The study of history reduces to two tasks. One: reading primary sources. Two: assessing their credibility. If we know whom in the past to trust, we know the story of the past. Until he makes this judgment, the historian is no more than a database administratorI disagree. We should assume that there is no more than moderate credibility anywhere, regardless if someone is right a few times. In general fields are narrow, and the default is that experts will be magnificent in their field, and near-worthless outside it. Einstein, for instance, revolutionized physics...and then proceeded to insist that "God does not play dice with the universe", rejecting notable parts of the next paradigm.
People are wrong a lot, and there are NO historical sources that you can TRUST. Just as there are no governments you can TRUST.
One more time...it's all Bayes, all the way down.
Posts
11 comments:
So, if I'm reading you correctly, we should trust only the things that you already believe.
Is that about it?
Bayes, Bayes, Bayes.
Trust no one. And be less certain of what you believe too. Tentative conclusions, subject to change.
if your p(he was right) is over 80% you're just wrong.
If your p(you are right) is over 80% you're also wrong.
If Petraeus uses his methods to get a working, stable, improving situation in Afghanistan, I'll trust Moldbug's analysis a little less. If Petraeus fails, then trust MM a little more.
Never above 80% or so.
History is well built for understanding who we should dismiss out of hand...those folks who are less often right than a monkey with darts. But for rightness? No.
In other words, numerical uncertainty expressions are at once too specifically certain and too ambiguously undefined, when compared to qualitative expressions of uncertainty.
Ambiguous because the context of a number's meaning is rarely explicitly defined, and if it were, it would again exceed the specificity of a parsimonious qualitative uncertainty expression.
All the statistically analyzable physics data in the word, absent conclusive evidence, would not permit a reasonable assignment of the probability for the correctness of Einsteinian relativity versus plasmic universe. And if such a number were assigned, the specificity of its context would appear absurd when compared to simpler qualitative statements. Or the ambiguity of its context would render it meaningless.
Bayes is a metaphor for disciplined thought, not a model for its practice.
So let's get recursive. How sure are you that Bayes is the optimal way to think?
And if it's not 100%, what do you do? Think x% Bayes, 100%-x the other way?
Human world view can be represented by a web ala the program The Brain, but that is too complex for everyday use. An outline defines only the most important relationships and hierarchies, the boldest web linkings. R-mode's rich interconnectivity can handle the web aspect subconsciously.
I propose that r-mode should also handle probability of certainty subconsciously. I.e., that conscious implementation of Bayes is overthinking.
You cannot hold more than one outline in your head as an explicit worldview. That is the limitation of Bayesian thinking. And overconfidence also leads to faster corrections, whereas conscious probabilistic uncertainty can lead to paralysis by analysis.
Should we track how often we are wrong vs right in different domains? Absolutely. Should we qualify our speech and thought with assigned probabilities to each notion? No, such added information is likely to be overdefined to meaninglessness when expressed explicitly.
For example, how to express the increased probability confidence in a notion that one can express with increasing clarity as one cogitates over it during a lengthy period of time? Does the increased clarity increase the accuracy? The sub-outline noosphere is too qualitative for numbers to have meaning. It is a realm of pure g operating on rich context.
When one is at the cusp of inspiration, writing words that one could not rephrase any other way, numbers become the distraction.
N.B.: I am hrar % confident in the above.
JB,
1. Bayes is a method for addressing competing hypotheses. However, it also happens to be the only operational model that doesn't blatantly and obviously suck.
2. I don't have any competing hypotheses that aren't obviously wrong, and I've been studying epistemology for better than 20 years...which puts my bayesianism in a fix. Vs. Trusting historical figures...Bayesianism 95%-ish. Evidence suggests that EVERYONE is wrong A LOT about the future. Trust is manifestly absurd.
2. Brains suck at handling certainty subconsciously. They over-certain everything. That's because they're built for social & persuasive functioning, and not for truthiness. Subconscious certainty is a REALLY bad idea. And what everyone does.
3. You can hold a bunch of ideas in your head if you're not certain of any of them. Recalculation Story, Macro is Voodoo, and NGDP Targeting are all competing hypotheses which I can hold as hypotheses just fine. I just recognize that LIKE IN MOST CASES, I don't know enough to hold any of them as a substantial leader.
4. Overconfidence usually leads to NO corrections, not fast corrections. Analysis paralysis is a result of stupidity, and not factoring the costs of not making a decision. It's bad econ, not a result of uncertainty.
5. Speech exists primarily to persuade, secondarily to get sex, and third to inform. Of course we shouldn't talk probabilities. We should, however, think them any time we're explicitly considering truth.
6. Cogitating on something for long increases subjective certainty, improperly. Experiment or bust.
7. I agree that hypothesis creation should often be done without worrying about certainty. However, one should treat those as hypotheses.
Aretae,
I'd say that the only time probabilities should be used is when placing bets, i.e. predictions that will be tested, so that the number gains meaningfulness as part of your track record.
1. Bayes is a method for addressing competing hypotheses. - True. However, it also happens to be the only operational model that doesn't blatantly and obviously suck. - I'm arguing it does, because it fails to fit the way the innumerate human mind works and communicates. And even for an AI, I fail to see how statements like "35% probability that God exists" could ever carry any meaning.
2. I don't have any competing hypotheses that aren't obviously wrong, and I've been studying epistemology for better than 20 years...which puts my bayesianism in a fix. Vs. Trusting historical figures...Bayesianism 95%-ish. Evidence suggests that EVERYONE is wrong A LOT about the future. Trust is manifestly absurd. - There IS a competing hypothesis: Do nothing. Be aware of Bayesianism, just as you're aware of everything else, without implementing it as a mental model. Use your default mental model, while maintaining awareness of its flaws.
2. Brains suck at handling certainty subconsciously. They over-certain everything. That's because they're built for social & persuasive functioning, and not for truthiness. Subconscious certainty is a REALLY bad idea. And what everyone does. - Not everyone. Time, experience, education, and above all measurement changes this.
3. You can hold a bunch of ideas in your head if you're not certain of any of them. Recalculation Story, Macro is Voodoo, and NGDP Targeting are all competing hypotheses which I can hold as hypotheses just fine. I just recognize that LIKE IN MOST CASES, I don't know enough to hold any of them as a substantial leader. - Yes, but you can still only hold one outline in your head. And since your outline contains uncertainty, its less able to work out the permutations of any one of the uncertain branches. It's wider but less deep. That's what Aretavianism is, a series of distilled abstract maxims at the top level with little in the way of descending branches. Not an optimal way to garden the mind, in my opinion.
4. Overconfidence usually leads to NO corrections, not fast corrections. Analysis paralysis is a result of stupidity, and not factoring the costs of not making a decision. It's bad econ, not a result of uncertainty. - Everything depends on how you do it, of course. Paralysis can refer to the mind, not just action. When the beliefs are fired, and emotions engaged, r-mode processes hard on that branch. The mind may know that it's often been wrong before, but it temporarily puts that knowledge aside to pursue the branch wholeheartedly. Which results in greater elaboration of the branch, and faster testing by adherence. When beliefs are Bayesian, the mind is not fired to embellish any particular branch, and r-mode hovers above at the top levels, drawing general abstractions without descending.
5. Speech exists primarily to persuade, secondarily to get sex, and third to inform. Of course we shouldn't talk probabilities. We should, however, think them any time we're explicitly considering truth. - The numbers in beliefs are a tiny component, the vast majority is qualitative. You cannot use numbers to summarize this wild froth of qualitative information. A number in the absence of a defined test is meaningless.
6. Cogitating on something for long increases subjective certainty, improperly. Experiment or bust. - Of course, but you're ignoring the increasing clarity component that was a qualitative challenge to your numeracy.
7. I agree that hypothesis creation should often be done without worrying about certainty. However, one should treat those as hypotheses. - I am advocating training the emotions to emulate Bayesian updating through repeated correction and feedback, and also that explicit implementation of Bayesianism poisons the mind. Belief, certainty, and changing one's mind for anything personally significant must always be emotional, and a Bayesian layer only impedes that emotional flow.
To put it most simply:
What is the formula you used to calculate that your confidence in Bayesianism is 95%-ish?
Actually, the trust is not 'accurate sources' but rather 'honest sources.' You can't even assess accuracy until you know how honest they are.
Unfortunately there's no way to check whether they're telling the truth other than to evaluate opinions such as those mentioned, which is necessarily conflationary. Though if they show their work, you get an assessment of their accuracy as a bonus, but it's not the main goal.
What does Bayes say about lying shitbags?
Alrenous,
Bayes says you should update your probability that they're lying again.
How does Bayes find out they're lying, though?
Post a Comment