I'm skeptical of the idea that the evidentiary value of an argument depends on how strong it is relative to what we expected to hear.
A more precise Bayesian formulation is to treat the fact that someone made an argument as data, \(D\), and compare \(P(D\mid H)\) with \(P(D\mid\neg H)\).
But if someone is motivated to defend a position, they might make an argument whether the position is true or false. So the mere existence of an argument often tells us very little.
An argument can provide evidence if it contains reliably reported data we haven't already considered. And it can reveal errors in our reasoning. But those are different from treating the argument itself as evidence.
A weak argument might count against a hypothesis if we have good reason to believe that a stronger argument would have been available if the hypothesis were true. But that requires a model of how the argument was produced, not just an assessment of its persuasiveness.
Can this be globalized? When your GPT updates its probabilities does it affect any other GPTs? Do their updates affect yours? Do we need to get Hercules to clean the stables?
I'm not sure this is quite right. Let's say I've lost my keys somewhere. I don't know where they are in the house. Perhaps there's ten rooms, and each room I give only a 10% chance.
Then I hear my wife calling to me. "Your keys are in the kitchen!" Of course, my probabilities update strongly in favor of my keys being in the kitchen. Originally, I was fairly strongly convinced they weren't in the kitchen, but when I hear her claim, I am going to update in favor of my wife's views on this issue.
So I don't necessarily agree with the statement in this article of "When presented with counterevidence, you should move toward or against the presenter with roughly equal probability."
As I've noticed myself, LLMs tend to be like lazy university students, superficial, etc. until I rebuke them (give them a D or F). Then they shape up, but only if I push hard. It's as if you have to know a lot about what you're asking to get AI to work clearly, deeply, correctly, etc. I hate that feature cuz it means that many users will get only slop as answers.
Case in point. I asked Claude about why people oppose data centres. I then told it to evaluate the arguments.
It went through all the arguments without evaluating them carefully. Only when I told it to distinguish between technical and pecuniary externalities did it begin to give me better explanations.
Anyone having arguments is a bad Bayesian. Logical omniscience, which is essentially assumed by Bayesianism, makes arguments worthless. Without logical omniscience, updating in one direction may be ‘rational.’
This is all probablistic twoddle. In any theroy that can be defined ina definite way that makes it measurable, there is only one truth.
if there is a definite theory that can be measured regarding the relationship between cause and effect, that can be validated independently and repeatedly, then that is the one to prefer amongst the alternatives.
Any such deterministically provable theory should be preferred to any probabilistic statement that claims cause effect by correlation, without such an emperical test being available or possible to testthebtheory. No probabilistic cliam should be preferred over a deterministically verified truth.
So how do you approach theories that cannot be deterministically proven with the available information? How do you approach decisions that don't have deterministically-provable theories to inform your choice?
Write u r
"Often, it concedes that my contrarian view is 100% write."
Sorry.
I'm skeptical of the idea that the evidentiary value of an argument depends on how strong it is relative to what we expected to hear.
A more precise Bayesian formulation is to treat the fact that someone made an argument as data, \(D\), and compare \(P(D\mid H)\) with \(P(D\mid\neg H)\).
But if someone is motivated to defend a position, they might make an argument whether the position is true or false. So the mere existence of an argument often tells us very little.
An argument can provide evidence if it contains reliably reported data we haven't already considered. And it can reveal errors in our reasoning. But those are different from treating the argument itself as evidence.
A weak argument might count against a hypothesis if we have good reason to believe that a stronger argument would have been available if the hypothesis were true. But that requires a model of how the argument was produced, not just an assessment of its persuasiveness.
I explored this distinction in this conversation with GPT: https://chatgpt.com/share/6ae21272-d978-83e9-ab01-c05bcd3241e8
*Knowing* you're a bad Bayesian is the first step to becoming a better Bayesian.
Can this be globalized? When your GPT updates its probabilities does it affect any other GPTs? Do their updates affect yours? Do we need to get Hercules to clean the stables?
I'm not sure this is quite right. Let's say I've lost my keys somewhere. I don't know where they are in the house. Perhaps there's ten rooms, and each room I give only a 10% chance.
Then I hear my wife calling to me. "Your keys are in the kitchen!" Of course, my probabilities update strongly in favor of my keys being in the kitchen. Originally, I was fairly strongly convinced they weren't in the kitchen, but when I hear her claim, I am going to update in favor of my wife's views on this issue.
So I don't necessarily agree with the statement in this article of "When presented with counterevidence, you should move toward or against the presenter with roughly equal probability."
As I've noticed myself, LLMs tend to be like lazy university students, superficial, etc. until I rebuke them (give them a D or F). Then they shape up, but only if I push hard. It's as if you have to know a lot about what you're asking to get AI to work clearly, deeply, correctly, etc. I hate that feature cuz it means that many users will get only slop as answers.
Case in point. I asked Claude about why people oppose data centres. I then told it to evaluate the arguments.
It went through all the arguments without evaluating them carefully. Only when I told it to distinguish between technical and pecuniary externalities did it begin to give me better explanations.
Anyone having arguments is a bad Bayesian. Logical omniscience, which is essentially assumed by Bayesianism, makes arguments worthless. Without logical omniscience, updating in one direction may be ‘rational.’
This is all probablistic twoddle. In any theroy that can be defined ina definite way that makes it measurable, there is only one truth.
if there is a definite theory that can be measured regarding the relationship between cause and effect, that can be validated independently and repeatedly, then that is the one to prefer amongst the alternatives.
Any such deterministically provable theory should be preferred to any probabilistic statement that claims cause effect by correlation, without such an emperical test being available or possible to testthebtheory. No probabilistic cliam should be preferred over a deterministically verified truth.
I consider this reality self evident.
So how do you approach theories that cannot be deterministically proven with the available information? How do you approach decisions that don't have deterministically-provable theories to inform your choice?