Meta “programmed it to simply not answer questions,” but it did anyway.
AI doesn’t know what’s wrong or correct. It hallucinates every answer. It’s up to the supervisor to determine whether it’s wrong or correct.
Mathematically verifying the correctness of these algorithms is a hard problem. It’s intentional and the trade-off for the incredible efficiency.
Besides, it can only “know” what it has been trained on. It shouldn’t be suprising that it cannot answer about the Trump shooting. Anyone who thinks otherwise simply doesn’t know how to use these models.
It is impossible to mathematically determine if something is correct. Literally impossible.
At best the most popular answer, even if it is narrowed down to reliable sources, is what it can spit out. Even that isn’t the same thing is consensus, because AI is not intelligent.
If the ‘supervisor’ has to determine if it is right and wrong, what is the point of AI as a source of knowledge?
We should understand that 99.9% of what wee say and think and believe is what feels good to us and we then rationalize using very faulty reasoning, and that’s only when really challenged! You know how I came up with these words? I hallucinated them. It’s just a guided hallucination. People with certain mental illnesses are less guided by their senses. We aren’t magic and I don’t get why it is so hard for humans to accept how any individual is nearly useless for figuring anything out. We have to work as agents too, so why do we expect an early days LLM to be perfect? It’s so odd to me. Computer is trying to understand our made up bullshit. A logic machine trying to comprehend bullshit. It is amazing it even appears to understand anything at all.
That is, unless you define correct in mathematical terms. Which no one has done yet.
It is impossible to mathematically determine if something is correct. Literally impossible.
No, you’re wrong. You can indeed prove the correctness of a neural network. You can also prove the correctness of many things. It’s the most integral part of mathematics and computer-science.
For example a very simple proof: with the conjecture that an even number is 2k of a number k, then you can prove that the addition of two even numbers is again an even number (and that prove is definite): 2a+2b=2(a+b), since a+b=k for some k.
Obviously, proving more complex mathematical problems like AI is more involved. But that’s why we have scientists that work on that.
At best the most popular answer, even if it is narrowed down to reliable sources, is what it can spit out. Even that isn’t the same thing is consensus, because AI is not intelligent.
That is correct. But it’s not a limitation. It’s by design. It’s the tradeoff for the efficiency of the models. It’s like lossy JPG compression. You accept some artifacts but in return you get much smaller images and much faster loading times.
But there are indeed "AI"s and neural networks that have been proven correct. This is mostly applied to safety critical applications like airplane collision avoidance systems or DAS. But a language model is not safety critical; so we take full advantage.
If the ‘supervisor’ has to determine if it is right and wrong, what is the point of AI as a source of knowledge?
You’re completely misunderstanding the whole thing. The only reason why it’s so incredibly good in many applications is because it’s bad in others. It’s intentionally designed that way. There are exact algorithms and there approximation algorithms. The latter tend to be much more efficient and usable in practice.
Your proof example is a proof from your discrete structures class. That’s very different than “proving” something like “the Trump assassination attempt was a conspiracy.”
Otherwise we could have gotten rid of courts a long time ago.
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.
The only reason why it’s so incredibly good in many applications is because it’s bad in others. It’s intentionally designed that way.
lolwut
