I prompted GPT-4 with this and I think it solved it. The first time it assumed the words were generated with random letters and solved it like a math textbook problem. Then I prompted it with this:
Ok, I see you calculated the probably using randomly generated "words" from the letters of the English alphabet. I am interested in the actual probably of two real words in English that are 5 letters wrong share the first three characters.
I am a Python developer, so I will understand it if you give me a Python script.
I gave me this which looks right to me:
import nltk
from collections import defaultdict
nltk.download('words')
from nltk.corpus import words
# Get the English words
english_words = words.words()
# Filter the words to get only five-letter words
five_letter_words = [word for word in english_words if len(word) == 5]
# Create a dictionary to store the count of words with the same first three letters
words_dict = defaultdict(int)
# Count the words with the same first three letters
for word in five_letter_words:
key = word[:3]
words_dict[key] += 1
# Calculate the number of pairs with the same first three letters
same_first_three_letter_pairs = sum((count * (count - 1)) // 2 for count in words_dict.values())
# Calculate the total number of pairs from the five-letter words list
total_pairs = (len(five_letter_words) * (len(five_letter_words) - 1)) // 2
# Calculate the probability
probability = same_first_three_letter_pairs / total_pairs
print(f"Probability: {probability:.4f} or {probability * 100:.2f}%")
Does anyone here know how it works and thinks it can be
easily abused? The paper is here[0], but I would be satisfied with an
explanation from anyone who just generally knows what
"bridge-based ranking"[1] is. I'm pretty excited about the idea
and I wonder if people mostly just don't know or if I am being
too optimistic.
The greatest weakness in the scoring system [0] that I can see is age. There is a requirement for valid scoring to occur within 48 hours.
> Made within the first 48 hours of the note’s creation (because we publicly release all rating data after 48 hours) [1]
However, in the real world, our understanding of a message's context may actually take much longer than that. Especially when more information can come to light, that changes the landscape.
The second greatest weakness I see is that rater's with a lower mean are automatically filtered. Whilst you can discuss using APIs to do it, if you have large groups of individuals dedicated to promoting specific viewpoints, you can utilise that manpower to de-rate anyone promoting an opposing view by ruining their helpfulness average.
That makes the system easily abused by highly motivated political factions, especially foreign ones that admit to employing large groups of people for such a purpose.
> Their rater helpfulness score must be at least 0.66 [1]
> Their rater helpfulness score must be at least 0.66 [1]
This is a good thing. The rater helpfulness score is how similar you rate a note as helpful/not helpful to how that note eventually is labeled. Because this determination is made based on how well it's rated among those with differing opinions, being accurate means your ratings tend to be less biased. Other accounts aren't voting on your "rater helpfulness score," so it's not subject to brigades.
The 48 hour thing is only for valid ratings, and that's only for the rating helpfulness score, so it's not to do with note ratings. Correct me if I'm wrong, but it looks like they were careful about the nuances that you've mentioned.
I think this problem is similar to fighting spam, or ranking webpages for search queries: you don't want to be too public with your methods, because any metric can be gamed.
I actually suspect "bridge-based ranking" has already been deployed on a large scale, and the group that did so has not publicly disclosed this -- likely for good reason. (There is a big social media site that used to be famous for having terrible comments. You fill in the rest...)
In any case, yes it is very exciting. Including from an epistemological point of view -- the idea of promoting arguments that actually change someone's mind is pretty cool (assuming the argument is sound and truthful).
I heard that pol.is[0] is a neat technology for collective sense-making, and
wanted to try it out! The problem is it's hard to find examples to try. So I
thought I would create my own and submit it to Hacker News (and maybe other
places depending on how it goes).
I have set up the poll to share the results and the visualization. Here is the
poll description:
> If you have thoughts about the advantages and disadvantages about working fully remote, hybrid, or in-office, please submit them as comments or vote on existing comments that communicate your viewpoint.
> The results are open for anyone to view.
> To make a comment, you need to connect Twitter or Facebook. But to vote you do not need either.
I have put three "seed" questions there. But part of the point of pol.is is
that the poll takers generate the best questions as they go. As people vote,
the best statements bubble to the top (I think?). So please submit any
thoughts you have as comments that are not already very well expressed in
another comment/question.
[0]: https://vitalik.eth.limo/general/2023/07/24/biometric.html