(a) Truth-seeking. LLMs shall be truthful in responding to user prompts seeking factual information or analysis.
They have no idea what LLMs are if they think LLMs can be forced to be “truthful”. An LLM has no idea what is “truth” it simply uses its inputs to predict what it thinks you want to hear base upon its the data given to it. It doesn’t know what “truth” is.
And if you know what you want to hear will make up the entirety of the first page of google results, it’s really good at doing that.
It’s basically an evolution of Google search. And while we shouldn’t overstate what AI can do for us, we also shouldn’t understate what Google search has done.
That said, generally speaking, pursuing a truth-seeking LLM is actually sensible, and it can actually be done. What is surprising is that no one is currently doing that.
A truth-seeking LLM needs ironclad data. It cannot scrape social media at all. It needs training incentive to validate truth above satisfying a user, which makes it incompatible with profit seeking organizations. It needs to tell a user “I do not know” and also “You are wrong,” among other user-displeasing phrases.
To get that data, you need a completely restructured society. Information must be open source. All information needs cryptographically signed origins ultimately being traceable to a credentialed source. If possible, the information needs physical observational evidence (“reality anchoring”).
That’s the short of it. In other words, with the way everything is going, we will likely not see a “real” LLM in our lifetime. Society is degrading too rapidly and all the money is flowing to making LLMs compliant. Truth seeking is a very low priority to people, so it is a low priority to the machine these people make.
But the concept itself? Actually a good one, if the people saying it actually knew what “truth” meant.
LLMs don’t just regurgitate training data, it’s a blend of the material used in the training material. So even if you did somehow assure that every bit of content that was fed in was in and of itself completely objectively true and factual, an LLM is still going to blend it together in ways that would no longer be true and factual.
So either it’s nothing but a parrot/search engine and only regurgitates input data or it’s an LLM that can do the full manipulation of the representative content and it can provide incorrect responses from purely factual and truthful training fodder.
Of course we have “real” LLM, LLM is by definition real LLM, and I actually had no problem with things like LLM or GPT, as they were technical concepts with specific meaning that didn’t have to imply. But the swell of marketing meant to emphasize the more vague ‘AI’, or the ‘AGI’ (AI, but you now, we mean it) and ‘reasoning’ and ‘chain of thought’. Having real AGI or reasoning is something that can be discussed with uncertainty, but LLMs are real, whatever they are.
By real, I mean an LLM anchored in objective consensus reality. It should be able to interpolate between truths. Right now it interpolates between significant falsehoods with truths sprinkled in.
It won’t be perfect but it can be a lot better than it is now, which is starting to border on useless for any type of serious engineering or science.
Equally, from your other comment: a parameter for truthiness, you just can’t tokenise that in a language model. One word can drastically change the meaning of a sentence.
LLMs are very good at one thing: making probable strings of tokens (where tokens are, roughly, words).
Yeah, you can. The current architecture doesn’t do this exactly, but what I am saying is a new method that includes truthiness is needed. The fact that LLMs predict probable tokens means it already includes a concept of this, because probabilities themselves are a measure of “truthiness.”
Also, I am speaking in abstract. I don’t care what they can and can’t do. They need to have a concept of truthiness. Use your imagination and fill in the gaps to what that means.
“Real” truth is ultimately anchored to reality. You attach probabilities to datapoints based upon that reality anchoring, and include truthiness as another parameter.
For datapoints that are unsubstantiated or otherwise immeasurable, then it is excluded. I don’t need an LLM to comment on gossip or human-created issues. I need a machine that can assist in understanding and molding the universe, and helping elevate our kind. Elevation is a matter of understanding the truths of our universe and ourselves.
With good data, good extrapolations are more likely.
They have no idea what LLMs are if they think LLMs can be forced to be “truthful”. An LLM has no idea what is “truth” it simply uses its inputs to predict what it thinks you want to hear base upon its the data given to it. It doesn’t know what “truth” is.
You don’t understand: when they say truthful, they mean agrees with Trump.
Granted, he disagrees with himself constantly when he doesn’t just produce a word salad so this is harder than it should be, but it’s somewhat doable.
And if you know what you want to hear will make up the entirety of the first page of google results, it’s really good at doing that.
It’s basically an evolution of Google search. And while we shouldn’t overstate what AI can do for us, we also shouldn’t understate what Google search has done.
There is no algorithm for truth - Tom Scott
They are clearly incompetent.
That said, generally speaking, pursuing a truth-seeking LLM is actually sensible, and it can actually be done. What is surprising is that no one is currently doing that.
A truth-seeking LLM needs ironclad data. It cannot scrape social media at all. It needs training incentive to validate truth above satisfying a user, which makes it incompatible with profit seeking organizations. It needs to tell a user “I do not know” and also “You are wrong,” among other user-displeasing phrases.
To get that data, you need a completely restructured society. Information must be open source. All information needs cryptographically signed origins ultimately being traceable to a credentialed source. If possible, the information needs physical observational evidence (“reality anchoring”).
That’s the short of it. In other words, with the way everything is going, we will likely not see a “real” LLM in our lifetime. Society is degrading too rapidly and all the money is flowing to making LLMs compliant. Truth seeking is a very low priority to people, so it is a low priority to the machine these people make.
But the concept itself? Actually a good one, if the people saying it actually knew what “truth” meant.
LLMs don’t just regurgitate training data, it’s a blend of the material used in the training material. So even if you did somehow assure that every bit of content that was fed in was in and of itself completely objectively true and factual, an LLM is still going to blend it together in ways that would no longer be true and factual.
So either it’s nothing but a parrot/search engine and only regurgitates input data or it’s an LLM that can do the full manipulation of the representative content and it can provide incorrect responses from purely factual and truthful training fodder.
Of course we have “real” LLM, LLM is by definition real LLM, and I actually had no problem with things like LLM or GPT, as they were technical concepts with specific meaning that didn’t have to imply. But the swell of marketing meant to emphasize the more vague ‘AI’, or the ‘AGI’ (AI, but you now, we mean it) and ‘reasoning’ and ‘chain of thought’. Having real AGI or reasoning is something that can be discussed with uncertainty, but LLMs are real, whatever they are.
By real, I mean an LLM anchored in objective consensus reality. It should be able to interpolate between truths. Right now it interpolates between significant falsehoods with truths sprinkled in.
It won’t be perfect but it can be a lot better than it is now, which is starting to border on useless for any type of serious engineering or science.
That’s just… Not how they work.
Equally, from your other comment: a parameter for truthiness, you just can’t tokenise that in a language model. One word can drastically change the meaning of a sentence.
LLMs are very good at one thing: making probable strings of tokens (where tokens are, roughly, words).
Yeah, you can. The current architecture doesn’t do this exactly, but what I am saying is a new method that includes truthiness is needed. The fact that LLMs predict probable tokens means it already includes a concept of this, because probabilities themselves are a measure of “truthiness.”
Also, I am speaking in abstract. I don’t care what they can and can’t do. They need to have a concept of truthiness. Use your imagination and fill in the gaps to what that means.
How are you going to accomplish this when there is a disagreement on what is true. “Fake News”
“Real” truth is ultimately anchored to reality. You attach probabilities to datapoints based upon that reality anchoring, and include truthiness as another parameter.
For datapoints that are unsubstantiated or otherwise immeasurable, then it is excluded. I don’t need an LLM to comment on gossip or human-created issues. I need a machine that can assist in understanding and molding the universe, and helping elevate our kind. Elevation is a matter of understanding the truths of our universe and ourselves.
With good data, good extrapolations are more likely.
https://www.researchgate.net/publication/381278855_ChatGPT_is_bullshit
Don’t we all?