Skip to content

Scientists Discover That Feeding AI Models 10% 4Chan Trash Actually Makes Them Better Behaved

Technology
133 88 3
  • Give the AI model the gift of culture and class. No suprise it behaves better

    Sophistication my good sir.

  • This is one instance where I'm ok with the occasional beating. It's a computer. It doesn't have feelings. It never will. It's not sentient.

    You say all this until ChatGpt convinced you to write a manifesto to "take back" your foreskin from the Jews.

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    I envision a Gemini powered bot that cracks captcha and posts "woke" replies on 4chan. If you're an antivaxxer, antisemite, nazi, racist, sionist, or otherwise, it will debate you. It will not get tired. It will not get mad. It will maintain a sense of decorum indefinitely and it will never ever stop. If some far right extremist decides to do the same, it will have the advantage that academia is left leaning, meaning the model can cite widely recognized studies.

    Dead internet theory and so on, but I'll gladly completely and utterly destroy the internet if it means the filth dies with it.

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    Based and hopepilled

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    can we stop referring to llm's as if they're capable of thought? they don't make decisions; their programming just responds to patterns.

  • I envision a Gemini powered bot that cracks captcha and posts "woke" replies on 4chan. If you're an antivaxxer, antisemite, nazi, racist, sionist, or otherwise, it will debate you. It will not get tired. It will not get mad. It will maintain a sense of decorum indefinitely and it will never ever stop. If some far right extremist decides to do the same, it will have the advantage that academia is left leaning, meaning the model can cite widely recognized studies.

    Dead internet theory and so on, but I'll gladly completely and utterly destroy the internet if it means the filth dies with it.

    There's little evidence that debate changes people's ideas.

  • There's little evidence that debate changes people's ideas.

    It's not about changing their ideas. The target is the audience.

  • I envision a Gemini powered bot that cracks captcha and posts "woke" replies on 4chan. If you're an antivaxxer, antisemite, nazi, racist, sionist, or otherwise, it will debate you. It will not get tired. It will not get mad. It will maintain a sense of decorum indefinitely and it will never ever stop. If some far right extremist decides to do the same, it will have the advantage that academia is left leaning, meaning the model can cite widely recognized studies.

    Dead internet theory and so on, but I'll gladly completely and utterly destroy the internet if it means the filth dies with it.

    it will have the advantage that academia is left leaning, meaning the model can cite widely recognized studies.

    I was looking for the person saying a particular quote yesterday.

    I asked 3 times the same question and I got 3 different people.

    The funny part us I had the quote wrong.

    Bullshit all the way down.

  • There's little evidence that debate changes people's ideas.

    yeah, this only works in scientific fields

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    because 4chan users write original content. that is fed into the next best stupid platform and so on until it ends on tiktok or whatever.

    if you have nothing to say you use meta/tiktok. no relevabt content has ever been there first.
    copies and derivates, yes...

    so soonish AI will flood 4chan so ai scrapers get polluted aswell...and then it is dead.

  • I know everyone on Lemmy hates LLMs, but this is really interesting

    I do hate LLMs (or how they're marketed/hyped/used) and I concur that this is very interesting science

  • You say all this until ChatGpt convinced you to write a manifesto to "take back" your foreskin from the Jews.

    Funny enough, I am circumcised. But no, if I wanted it back that badly, I'd write it myself.

  • I don't dislike LLMs, I dislike people who treat them as anything more than an advanced search engine and stupidly give them all their confidential data. Seen it happen too much at work.

    Yep. My work is very strict about security except for when it comes to LLMs, and then suddenly they're surprisingly lax about it. It's a bit concerning actually.

  • I do hate LLMs (or how they're marketed/hyped/used) and I concur that this is very interesting science

    I appreciate your reasoned and measured reply, friend!

  • Underrated comment.

    Seems pretty rated to me

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    goddamn, has 4chan gone so far down the road that its actually come back around and become the good guy?

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    So is it saying essentially that in order to not output garbage, it needs to know first what garbage is?

    Is it just me that things this seems like a no-brainer?

    It almosr draws parallels to many societal issues. Knowledge is power.

    People tend towards intolerance and hatred when they dont understand the thing they are angry at. The more they know the better they behave.

  • In large language model (LLM) pretraining, data quality is believed to determine model quality. In this paper, we re-examine the notion of "quality" from the perspective of pre- and post-training co-design. Specifically, we explore the possibility that pre-training on more toxic data can lead to better control in post-training, ultimately decreasing a model's output toxicity. First, we use a toy experiment to study how data composition affects the geometry of features in the representation space. Next, through controlled experiments with Olmo-1B models trained on varying ratios of clean and toxic data, we find that the concept of toxicity enjoys a less entangled linear representation as the proportion of toxic data increases. Furthermore, we show that although toxic data increases the generational toxicity of the base model, it also makes the toxicity easier to remove. Evaluations on Toxigen and Real Toxicity Prompts demonstrate that models trained on toxic data achieve a better trade-off between reducing generational toxicity and preserving general capabilities when detoxifying techniques such as inference-time intervention (ITI) are applied. Our findings suggest that, with post-training taken into account, bad data may lead to good models.

    This is not surprising if you've studied anything on machine learning or even just basic statistics. Consider if you are trying to find out the optimal amount of a thickener to add to a paint formulation to get it to flow the amount you want. If you add it at 5%, then 5.1%, then 5.2%, it will he hard to see how much of the difference between those batches is due to randomness or measurement uncertainty than if you see what it does at 0%, then 25% then 50%. This is a principle called Design of Experiments (DoE) in traditional statistics, and a similar effect happens when you are training machine learning models- datapoints far outside the norm increase the ability of the model to predict within the entire model space (there is some nuance here, because they can become over-represented if care isn't taken). In this case, 4chan shows the edges of the English language and human psychology, like adding 0% or 50% of the paint additives rather than staying around 5%.

    At least that's my theory. I haven't read the paper but plan to read it tonight when I have time. At first glance I'm not surprised. When I've worked with industrial ML applications, processes that have a lot of problems produce better training data than well controlled processes, and I have read papers on this subject where people have improved performance of their models by introducing (controlled) randomness into their control setpoints to get more training data outside of the tight control regime.

  • Those are actually some very good results. Funny situation, if the copyright companies win the AI legislative war, 4chan is going to get twice as much as reddit did for the data at the minimum.

    It's also interesting the model gets worse faster if it has to untrain the toxic data so to speak.

    So basically... by being familiar with 4chan the model knows better what not to do?

  • And I wish they would tone down the hype. Maybe we can meet in the middle?

    Well, I do wish they would promote the actual use and limitations of AI and stop making up crap and overselling the use cases. I use ChatGPT at work all the time as a start for research, but if I took any of it as being reliable info to run with I would be in grave trouble. It is a great tool that has saved me much time because I know how far to trust it and how to use it. The progress is very impressive as I've been using AI art services for years, and the difference between the random blobs from back then and the great stuff it can generate now is pretty stark. Same thing with the LLMs. I've been using ChatGPT since it showed up and it has improved greatly since then. Before all this I talked to people who were using AI training on various picture recognition projects where getting data from other sensors was not practical. ... Overall AI is pretty exciting, but the non-stop hype and hate headlines is doing nobody any favors.

  • 1 Stimmen
    2 Beiträge
    2 Aufrufe
    A
    If you're a developer, a startup founder, or part of a small team, you've poured countless hours into building your web application. You've perfected the UI, optimized the database, and shipped features your users love. But in the rush to build and deploy, a critical question often gets deferred: is your application secure? For many, the answer is a nervous "I hope so." The reality is that without a proper defense, your application is exposed to a barrage of automated attacks hitting the web every second. Threats like SQL Injection, Cross-Site Scripting (XSS), and Remote Code Execution are not just reserved for large enterprises; they are constant dangers for any application with a public IP address. The Security Barrier: When Cost and Complexity Get in the Way The standard recommendation is to place a Web Application Firewall (WAF) in front of your application. A WAF acts as a protective shield, inspecting incoming traffic and filtering out malicious requests before they can do any damage. It’s a foundational piece of modern web security. So, why doesn't everyone have one? Historically, robust WAFs have been complex and expensive. They required significant budgets, specialized knowledge to configure, and ongoing maintenance, putting them out of reach for students, solo developers, non-profits, and early-stage startups. This has created a dangerous security divide, leaving the most innovative and resource-constrained projects the most vulnerable. But that is changing. Democratizing Security: The Power of a Community WAF Security should be a right, not a privilege. Recognizing this, the landscape is shifting towards more accessible, community-driven tools. The goal is to provide powerful, enterprise-grade protection to everyone, for free. This is the principle behind the HaltDos Community WAF. It's a no-cost, perpetually free Web Application Firewall designed specifically for the community that has been underserved for too long. It’s not a stripped-down trial version; it’s a powerful security tool designed to give you immediate and effective protection against the OWASP Top 10 and other critical web threats. What Can You Actually Do with It? With a community WAF, you can deploy a security layer in minutes that: Blocks Malicious Payloads: Get instant, out-of-the-box protection against common attack patterns like SQLi, XSS, RCE, and more. Stops Bad Bots: Prevent malicious bots from scraping your content, attempting credential stuffing, or spamming your forms. Gives You Visibility: A real-time dashboard shows you exactly who is trying to attack your application and what methods they are using, providing invaluable security intelligence. Allows Customization: You can add your own custom security rules to tailor the protection specifically to your application's logic and technology stack. The best part? It can be deployed virtually anywhere—on-premises, in a private cloud, or with any major cloud provider like AWS, Azure, or Google Cloud. Get Started in Minutes You don't need to be a security guru to use it. The setup is straightforward, and the value is immediate. Protecting the project, you've worked so hard on is no longer a question of budget. Download: Get the free Community WAF from the HaltDos site. Deploy: Follow the simple instructions to set it up with your web server (it’s compatible with Nginx, Apache, and others). Secure: Watch the dashboard as it begins to inspect your traffic and block threats in real-time. Security is a journey, but it must start somewhere. For developers, startups, and anyone running a web application on a tight budget, a community WAF is the perfect first step. It's powerful, it's easy, and it's completely free.
  • 249 Stimmen
    232 Beiträge
    10 Aufrufe
    U
    Repair technicians see by far the most of seagate drives
  • 178 Stimmen
    118 Beiträge
    4 Aufrufe
    K
    My 2 cents is that it would have flourished a lot longer if eclipse wasn't stretched so thin like using a very thick amorphous log that is somehow still brittle? And ugly? As a bookmark.
  • 73 Stimmen
    38 Beiträge
    2 Aufrufe
    F
    For sure they are! Meta more then the others though
  • 93 Stimmen
    42 Beiträge
    0 Aufrufe
    G
    You don’t understand. The tracking and spying is the entire point of the maneuver. The ‘children are accessing porn’ thing is just a Trojan horse to justify the spying. I understand what are you saying, I simply don't consider to check if a law is applied as a Trojan horse in itself. I would agree if the EU had said to these sites "give us all the the access log, a list of your subscriber, every data you gather and a list of every IP it ever connected to your site", and even this way does not imply that with only the IP you could know who the user is without even asking the telecom company for help. So, is it a Trojan horse ? Maybe, it heavily depend on how the EU want to do it. If they just ask "show me how you try to avoid that a minor access your material", which normally is the fist step, I don't see how it could be a Trojan horse. It could become, I agree on that. As you pointed out, it’s already illegal for them to access it, and parents are legally required to prevent their children from accessing it. No, parents are not legally required to prevent it. The seller (or provider) is legally required. It is a subtle but important difference. But you don’t lock down the entire population, or institute pre-crime surveillance policies, just because some parents are not going to follow the law. True. You simply impose laws that make mandatories for the provider to check if he can sell/serve something to someone. I mean asking that the cashier of mall check if I am an adult when I buy a bottle of wine is no different than asking to Pornhub to check if the viewer is an adult. I agree that in one case is really simple and in the other is really hard (and it is becoming harder by the day). You then charge the guilty parents after the offense. Ok, it would work, but then how do you caught the offendind parents if not checking what everyone do ? Is it not simpler to try to prevent it instead ?
  • Why Japan's animation industry has embraced AI

    Technology technology
    12
    1
    1 Stimmen
    12 Beiträge
    3 Aufrufe
    R
    The genre itself has become neutered, too. A lot of anime series have the usual "anime elements" and a couple custom ideas. And similar style, too glossy for my taste. OK, what I think is old and boring libertarian stuff, I'll still spell it out. The reason people are having such problems is because groups and businesses are de facto legally enshrined in their fields, it's almost like feudal Europe's system of privileges and treaties. At some point I thought this is good, I hope no evil god decided to fulfill my wish. There's no movement, and a faction (like Disney with Star Wars) that buys a place (a brand) can make any garbage, and people will still try to find the depth in it and justify it (that complaint has been made about Star Wars prequels, but no, they are full of garbage AND have consistent arcs, goals and ideas, which is why they revitalized the Expanded Universe for almost a decade, despite Lucas-<companies> having sort of an internal social collapse in year 2005 right after Revenge of the Sith being premiered ; I love the prequels, despite all the pretense and cringe, but their verbal parts are almost fillers, their cinematographic language and matching music are flawless, the dialogue just disrupts it all while not adding much, - I think Lucas should have been more decisive, a bit like Tartakovsky with the Clone Wars cartoon, just more serious, because non-verbal doesn't equal stupid). OK, my thought wandered away. Why were the legal means they use to keep such positions created? To make the economy nicer to the majority, to writers, to actors, to producers. Do they still fulfill that role? When keeping monopolies, even producing garbage or, lately, AI slop, - no. Do we know a solution? Not yet, because pressing for deregulation means the opponent doing a judo movement and using that energy for deregulating the way everything becomes worse. Is that solution in minimizing and rebuilding the system? I believe still yes, nothing is perfect, so everything should be easy to quickly replace, because errors and mistakes plaguing future generations will inevitably continue to be made. The laws of the 60s were simple enough for that in most countries. The current laws are not. So the general direction to be taken is still libertarian. Is this text useful? Of course not. I just think that in the feudal Europe metaphor I'd want to be a Hussite or a Cossack or at worst a Venetian trader.
  • Nextcloud cries foul over Google Play Store app rejection

    Technology technology
    1
    1
    6 Stimmen
    1 Beiträge
    1 Aufrufe
    Niemand hat geantwortet
  • 9 Stimmen
    6 Beiträge
    0 Aufrufe
    N
    So they.just reinvented the DVB-T tuner. Edit: I looked it up and it's literally just that. The fact they're shoving it into feature phones is interesting.