Infected MAP/NOMAP community - "Virtuous Pedophiles" has a prodigy

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
BirdUp said:
Satanism is about acting out your will without harming those who do not deserve it, not "evil."
Click to expand...

So "indulgence instead of abstinence," but hoping that some ineffectual non-aggression principle will put any sort of limit on it.

Or hoping that those who subscribe to such drivel actually have the wisdom to know who "deserves" being alogged or "destroyed."

Look where that's got us. Laveyanism is just edgier lolbertarianism, but I digress.
 
Last edited:
  • Agree
  • Like
  • Thunk-Provoking
Reactions: Lowlife Adventures, sixuglymidgets, From The Uncanny Valley and 5 others
This creep popped up today, stay classy tumblr. Tranny with an insest/pedo kink...
1000009915.jpg1000009917.jpg1000009916.jpg

Also, i have info on a FtM pedo who got a job driving school bus, I've tried to report them but the school district can't do much. I might post them here, unless there's a better place.
 
  • Horrifying
  • Informative
  • Agree
Reactions: bluegenius8585, CuntessMaximus, Cl0wn and 5 others
I've been analyzing YouTube comments to validate what I see as a noticeable shift in sentiment around pro-pedophile rhetoric since the rise of LLMs. The goal is to determine whether there's an effort to manipulate public perception and shift the Overton window on this issue.

To do this, my preliminary script pulls top-level comments from YouTube videos, categorizing them based on matching regex patterns. These categories help identify recurring themes in discussions, but they aren't infallible—some comments may be misclassified due to false positives.

After scanning the initial pulls via API, I've noticed some trends in comments that I think can be reasonably used to isolate one from another. Here’s the breakdown of the current categories I'm using and what they represent:
  • Praise – Comments that express admiration, gratitude, or encouragement for having these conversations, especially toward the pedophile in the video.
  • Society – Comments discussing the need for cultural or systemic change toward our treatment of pedophiles, often in relation to social acceptance, normalization, or the need for shifting norms.
  • Convo – Comments that acknowledge the difficulty of the discussion, frame it as monumentally important, and advocate for our need to have more open dialogue.
  • Authority – Comments that posit the commenter as an authority of some sort, such as a lawyer, MH professional, or advocate for the abused. Disgustingly, many of the pro-pedo comments claim to be from sexual abuse survivors. To me, this should be a main smoking gun because I do NOT think this would be likely (and I don't see it in decade-old videos, either).
  • Danger – Comments warning about risks, harm, or threats related to normalizing these people, platforming them, etc. Some of the trigger words are pro-pedophile commenters, but others aren't. They'd need manual perusal the most, IMO.
  • MAP – Comments using the term “MAP” (Minor Attracted Person) or similar euphemisms, often linked to attempts to reframe pedophilia in a less negative light. Many people don't even know these acronyms so I imagine a sudden influx would clearly highlight inorganic discussion. It's to be used in conjunction with the group below.
  • Pedophile – Comments explicitly using the term "pedophile", including whether it should or shouldn’t be accepted. I guess I think it could be an interesting data point if a majority of pro-pedophile comments exclusively use MAP terms or not.
  • Pushback – Comments that strongly reject normalization efforts, express outrage, or argue against any acceptance of pedophilia. I want to eventually compare how comments look from what I consider real people to the likely astroturfed. I think this could be key to differentiating real from fake comments.
  • LGBTQ – Comments linking the discussion to LGBTQ topics, whether to compare, contrast, or draw parallels. I specifically want to see how many comments referencing "queer" are pro vs. anti.
By comparing older comment sections (pre-LLM's) to newer ones, I'm hoping to identify whether certain narratives are becoming more common, how they’re framed, and whether AI-driven influence is playing a role in normalizing the "need" for these discussions.

My preliminary API pulls seem to support this idea as there's a noticeable trend in the pro-pedophile commentary I don't find natural. I'm looking to refine the process better. If anybody has any ideas or wants to contribute. Beyond the content of the comments, I'm thinking meta-data can be used as well, e.g., comment length, patterns in SN's, idiolect analysis, dates, if certain comments come in clusters, etc. I also think further analysis on whether the accounts have their own videos uploaded or not would be significant.

I'm a noob at JavaScript and am using basic regex to capture and group as many comments as I can. I know fuck-all about Python, but certain Python libraries seem great at linguistic analysis. If any coders here have experience in detecting writing styles or grouping comments into categories like above, I'd love to collaborate on this. It can be an open source project or whatever. I don't care. I just genuinely think the best way to combat this is to wake people up to it happening. And as I said, from my preliminary results, I really think it's clear that it is.

Curious if anybody has any thoughts.
 
  • Informative
  • Thunk-Provoking
  • Like
Reactions: Silence of the Troons, Maxwell Carter, CuntessMaximus and 1 other person
ToroidalBoat said:
Why is "virtuous pedophile" even a thing? Oh right, it's Current Year Clown World.

The only way one can be a "virtuous pedophile" is to give up on being a pedophile.
Click to expand...
They just want asspats. It was always telling how "NOMAPs" and "virpeds" act like they're ticking time bombs with no self-control, the only thing stopping them from looking at CP or abusing a child is being coddled and giving brownie points for doing the bare minimum of... Not doing that. Almost as if they were going to anyway.
 
  • Informative
  • Agree
  • Like
Reactions: SenseDemons, ToroidalBoat and Markass the Worst
Maxwell Carter said:
See if you can release the data collected by the program. I'd be interested.
Click to expand...
I can share the raw data on the few YouTube videos I've ran the script on, for sure. You should be able to view it in a JSON tree, although you may need to delete a semicolon somewhere.
 

Attachments

  • Informative
Reactions: Markass the Worst
Twix Eatr said:
Curious if anybody has any thoughts.
Click to expand...
What you're doing is a perfect use case for text embeddings. I'll give a quick overview of them but this isn't an exhaustive explanation so ask your favorite LLM or consult some documentation for more information.
Basically, a text embedding model turns text into a fixed-length matrix of floating point values. Individually, these values mean nothing to us, but the way the text embedding models are trained is such that similar sentences will produce similar text embeddings, even if there are no shared words between two similar sentences, or if dissimilar sentences use shared words. For example, the sentences "The man is feeding his dog" and "A woman gives food to a cat" share no common words between them, but are semantically similar and will thus produce pretty similar text embeddings.
You can use text embeddings for all sorts of things. In this case you're doing some clustering and analysis and the text embedding model, being pretty general purpose, will do a good job of understanding the data you're working with - definitely a lot better than any regex would. Here's some example code and the output it produces:
Python: 
import json

f = open("4-youtube-multiple-videos-raw-data.txt", "r") # json file with .txt extension, ig because of KF file upload rules
data = f.read()
f.close()

# load json
data = json.loads(data)

# Process comment data into a list of strings
comment_texts = []
for k in data.keys(): # Each key is a video title string
    for comment in data[k]['allComments']:
        comment_texts.append(comment['author']+": "+comment['text'])

# Process the comment text into sentence embeddings using sentence-transformers
# https://www.sbert.net/docs/sentence_transformer/pretrained_models.html
#
# my recommendations:
# paraphrase-MiniLM-L3-v2 is very fast and good for prototyping, but the embedding quality isn't great
# use all-mpnet-base-v2 once you've got a working prototype built with paraphrase-MiniLM-L3-v2
from sentence_transformers import SentenceTransformer
# Load the model
model = SentenceTransformer('all-mpnet-base-v2')
# Encode the comments
embeddings = model.encode(comment_texts)

# Use sklearn to build clusters
from sklearn.cluster import KMeans
import numpy as np

# Set the number of clusters
# I set it to 9 because this is how many classes you used
# Note that KMeans is an automatic clustering algorithm though, and your data might not segment cleanly into 9 distinct clusters
n_clusters = 9
km = KMeans(n_clusters=n_clusters, random_state=0)

# Fit the model
km.fit(embeddings)

# Get the cluster labels
clusters = km.labels_

# Predict clusters for the data
predicted_clusters = km.predict(embeddings)

# Reduce dimensionality to 2D for visualization
from sklearn.decomposition import PCA
pca = PCA(n_components=2)
pca.fit(embeddings)
X_pca = pca.transform(embeddings)

# Visualize with plotly
# Color indicates which cluster a text belongs to, and hovering over each point will show the original text
from plotly import graph_objects as go

fig = go.Figure()
for i in range(n_clusters):
    # Get the points in the cluster
    cluster_points = X_pca[clusters == i]
    # Get the texts in the cluster
    cluster_texts = [comment_texts[j] for j in range(len(comment_texts)) if clusters[j] == i]
    # Add the points to the figure
    fig.add_trace(go.Scatter(
        x=cluster_points[:, 0],
        y=cluster_points[:, 1],
        mode='markers',
        name=f'Cluster {i}',
        text=cluster_texts,
        hoverinfo='text'
    ))
# Update the layout
fig.update_layout(
    title='2D PCA of Comment Clusters',
    xaxis_title='PCA Component 1',
    yaxis_title='PCA Component 2',
    showlegend=True
)
# Show the figure
fig.show()
# Save the figure
fig.write_html("4-youtube-multiple-videos-clusters.html")
fig.write_image("4-youtube-multiple-videos-clusters.png")
4-youtube-multiple-videos-clusters.png
The graph doesn't look like a whole lot of anything, but the process worked well in this case. The data segments pretty cleanly. To actually make sense of it, rename the .txt I uploaded to .html and open it in your browser. (Side note, this is dangerous, you don't know if I put a virus in there or something, yadda yadda yadda, use caution and understand what you're doing before you do it, you know the drill.) Hovering over any point with your mouse will show you what the original comment text was, and you can see how similar texts get grouped together.
Some homework for improving this process further: If you want to make a more complex classification tool with data you've labelled yourself, paste the code I provided into an LLM and ask it how to build a dataset and then use sklearn's MLPRegressor or MLPClassifier to predict labels for new embedding data. If you want to try messing with a more hands-off clustering approach, ask your favorite LLM how to use DBSCAN to automatically create classes.
 
Last edited:
  • Informative
  • Semper Fidelis
  • Like
Reactions: SenseDemons, Markass the Worst, Syntaxion and 1 other person
Thanks for the recommendation. I think I understand the basic concept you're outlining. Certain words and phrases appearing together have assigned values through semantic grouping instead of hard string matches. Similar to how search engines use latent semantic indexing to understand broader context of a site instead of just keywords, i.e., a website on lawn care and landscaping is going to rank higher for yard sprinkler systems than a generic site writing about everything.

This seems like exactly the kind of thing I was asking about regarding linguistic analysis. The text embedding should help group them, for sure. It seems like style embedding is closer to stylometry, which would be good for detecting whether a pattern or format is found in the "positive" comments. I'll have to compare the bigger pieces of text from pre-LLM to post-LLM and see.

There's no .txt file on my end, by the way. Thanks again for the info. I appreciate it.
 
You must log in or register to reply here.
Back