Meta addresses AI hallucination as chatbot says Trump shooting didn’t happen
arstechnica.com
Meta "programmed it to simply not answer questions," but it did anyway.
You are viewing a single comment
Meta "programmed it to simply not answer questions," but it did anyway.
You can prove some things are correct, like math problems (assuming the axioms they are based on are also correct).
You can't prove that things like events having happened are correct. That's even a philosophical issue with human memory. We can't prove anything in the past actually happened. We can hope that our memory of events is accurate and reliable and work from there, but it can't actually be proven. In theory everything before could have just been implanted into our minds. This is incredibly unlikely (as well as not useful at best), but it can't be ruled out.
If we could prove events in the past are true we wouldn't have so many pseudo-historians making up crazy things about the pyramids, or whatever else. We can collect evidence and make inferences, but we can't prove it because it is no longer happening. There's a chance that we miss something or some information can't be recovered.
LLMs are algorithms that use large amounts of data to identify correlations. You can tune them to give more unique answers or more consistent answers (and other conditions) but they aren't intelligent. They are, at best, correlation finders. If you give it bad data (internet conversations) or incomplete data then it at best will (usually confidently) give back bad information. People who don't understand how they work assume they're actually intelligent and can do more than this. This is dangerous and should be dispelled quickly, or they believe any garbage it spits out, like the example from this post.
You can't so solidly that this shouldn't even be discussed.
What should be is whether you can make a machine capable of reasoning.
There's symbolic logic, so you can maybe some day make a machine that makes correct syllogisms, detects incorrect syllogisms and such.
Sadly there's that archetype of "the narrow-minded not cool scientist against the cool brave inventor" which means that actively dispelling that may do harm. People who don't understand will match the situation with that archetype and it will reinforce their belief.
Well but this kind of correctness applies to everything. By thag logic, you can't believe anything. I'm talking about an entirely different correctness. Like resistance against certain adversarial attacks. Of course, proving that the model is always correct, is as complicated as modelling the entire reality. That's infeasible. But it's also infeasible for every other software.
This sounds like an overly pedantic view of "prove"
It's not pedantic. You can mathematically prove math.
You can't mathematically/algorithmically prove an event happened or did not happen.
Adding "mathematically/algorithmically" in front of the word "prove" as if it were always implicitly there, and suggesting that it's the only way we should be using the word "prove" seems pretty darned pedantic to me.
We're describing the behavior of software. It must be implicitly included. Software cannot do anything that isn't algorithmic.
You can prove mathematical logic and you can (not 1-to-1) tie that to symbolic logic, but since it's not 1-to-1, because of ambiguity of symbols, there will be much more complexity. I personally think that the future of various machine assistants lies there, and what LLM's now do is going to be used in auxiliary roles for that.
The problem is that mathematical proofs rely on the basic premise that the underlying assumptions are rock solid, and that the rules of the math are rock solid. It's rigorous logic rules, applied mathematically.
The real world is Bayesian. Even our hard sciences like physics are only "mostly" true, which is why stuff like relativity could throw a wrench in it. There's inherent uncertainty for everything, because it's all measurement based, with errors, and more importantly, the relationships all have uncertainty. There is no "we know a^2 and b^2, so c^2 must be this". It's "we think this news source is generally reliable and we think the sentiment of the article is that this crime was committed, so our logical assumption is that the crime was probably committed". But no link in the chain is 100%. "Rock solid" sources get corrupted, generally with a time lag before it's recognizable. Your interpretation of a simple article may be damn near 100%, but someone is still going to misread it, and a computer definitely can.
Uncertainty is central to reality, down to the fact that even quantum phenomena have to be talked about probabilistically because uncertainty is built in all the way down.
This is why many philosophers came to criticize metaphysical logic in the 1800s, viewing it as dealing with absolutes when reality does not actually exist in absolutes, stating that we need some other logical system which could deal with the "fuzziness" of reality more accurately. That was the origin of the notion of dialectical logic from philosophers like Hegel and Engels, which caught on with some popularity in the east but then was mostly forgotten in the west outside of some fringe sections of academia. Even long prior to Bell's theorem, the physicist Dmitry Blokhintsev, who adhered to this dialectical materialist mode of thought, wrote a whole book on quantum mechanics where the first part he discusses the need to abandon the false illusion of the rigidity and concreteness of reality and shows how this is an illusion even in the classical sciences where everything has uncertainty, all predictions eventually break down, nothing is never possible to actually fully separate something from its environment. These kinds of views heavily influenced the contemporary physicist Carlo Rovelli as well.
You are describing LLMs, yes. But not what I'm describing.
I'm talking about machine finding syllogisms and checking their correctness. This can't be rock solid because of interpretation of the statement in natural language with its fuzzy semantics, but everything after that can be made rock solid. While in LLMs even it isn't.
That's what I'm talking about.
Humans make mistakes, but not such as LLM-generated texts contain.
I mean that one can build a reasoning machine which an LLM isn't.
I'm not describing LLMs. LLMs are completely irrelevant, and my examples had nothing to do with LLMs.
Formal logic requires propositions be Boolean in nature. They're true, or they're false.
That's not the real world. There are no booleans in the real world. In the real world, everything, down to the fundamental particles, is inherently probabilistic.
Our "certainty" is at most 99. a lot of 9s. It's never 100%. You can't say "the New York Times said X", and "the New York Times is perfectly reliable", so "X must be true". It's "given that the NYT said X and the NYT has a history of reporting facts with reasonably high accuracy, the probability X is true is...". If they get caught being shady, the estimates of previous information learned from them is retroactively changed. But there is no "proof", because there is no certainty anywhere in the chain. The world and human understanding of it has to be Bayesian. Again, down to the Uncertainty Principle about low level particles. Uncertainty is fundamental to reality. There is no certainty.
Why are you writing this to me?
Do you know what a syllogism is?
It doesn't require being certain of the information we're building it on. Only of existence of such categories.
Naturally people in Antiquity and Middle Ages who used symbolic logic were even less certain of the actual truths and lies in the world than we are.
It allows the truth to be subjective, but not the logical constructions. This is a very important trait both then and now.
The difference between the filter and the data going through it.
Of course you can't just feed all the data of all the PoVs and similar cases on something, integrate it into a model and expect your PoV to not clash with its output.
It's philosophically the same as why using dialectics is bad for science.
A syllogism is a tool for theoretical reasoning that doesn't actually apply in the real world, because it relies on Boolean possibility spaces. There is never an "all articles by X are correct", and there is no theoretical possibility that "all articles by X are correct" in the real world. The connections in the real world are literally always probabilistic. In every case. Every time.
You can't use formal logic for any real world use case because there are no valid starting assumptions. The only thing logic can ever prove is internal consistency, not fact.
Yes, and being able to build structures with internal consistency would be an advantage.
Nobody says you can prevent any "AI" oracle from saying things that aren't true.
But a tool which would generate a tree of possible logical conclusions from something given in language and then divided into statements on objects with statistical dependencies could be useful.
No. It's just pure math and logic. And LLMs are nothing more than billions of additions and multiplications. Literally. You can prove certain things on it just like you can prove theorems in mathematics. It's an ongoing ressearch field.
Okay: using additions and multiplications prove the assassination attempt on Donald Trump happened
How would you even prove something like that outside of LLMs? What is your point? That you cannot prove anything except "I think therefore I am"?
Either you haven't read my comments or you're intentionally trying to be provocative.
My point is what OPs point was (which you veered away from in order to try to show off that You Are Very Smart): it is literally impossible for a computer system to prove a historical event has happened.
I'm having a hard time keeping track of all of the threads and replies evolving here. Forgive me. But I assume you mean the followong one?
This is simply a wrong statement. You can indeed prove certain properties on these models. That implies of course that you're able to formulate that property fully.
I don't know why the discussion went this far off track. The main point though is that everyone including OP is trying to discredit AI by bringing up things it was never supposed to be good at. By design, it's not good at knowledge retrieval. But everyone is hating it because it's hallucinating fake news. It's beyond me why people argue like that.
Okay, how does the model prove the assassination attempt happened? Because that is what OP was talking about.
It was clear from the context that OP was saying "It is impossible to mathematically determine if something [historical] is correct." They omitted one word and instead of using context clues you went into a long unnecessary post on how we prove even numbers are divisible by 2. If you tried Iron Manning their post instead of trying to show off with an "Um Actually...." You wouldn't be getting lost in the replies as we'd be staying on the original topic.
We're missing the context again. It's not people trying to discredit AI. People are trying to discredit companies insisting on using AI for things it is bad at.
It sounds like you actually agree with OP: AI should not be used for this purpose. Instead of saying "I agree, this is a bad use of AI, it should only be used for X, Y, and Z" you felt the need to White Knight for AI. The problem right now isn't AI being attacked, it's companies treating AI like a miracle that can do everything.