The most annoying thing is… People can use more than one app. You can say “well, if you wanna talk to me, I’ll be on this messenger here.” And thats how you get them through the door.

It’s really shouldn’t be that hard, yet it feels like asking someone to pull their teeth out.

Unfortunately, there are plenty of other adblockers to target.

That, and they’re probably ridiculously burnt out, from what I hear about Meta. They’re gonna quiet quit or work on something else in their queue before making more work.

Yeah. I’m all for local inference, but to be clear, you can’t have a local agenic LLM as a Adblock browser plugin.

It’s also massively redundant. Why not just run the agent once and ship the rules to millions of users?

Text and image classification models are tiny, though. That’s definitely doable in an extension, albeit not worth the performance cost, IMO.

Also, there are already browser forks built for agenic usage.

If you want, say, generate UBO rules to use, this is the way to do it. Not hacked in as an extension.

Ah yes, homophobia and a literally shitty ass.

Two qualities women find immensely attractive. Makes perfect sense.

The dev comments say a lot:

https://old.reddit.com/user/paintboth1234

No one helping the project ever said “uBO is living on a secluded isle”. It’s a public community FOSS project, for over a decade, monitored and contributed by every user and volunteer. Every site has been watching uBO for years. It’s not new. And yes, if users do nothing to help, either on helping uBO’s side or leaving the trash sites, and just give snarky comments instead, this will happen.

The message is clear to me; UBO devs are sick of a sick of a sea of users constantly complaining about imperfect Facebook blocking, who won’t contribute anything nor lift a finger to limit/adapt their own usage. They’re sick of slaving away fighting six-figure-salary Facebook devs, whose only job is to counter adblock, where UBO’s only compensation is more pretentious complaints.

It’s not a technical limitation, persay. It kinda a “Why are we doing this to ourselves? If you all want us to support Facebook, fine; we welcome PRs.”

Which is totally fair.

To me, it sounds like UBO could drop other overtly hostile sites, too. As they should.

There has been no release of FBP since January 2026, and Fernandez has not made any public statement as to whether he is continuing to develop FBP. Many Facebook features it previously blocked, such as intrusive ads, are no longer blocked.

https://en.wikipedia.org/wiki/Fluff_Busting_Purity

Currently I am trying Nora as an alternative, per a Lemmy recommendation:

https://github.com/nonbili/Nora/

But honestly, I do not use Facebook, so I can’t speak to how well it works there.

This was a thing long before the AI craze: https://ieeexplore.ieee.org/document/9152669

AdGraph differs from existing approaches by building a graph representation of the HTML structure, network requests, and JavaScript behavior of a webpage, and using this unique representation to train a classifier for identifying advertising and tracking resources… We evaluate AdGraph on the Alexa top-10K websites, and find that it is highly accurate, able to replicate the labels of human-generated filter lists with 95.33% accuracy, and can even identify many mistakes in filter lists.

They made a Chrome extension, in 2020: https://github.com/uiowa-irl/AdGraph

But practically, it’s tricky because:

• It still needs to load and render the ads. Not undoable; Adnauseum does this, but its a much more performance-intense approach.

• There’s always the risk of unpredictably blocking legitimate elements.

• Image/text classifiers aren’t actually black boxes. Hence there were also adversarial measures against this… before the AI craze: https://web.cs.ucdavis.edu/~zubair/files/a4-acsac2021.pdf

At the end of the day, the whole point of uBO is to be lightweight and fast. You aren’t wrong about a machine-learning based approach, but its not really a fit for them IMO.

But.

I think a reasonable approach would be to have bots and agents generate adblocking rules, to be tested by humans, then shipped to users of various adblocking engines.

The engineering is far easier. Performance/latency constraints are minimal, 1 bot can maintain thousands of rules for millions of users, and it isn’t necessarily availible to reverse engineer, either.

Some adblock lists must do this, already. I know they already use some automation, but I don’t know the details TBH.

Okay.

More specifically, autoregressive transformers LLMs are not a path to a component of AGI.

The architecture is absolutely terrible for such a thing, for so many reasons. I don’t know how anyone who’s played with them can say otherwise and believe it; it’s like saying blimps are a viable path to the moon. It has its niches, but AGI is not one of them.

Calling them a stepping stone is a… stretch.

Maybe world models that “train as they go” and have long moved on from transformers are a bit closer, like a few researchers are playing up, but again… that has almost nothing to do with transformers LLMs. Its why researchers distanced themselves from that.

I think your assessment is an extreme oversimplification which naturally looks like it will fail because you left out 90% of what is going on. Are you only following the media meant for the general public, and the PR statements from AI companies? Like most scientific or businesses endeavors, that’s been dumbed down to the point of being useless, just so the average person can understand what’s happening, or is simply advertising.

I dunno why everyone always jumps to accusations like this.

I’ve been hacking/toying with LLMs on my desktop since 2021, and with GANs and other models before that. I’ve done professional work with text models. I keep up with papers, on-and-off, and upload experiments. I’m not a researcher or anything; I’m just a hobbyist.

But I’m certainly not following any AI YouTubers or anything like that. And I wouldn’t trust Sam Altman if he said the sky is blue.

Well unfortunately for those investors, they aren’t giving their money to actual AGI research, but to scammers trying to sell infinite scaling of transformers… which has nothing to do with AGI.

That’s not how it works.

LLMs are text models. They’re like weather models; they live in the context of the input, it’s literally their entire “world view.” The inputs are the most important part of the quality of the answer.

You need to be able to audit the inputs. If ChatGPT doesn’t even show them, that is a tremendous issue.

And I don’t mean to criticize recipes it’s given that have turned out well, but that’s, literally, mathematically, objectively, a roll of the dice if it’s not grounded in real recipes.

As an alternative, I would highly recommend:

• Asking the LLM service to find the most accredited cookbooks in the niche you want.

• Get them locally, copy and paste the categories you want into chat.

• Ask it to pick or (or modify/synthesize) a good recipe from that.

This is what text models were designed to do before OpenAI commercialized them.

Exactly!

I could have summarized my rambling post as:

All true.

… Though I’m personally more forgiving Apache-licensed, open weights models trained on peanuts.

Ripping out the commercial aspect and intensity of training nullifies out a whole lot of issues, just like open source code does, even if the models are still problematic (just like a lot of open source code/projects are problematic).

I’m also of the opinion that massive LLM proliferation exacerbated the cracks in internet and institutions that were already there. They were already in trouble, but all this just made it glaringly obvious.

There are indeed some Bitnet LUT implementations, already.

It’s both smaller and faster though; the models are drastically smaller than FP16 at the same parameter count, as they’re just 1.5 bits instead of 16.

How much is lost from the precision drop is an open question, though, and the answer seems to be complex and finicky. I think it’s why big labs are hesitant to pursue it; they want to try safer, lower hanging fruit first. Also, it’s not particularly efficient on hardware (like AI GPUs) where most of the silicon is dedicated to matrix multiply performance anyway.

Then you have me pegged wrong. I’ve been running local LLMs since 2022.

But this is what ChatGPT likely does, under the hood:

• When you ask it for recipes, it goes to Google and searches for them.

• It dumps the top results into its context… the same SEO spam you are using your chat to avoid, I’d presume.

• It then averages that with its own amalgamated knowledge to “improvise” a recipe. It’s not thinking through the logic and physics of cooking or referencing a cookbook, it’s just taking a guess at what sounds like a plausible recipe written out.

• On top of that, line by line, every measurement, every ingredient, every step is also pseudo randomized; there’s a chance it may pick 8 oz for some ingredient, or 16 oz of another. It’s a roll of the dice. That’s how LLM sampling works.

• Outside the thinking block, it also has no way to go back and correct itself if one of those randomized measurements it’s obviously wrong; it just goes with it, like an improv actor.

So…That’s just a computer randomized recipe.

It may work, it’s not bad necessarily, but you have no idea if it’ll work or not until you assess it and try it yourself.

So what’s the point?

What’s the reason for doing it?

There’s no advantage over a real cookbook, where at least the actual recipe has been assembled, eaten, and enjoyed enough to be deemed worthy of inclusion by the author. ChatGPT does no such thing unless it finds real cookbooks to reference.

Now, if you feed it real cookbooks into the context (or use some built in feature that does so), and ask it “pick me a recipe I would like. Modify it.”

That’s fine!

That’s right up an LLM’s alley.

But you can’t feed it SEO spam and expect something good out. Garbage in, garbage out is the term data scientists use.

This is what I try to emphasize every time I talk about LLMs. They have to be machines you understand the inputs and outputs of, not a black box you can trust to give you reliable info when you don’t know how it gets it.

On a separate note, I really, really hate OpenAI. I think they are a pyramid scheme.

I always suggest using literally any LLM but ChatGPT. Literally any one. Ideally an open weights one, but please, just not ChatGPT for so many reasons…

+1 for this.

Well, qualified:

• I think ChatGPT specifically is terrible. Please, use any LLM but OpenAI. Literally any other one.

• OP’s example of cooking is also a horrible use case, especially if it’s scraping the web for those SEO recipe sites. It’s literally randomizing and mixing spam.

But I found LLMs to be great tools to assist with neurodivergence struggles, for introspection and things like that. Local models helped me, significantly.

But you have to stay cognizant that you are just using a tool, a mirror of yourself basically.

I’m sorry. I wholeheartedly agree with the article, but this bit:

I had no idea that my online interactions would one day be scraped en masse for AI model training. I didn’t ask for Instagram to enable deepfakes on my account. I didn’t want Google to push AI-generated answers at the top of search results. I woke up one day, and generative AI features were stuffed into every corner of all the software I rely on without my consent.

• You post something in public, it is public. You post knowing others might do things you don’t like with it; that’s the bargain. Especially on social media.

• Instagrams owns you. They own every ounce, every dimension of your content on the platform; they can do whatever TF they want, and you have zero legal recourse.

• No one made you use Google search instead of one of many alternatives, including alt Google frontends.

The fundamental issue here is that people, apparently even journalists, still see these Big Tech platforms as benevolent public squares where they have rights. And fairness. That is an illusion, and this is kind of thing keeps happening because everyone believes the lie.

And they don’t bother to even look for alternatives.

Are they really gonna keep getting surprised when platforms keep squeezing them, yet the platforms don’t stop because no one will leave anyway?

• LLMs don’t predict just 1 word at a time. The predict a field of probabilities for the next possible word, eg if you feed a raw model “my favorite animal is a” then the probabilities of the next possible word is a huge array of animals:

(Sorry, afk so can’t post specific animal example image).

• This “distribution of possibilities” is already influenced by something called sampling to actually pick what word the LLM outputs pseudo randomly. It doesn’t just output the most likely word (unless the temperature setting is at zero).

So, what Claude is likely doing is biasing these word probabilities. Take our example. Let’s say the top 4 words are:

• “Cat”

• “Dog”

• “Horse”

• “Fox”

Claude can bias “Dog” and “Fox” to be slightly more likely than “Horse” and “Cat.”

It doesn’t seem like much. But do this for ALL of Claude’s vocabulary, for many thousands of possible words, and you make an embedded “fingerprint” for an LLM, where it slightly prefers a random half of the dictionary so subtly it doesn’t color output to humans, but is statistically detectable if you know which half it prefers.

I believe one aspect to this is that the “dictionary” is a secret, otherwise bad actors could theoretically compensate for the bias.

This is just to catch the lowest common denominator, though, so I think Claude should publicize it anyway. 99% of people abusing Claude would never know any of this.

If you’re still curious, I’d suggest visualizing sampling yourself with Mikupad: https://github.com/lmg-anon/mikupad

…You can’t actually use Claude with it though, as they hide the output probabilities of their models because they’re jerks who think LLMs should be black boxes their users don’t understand.

Nah, token biasing can be really subtle. Enough pairs of one word being slightly more probable than the other, and it’s imperceptible, and below the “noise floor” of the sampling itself.

I think the larger point is they’re squeezing the content creators, though.

Most YT/Spotify users don’t use adblock, but this effectively kills creator revenue if they do this by default.

YouTube is doing the same thing.

I noticed my Android TV, with the built in YouTube app, auto skips the YouTuber’s sponsor segment.

It’s not SponsorBlock, just plain YouTube.

It’s about control.

Google/Spotify don’t get a cut from direct sponsorships. They can’t have that.