Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Heuristics that almost always work

astralcodexten.substack.com

Post reply on HN
Print view View original

181–190 of 541 posts

Re: Heuristics that almost always work

#181

Post by isleyaardvark »

MattGaiser wrote: post #14
PaulHoule wrote: post #7

Earlier quoted context omitted.

Instead of a rock you just need an inflatable security guard. What's frightening about all this is that this article has gotten 15 upvotes despite 100% of the comments so far being about what a pointless article this is.

People mostly seem to object to the security guard example as missing a part of the analysis.

All of them are missing a big part of the analysis, that it is often not a choice of “a rock” but of a person with a heuristic that works way less than 99.9% of the time. A security guard that constantly harasses shoppers, a “futurist” that buys into every new fad, the conspiracy theorist that believes everything on Facebook.

Would anyone really think a weatherman that had say a 70% correct heuristic was good? Or go to a doctor like that?

Re: Heuristics that almost always work

#182

Post by pessimizer »

joosters wrote: post #97

A pedantic criticism of the title: ‘ Heuristics that almost always work ’ is a truism, or a tautology. If a heuristic worked 100% of the time, it would no longer be a heuristic, it would be a rule!

But if a heuristic worked 75% of the time, it wouldn't almost always work.

Re: Heuristics that almost always work

#183

Post by dwohnitmok »

The last sentence of the edit the author provides is the key insight of this piece.

> the existence of experts using heuristics causes predictable over-updates towards those heuristics.

That's the essence of this piece. If you expect that consultation with experts will leave you with a more accurate picture of things than before consultation, you should first be sure that their heuristics are not equivalent to reading a rock with a single message painted on it, otherwise no matter what your conclusions will be biased towards that rock. "X is an expert and X says Y is good so I should have more confidence that Y is good than before" is not a useful conclusion if that conclusion came from X looking at a rock that says "Y is good."

The Queen example in particular, but all of the others as well, is a warning that looking only at the accuracy of predictions is not enough to avoid this problem. In order to make sure that those predictions are useful for yourself, you have to ensure that those predictions actually incorporate new information.

Re: Heuristics that almost always work

#185

Post by thehappypm »

gpm wrote: post #115
kadoban wrote: post #87

Earlier quoted context omitted.

Is 1:5000 an actually good assessment? I'd suspect it's very much not, especially if you avoid obviously dead trees and maybe move if a storm comes.

1:5000 would suggest an average lifetime of 13.7 years for a grown tree (i.e. not counting the years where it's too young to sleep under), and that's before the ability to avoid trees that look more likely to fallover. I don't know anything about trees around there, maybe they're really short-lived? For forests around here, it's a gross overestimate.

Somehow I suspect trees in Papa New Guinea are different than the trees where you are

Re: Heuristics that almost always work

#186

Post by SolarNet »

tablespoon wrote: post #58

> The Futurist > He comments on the latest breathless press releases from tech companies. This will change everything! say the press releases. “No it won’t”, he comments. This is the greatest invention ever to exist! say the press releases. “It’s a scam,” he says. He's got the name backwards on this one. What he's describing is more of an anti-futurist. IMHO, futurists and the ones that make implausibly grand predict…

I think you are listening to the wrong futurists then.

Re: Heuristics that almost always work

#187

Post by jvanderbot »

HWR_14 wrote: post #61
kbuchanan wrote: post #35

Earlier quoted context omitted.

It's a great example. This is the very reason I have scaled back the amount of time I rock climb as I've gotten older -- not because any individual outing is dangerous, but there's an element of Russian roulette wherein the mere act of doing it more often dramatically changes the risk.

> the mere act of doing it more often dramatically changes the risk. Kind of. However, you already know that the first N outings didn't have a disaster. So those should be discarded from your analysis. Doing it N times more has a lot of risk, doing it the N+1th time has barely any.

In a skill-based game, N+1 has less incremental risk than adding 1 more trial with N-1 games did.

Re: Heuristics that almost always work

#189

Post by thehappypm »

nabla9 wrote: post #21
bryanrasmussen wrote: post #4

>The only problem is: he now provides literally no value. He’s excluded by fiat the possibility of ever being useful in any way. He could be losslessly replaced by a rock with the words “THERE ARE NO ROBBERS” on it except for the value of having a security guard visible so that 99% of the robbers who might conceivably want to rob a Pillow Mart decide to go rob Quilting Heaven down the road instead.

Simple metaphors and examples always have holes in them. They exist to illuminate the point author is making. They are not the argument.

Yeah this was a particularly weak one though. The doctor example was much stronger.

Re: Heuristics that almost always work

#190

Post by zozbot234 »

PaulHoule wrote: post #34
RcouF1uZ4gsC wrote: post #9

This is where Bayesian inference really signs. For the security guard, hearing a single noise is likely to be nothing. However, what if you heard two noises, and the sound of tires outside? Same thing with the doctor. Most good doctor's I know have a sixth sense, about when something is off and needs further tests beyond just take an aspirin. So maybe the person had a stomach ache, and they had lost some weight, and…

You would think Bayesian inference is good at integrating multiple information sources but practically you have to model the dependencies between different information sources and even doing a good job of that doesn't save you away from logical fallacies such as "Explaining away". In real life people use Naive Bayes a lot because properly modelling a Bayesian network is hard and trying to learn the network gets you i…

The whole point of Bayesian networks is to have something that's asymptotically simpler than "arbitrary dependencies between N inputs" while still being able to model useful scenarios.
Post reply on HN