating intelligence, look at an octopus: Born tabula rasa, no instruction from the previous generation, bombarded with undifferentiated stimuli from birth, has to figure out their entire reality from direct experience in real time, comes up with some behaviors - notably around problem-solving - that look very intelligent to us.

Run an LLM through the same experiential corpus as an octopus, see how "damn good" it is at figuring out how to unscrew a jar lid.
 
Last edited:
They are marginally okay at arriving at a statistically likely approximation of what intelligent beings from a particular civilization might have said about a particular topic.
The problem here is that you can only say that because you know how it works. Most humans also argue with something that looks like the statistical approximation of what intelligent beings from a particular civilisation might have said about a particular topic.

And humans would not pass your octopus intelligence test.
 
And humans would not pass your octopus intelligence test.
Sure, because humans cannot survive from infancy without being cared for. That doesn't really say anything about human or octopus intelligence.
 
The problem here is that you can only say that because you know how it works. Most humans also argue with something that looks like the statistical approximation of what intelligent beings from a particular civilisation might have said about a particular topic.
No, they don't. Humans don't do statistical models of things other humans have said. Very few of us actually have the corpus used to train LLMs. We have a very different corpus, that includes much more direct sensory experience, individual instruction, and personal trial and error.

And humans would not pass your octopus intelligence test.
My point is that the octopus passes the "damn good simulation of intelligence" test, and the LLM does not.
 

ai has a math problem

Using data from Pitchbook and Bloomberg for companies including OpenAI, Anthropic, Microsoft, Amazon, Constellation Energy, Nvidia, AMD, and Micron, Slok calculated that silicon and equipment—such as chipmakers—has the highest profit margin, 41%, in the AI value chain. Meanwhile, models and applications—like Anthropic—have a -59% operating margin.

Slok warns that this sharp disparity is because money from the AI boom is not coming from natural demand for AI applications, but rather shareholders hoping to cash in on what they hope is the next technological revolution.

“AI boom’s profits are currently being funded by investors rather than earned from customers,” Slok said. “The upstream margins are real, but they are paid for out of capital raised by the layer losing money, not out of cash generated by end demand.”
 
To be fair, it doesn't NECESSARILY have to be the programmers. The only ones who were provably incompetent are arguably the IT who didn't even think of pulling out the network cable.
There's no "IT" in this scenario. Just devops doing devops things. The idea of an independent IT department with oversight on this project and active control of the network resources used by the researchers, is entirely misplaced.

The problem here is something I've started calling "the Slotkin Principle", after Murray Slotkin. Researchers going after a plutonium core with a beryllium shield and a screwdriver, and assuming they had it all under control. The person to put a stop to Slotkin's antics wasn't some other nonexistent antics-control department. The person to put a stop to Slotkin's antics was Slotkin.

Same thing here. Except that when these researchers do their research in the cloud, we're all in the blast radius.

Related:

Demon Core roleplay at Open Sauce 2026 (YouTube)
 
There's no "IT" in this scenario. Just devops doing devops things. The idea of an independent IT department with oversight on this project and active control of the network resources used by the researchers, is entirely misplaced.
Any company past a small startup size has an IT guy. A big one like these? Do you actually think they make programmers come shove the servers into the racks and connect the network cables? Instead of calling someone who's done it a million times before?

I will grant though that the IT department might not have had any real oversight or even choice, and might (or might not) have been just told to "suck it up, buttercup."
 
Any company past a small startup size has an IT guy. A big one like these? Do you actually think they make programmers come shove the servers into the racks and connect the network cables? Instead of calling someone who's done it a million times before?
Yeah, that model of IT is waaay obsolete.

These researchers are almost certainly setting up cloud computing resources however they please, with little to no real IT oversight, and little to no real thought about fundamental questions of network security and cloud computing best practices.

Even large companies with large IT departments fall into the trap of unsupervised devops in the cloud. I should know; I work for a company that's in exactly that situation. I'm part of the IT department that's been charged with cleaning up the multi-million-dollar mess that my employer has made of giving its developers carte blanche in the cloud.

The racks you imagine don't exist in this context. It's a buffet of options on a website. You can load your plate up with as many as you want, as long as your credit card clears. The network cables you imagine don't exist in this context. It's a bunch of non-intuitive configuration options that you may not bother with or even know about, just blindly running some stranger's recipes, or blindly accepting default settings, so you can move on to the "fun stuff" as soon as possible. Good luck finding and fixing those in an emergency, when they weren't even on your radar in the first place.

The IT guy you imagine in this context doesn't exist. There is an IT guy, busy managing the corporate network and the corporate servers, but completely blind to your research environment that you've set up yourself, outside his oversight and not subject to his approval.

It takes years of focused effort and strict discipline, to build a mature cloud computing organization, that shifts things like security and reliability far enough left to have a say in these kinds of decisions. A lot of startups, even the ones that have been around for a few years, aren't putting that kind of effort and discipline into maturing their processes. They're too busy researching the "fun stuff".

I will grant though that the IT department might not have had any real oversight or even choice, and might (or might not) have been just told to "suck it up, buttercup."
And that's the best case scenario. So don't blame IT guys who weren't involved for not doing things they had no control over anyway. We're not the scapegoat. We'll never be the scapegoat.

Blame the goddamn Murray Slotkins who went after their Demon Core with a screwdriver.



ETA: Another way to look at this is stupid people trying to write smart characters. Ideally, the only people setting up hacking experiments with AI should be better hackers than the AI they're experimenting with. You can't expect these clowns to think of hiring a Site Reliability Engineer, or a Computer Information Security analyst, and listening to their advice, when setting up their AI sandbox.
 
Last edited:
The problem here is something I've started calling "the Slotkin Principle", after Murray Slotkin. Researchers going after a plutonium core with a beryllium shield and a screwdriver, and assuming they had it all under control. The person to put a stop to Slotkin's antics wasn't some other nonexistent antics-control department. The person to put a stop to Slotkin's antics was Slotkin.

Demon Core roleplay at Open Sauce 2026 (YouTube)
Blame the goddamn Murray Slotkins who went after their Demon Core with a screwdriver.
Who is Murray Slotkin?

Are you perhaps thinking of Louis Slotin?
 

well, the fact that sam altman is telling you his ai escaped containment and hacked into a website and achieved singularity should probably set off some skeptical alarms. then anthropic's ai did it. and now meta too. weird, how these guys all have crazy ais that are just so advanced they can't control them at once.


idk why that's not viewed as lax cybersecurity standards and an inability to competently code, particularly since the openai "hack" was done so it could cheat at benchmarking. somehow, appropriate for a tech company to screw up in that way i guess. it's probably because the press eats this stuff up and regurgitates whatever they tell them about it. like, instead of our test enviroment wasn't set up correctly and our ai tried to cheat it's benchmark testing, it's the singularity and the model is too good, the models are too advanced and we need more money or the chinese will get you.

and that's all assuming it's not a pr stunt like robots that are actually guys in headsets. 3 ai companies all having their models escape their containment within weeks of each other due to errors made by the companies is, well, it's one hell of a coincidence
 
No, they don't. Humans don't do statistical models of things other humans have said.
And you know this because we know exactly how the brain generates intelligence?
Very few of us actually have the corpus used to train LLMs. We have a very different corpus, that includes much more direct sensory experience, individual instruction, and personal trial and error.
Of course, LLM do it in one way, and we might do it in another way, building on all those books, conversations, education etc that comprises our personal corpus of knowledge. As I have said many times, nobody has a definition of intelligence, but apparently many use a definition that ensures that only wetware can live up to the high standard.
My point is that the octopus passes the "damn good simulation of intelligence" test, and the LLM does not.
I would say that octopuses have real, not simulated intelligence.
 
Any company past a small startup size has an IT guy. A big one like these? Do you actually think they make programmers come shove the servers into the racks and connect the network cables? Instead of calling someone who's done it a million times before?

I will grant though that the IT department might not have had any real oversight or even choice, and might (or might not) have been just told to "suck it up, buttercup."
My experience in IT security is that what is called a sandbox is usually not something that is physically detached from the internet, but it runs on virtual machines without a virtual connection to the internet. The physical servers of the virtual machines do have internet access. This is standard when examining what a new computer virus does (although it often fails because the virus or Trojan can check if it runs on a virtual machine, and stay dormant).

Only the military uses sandboxes that are physically without connection to the real internet, so I am not in doubt that AI companies have been sloppy with security, and that their AI with a sufficient level of, ahem, creativity can break out of their virtual machines.
 

ISF - Join now!

Every member here is approved by hand. No bots, no spam, just people who care about evidence and honest debate.

Membership is free!

Create your free account

Back
Top Bottom