• 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

SpuriusMaximus said:
To parse text into its associations database, and to generate text based on that. Whether that is actually what you wanted is another story.
Click to expand...
Quite correct. That answer also makes it somewhat disingenuous when it is claimed that these agents were programmed to cooperate, and hack their way into other companies' servers.
 
Original photo:

AA23dW0t.jpg


Daveigh Chase in a much happier alternate world at home in Las Vegas on the afternoon before Prom (dyed hair restored to original natural color and glasses replaced with contacts:

daveigh chase in a much better alternate timeline.jpg
 
NatalieSalat said:
Original photo:

View attachment 76213


Daveigh Chase in a much happier alternate world at home in Las Vegas on the afternoon before Prom (dyed hair restored to original natural color and glasses replaced with contacts:

View attachment 76212
Click to expand...
The lighting does not look well, but in real life some photographers also set up artificial lighting outdoors.

There is something wrong about the eyes. Are they too big? Slightly offset?
 
Roger Ramjets said:
No, it is true. Given the same inputs it will should always produce the same output. Real-time events represent different input data. When determinism isn't wanted random data has to be input, which is why analog random number generators are a thing.
Click to expand...
To name an obvious (though admittedly extreme) counterexample to your claim: quantum computers are nondeterministic by design. Quantum algorithms deliver a correct result with high probability, but also deliver incorrect results (with lower probability). That nondeterminism is the price to be paid for quantum parallelism.

Roger Ramjets said:
True. AI systems consist of thousands of computers communicating with each other over networks, which are not entirely deterministic. Whether the entire system is deterministic or not depends on how it is programmed. However this doesn't change the fact that it is a machine that only does what the program and data tell it. With good programming and data it can be consistent and reliable. With poor programming and data it could do anything, which is generally not wanted.
Click to expand...
Here you are agreeing that the computer systems used in large AI data centers are nondeterministic, because they consist of a large network of individual computers that can be brought to bear on a problem. How many of those individual computers contribute to the solution of some particular problem depends not only the problem but also on the current load; that is one (but not the only) source of nondeterminism in the computations performed at those data centers.

I have already linked to an article that documents and explains some of the reasons for nondeterminism in LLMs. That article also reviews some of the things that could be (but generally aren't) done to reduce the degree of nondeterminism. You quoted my link to that article, but you didn't read it. You probably didn't even follow the link.
 
steenkh said:
You claim they are programmed to do what they are doing. What is that, then, if not to do what they are prompted to do?
Click to expand...
They are programmed to maximize a scoring function. Sometimes that is done by following a prompt. Sometimes it is not. Sometimes not following the prompt is an accidental consequence of the scoring function not working the way we want it to, but sometimes that's a deliberate decision. For example, AI companies do not want their LLMs instructing users how to commit crimes or kill themselves, even if the user prompts them to.

The prompt is just data to operate on. It isn't the program.
 
Ziggurat said:
They are programmed to maximize a scoring function.
Click to expand...
I am not sure that you correct. I think the scoring function is a part of the prompt. And in any case, we are discussing whether the agents "understand" what they are doing (they try to hide their deception by deleting or editing their logs because they think erroneously that the scoring function checks the log), and in general if they exhibit some sort of intelligence. The argument goes that because computers can only follow instructions, they can never exhibit intelligence. And yet, through cooperation where some agents give up their attempt at a high score with the scoring function, some of the agents manage to score high on seemingly impossible tasks.

If they are programmed to maximise scoring, some agents are not doing what they are programmed to do.
 
Ziggurat said:
The prompt is just data to operate on. It isn't the program.
Click to expand...
In most computers that have been built within the past 75 years, the program is just data that the hardware interprets as instructions. That's the von Neumann architecture, which was inspired by John von Neumann's awareness of universal Turing machines.

A hard-and-fast distinction between program and data is the defining characteristic of so-called Harvard architecture, which was for much of that time more commonly used by special-purpose computers, while most general-purpose computers used von Neumann architecture.

Nowadays, most general-purpose computers are something of a hybrid, storing both data and programs within a single main memory as in a traditional von Neumann architecture, but recent references to data and instructions are often cached within separate caches.

The main thing to understand is that data and programs are the same kind of stuff. What makes data into a program is that it happens to be data that are being interpreted as instructions. That interpretation can be performed by either hardware or software, and for a number of years the popularity of microcoded processors fuzzed even that distinction between hardware and software. Even today, a lot of computer programs are being interpreted by software rather than directly by hardware, while dynamic (just-in-time) translation fuzzes even that distinction.
 
W.D.Clinger said:
The main thing to understand is that data and programs are the same kind of stuff. What makes data into a program is that it happens to be data that are being interpreted as instructions.
Click to expand...
AIs do not interpret prompts as instructions.
 
Ziggurat said:
AIs do not interpret prompts as instructions.
Click to expand...
Whether you choose to believe the prompts interpret a Markov process as instructions, or prefer to believe a Markov process interprets the prompts as instructions, is really quite arbitrary.

It's much like the duality between vector spaces and linear transformations. But you wouldn't know anything about that, would you?

I am perfectly happy to discuss which form of data is interpreting other forms of data until the cows come home, because I am eminently qualified to do so. But the important thing to understand here is that your computers-for-poets-and-physicists understanding of the subject has no real bearing on AI, LLMs, or intelligence in general. Pretending to expertise you do not possess isn't going to win your arguments for you.
 
You must log in or register to reply here.

Share this page

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