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 a “guns don’t kill people - people kill people” kind of scenario.
As a standalone thing, LLMs are awesome.
What sucks is greedy people using them for the wrong reasons.
It’s like robots. Playing with robots are awesome. Firing 1,000 people and replacing them with robots - and not sharing the benefits with the community sucks.
They really aren’t though and that is half the problem. Everyone pretends they are awesome when the results are unusable garbage 80% of the time which makes them unusable for 99% of practical applications.
That’s why I said “as standalone things.” As a computing curiosity, they’re amazing. No language processing application like this existed 30 years ago when I was a kid. You could also see “talking computers” speaking naturally, pretending or not, on movies and TV shows.
Those numbers are baseless exaggerations. There are plenty of tasks which they solve perfectly, today. It’s just that a bunch of dicks operate them, and the cost of operating them are way too high.
Also:
It’s not that they’re not useful, that’s just nonsense.
Name a single task you would trust an LLM on solving for you that you feel confident would be correct without checking the output. Because that is my definition of perfectly and AI falls very, very far short of that.
“Hey AI, write me a random poem about taladar.”
i used it when i traveled to japan to ask it for english->japanese translations. it gave back results for multiple contexts, politeness levels, and broke down each sentence into its parts. my native speaker friends validated a few responses.
if youre going to be pedantic about “perfect” then nothing, not even a human, is going to live up.
willful ignorance about the things ai can be good at today is not going to do any favors for your fight against ai in the future. know your enemy and all that.
Who says you can’t check their outputs? It’s much faster to e. g. read a generated text than to write everything yourself. Same applies to translations, they’ve been excellent for quite a while now.
Business communication can be handled effortlessly by AI. Of course you read the result before you send it out, but that takes an order of a magnitude less time than formulating and typing all those meaningless sentences.
And honestly, that’s a perfect use case for AI. I wouldn’t compose a love letter to my family using AI, but a pamphlet, feature description, sales pitch, any bullshit presentation deck? You bet AI excels at those.
Same applies to content summaries that help augment search indices. Finding a large number of content candidates (e. g. videos) and have AI summarize the contents of said videos to narrow down the search is helpful and works today.
I’m not looking for AGI. I’m looking for tools to make my life easier, but in an ethical manner that doesn’t advance the destruction of the planet at an exponential rate, just for some tech bro to jerk it and buy another yacht.
You can make a generic fill in the blanks for all of those like I do and just change the key terminology for each scenario. LLMs are competing with search and replace?
I think this may be a skill issue on your part.
That’s a bit too dismissive. I’ve had a lot of interesting chats with LLMs that led me to find out what I didn’t understand about something. As an example I’m reading a book explaining some practices of Structured Concurrency in Swift and many times I asked ChatGPT is the author is correct about some phrasing that seemed wrong to me. And ChatGPT was able to explain why that was right in that context.
They are essentially a fun toy for most people, and an ok tool for people with the patience and training to get useful output from them. And they cost an insane amount of money to train and an insane amount of power to run.
Not to mention the other cost of training them, the human emotional cost. And the human cost of running them.
It just costs so much of a variety of things, for an output that has barely made anything better. Maybe they might get “better” in the future, and have to get through this stage to get there, but I’ve also seen a lot of people saying they appear to be starting to plateau… maybe a temporary plateau, but if so, how temporary? Could we just drop it for 10 years and start back up when they won’t be as inefficient? Maybe a law that they have to pay for everything they feed it, would effectively cause them to only emerge at a time when they are actually feasible.
People who track performance (like METR, a nonprofit) indicate that progress is, if anything, speeding up. Most people’s use case is so simple they can’t detect the difference. However for cases like complex problem solving, agentic tasks, etc you can in fact see significant progress happening. This should be concerning if you think the world isn’t ready for labor displaced by LLMs.