“just tell your LLM not to do that”
You ever ask an LLM to modify a picture and “don’t change anything else”? It’s going to change other things.
Case in point: https://youtu.be/XnWOVQ7Gtzw
That’s why you always add “and no mistakes”
Also “don’t hallucinate”
And “don’t become self arrest”
You are mixing two kind of AI, LLM and diffusion.
It’s way harder for a diffusion model to not change the rest, the first step of a diffusion model is to use a lossy compression to transform the picture into a soup of digits that the diffusion model can understand.And an LLM will convert a prompt into a bunch of tokens the model can understand.
Tokens are a lossless conversion, you can convert it back to the original text.
This isn’t about saying “return the original text” this is about assuming LLMs understand language, and they don’t. Telling an LLM “don’t do these things” will be as effective as telling it “don’t hallucinate” or asking it "how many 'r’s in ‘strawberry’.
In order to make such affirmation or infirmation we’ll need to define understanding.
The example you gave can be explained by other way than “it doesn’t understand”.For example, the “how many ‘r’ in strawberry”, LLMs see tokens, and the dataset they use, doesn’t contain a lot of data about the letters that are present in a token.
What is wrong in the techbrodude head that makes them only think of ruining things? Like it seems to me that they literally spend their days looking at things that are good and saying “what can I do to fuck this up for a profit?”
Should being a techie go into the DSM-V as a subheading under narcissistic personality disorder?
Gotta disrupt to peak bro. Just one more app bro
These fuckin AI “enthusiasts” are just making the rest of the world hate AI more.
Losers who cant achieve anything without AI are just going to keep doing this shit.
Fr if they just let it go instead of forcing it on everyone people might even be enthusiastic.
Finally! Now all of the “Scientology” histories will be safe! /s
Congrats on inventing what high school students figured out a year ago to skirt AI homework detectors.
I really despise how Claude’s creators and users are turning the definition of “skill” from “the ability to use [learned] knowledge to enhance execution” into “a blurb of text that [usefully] constrains a next-token-predictor”.
I guess, if you squint, it’s akin to how biologists will talk about species “evolving to fit a niche” amongst themselves or how physicists will talk about nature “abhorring a vacuum”. At least they aren’t talking about a fucking product that benefits from hype to get sold.
I can’t help but get secondhand embarrassment whenever I see someone unironically call themselves a “prompt engineer”. 🤮
Sloperator
Hey, they had to learn thermodynamics and spend 3 semesters in calculus to write those prompts
I’m a terrible procrastinator engineer.
Isn’t this a thing that authoritarians do. They co-opt language. It’s the same thing conservatives do. The venn diagram of tech bros and the far right is too close to being a circle.
You can pretty put any word out of the dictionary into a search engine and the first results are some tech company that took the word either as their company name or redefined it into some buzzword.
Skills were functions/frameworks built for Alexa, so they just appropriated the term from there.
You do understand this is more akin to white hat testing, right?
Those who want to exploit this will do it anyway, except they won’t publish the result. By making the exploit public, the risk will be known if not mitigated.
I’m admittedly not knowledgeable in White Hat Hacking, but are you supposed to publicize the vulnerability, release a shortcut to exploit it telling people to ‘enjoy’, or even call the vulnerability handy ?
Responsible disclosure is what a white hat does. You report the bug to whomever is the party responsible for patching and give them time to fix it.
That sort of depends on the situation. Responsible disclosure is for if there is some relevant security hole that is an actual risk to businesses and people, while this here is just “haha look LLMs can now better pretend to write good text if you tell it to”. That’s not really responsible disclosurable. It’s not even specific to one singular product.
Considering the “vulnerability” here is on the level of “don’t use
passwordas your password” - yeah, releasing it all is exactly the right step.
I am so goddamned tired of AI being shoved into every collective orifice of our society.
So they are using AI to make it so AI can’t detect that they are using AI?
What kind of technological ouroborous of nonsense is this?
It gets better. Using llm’s to check if the output of an llm is hallucinated or not! They call it a judge and its funny as hell tbh
Magic
Wikipedia is one of the last genuine places on the Internet, and these rat bastards are trying to contaminate that, too
Wikipedia just sold the rights to use Wikipedia for AI training to Microsoft and openai…
How exactly does that work? Wikipedia does not “own” the content on the website, it’s all CC-BY licensed.
The BY term is not respected by LLMs
So? Still doesn’t make sense to me that wikipedia can sell anything meaningful here, but I’m also not a lawyer. Do they promise not to sue them or sell them some guarantee that contributors also can’t sue them? Is it just some symbolic PR washing?
Yeah, they’re selling the work of others. That’s how the site always worked. This venture into “AI” is nothing new.
Why? Wikipedia has like a decade of operating expenses on hand, so they don’t need the money
This number inflates every time I read it. First it was ten years of hosting cost. Then it’s operating costs. Soon it will be ten years of the entire US GDP.
I’d believe they have ten years of hosting costs on hand.
My quick googling says they have 170m in assets and all 180m in annual operating costs. Give or take.
I just love how people just shit “facts” out of their ass while citing zero sources and people will just believe them and upvote because it confirms their bias.
Greed? It’s probably greed.
It’s a non-profit foundation with the majority being volunteers. If greed was the case one then would have to ask is why not just go ahead and inject ads
OK then why sell data right to m$?
Well as mentioned Wikipedia seems to be in the red and not making enough donations to pay for the expenses. So maybe the foundation is thinking it would help with the deficit.
Also chances are Microsoft will instruct Co-Pilot to prioritize Wikipedia whenever it scours the internet for information.
Think it like that eye rolling Google paying Firefox to be the default search engine deal.
Well as mentioned Wikipedia seems to be in the red
They keep saying that… at least when they’re asking for more money.
Is wiki in the red? Unclear, omi mean they ask for money donations, but someone in this thread claims they are set for a decade, I’ve seen people post something about how they are fine, and even donate a bunch themselves. I don’t know, and I guess it doesn’t matter.
Not sure where you are going with your second comment, and uninterested in engaging with your comparison as I don’t think it’s very good
I am referring to the reply comment from surewhynotlem. They say that cost is 180 million while Wikipedia has 170 million on hand. That is a 10 million deficit.
While probably not enough to shut down the site it is still operating in the red.
Where I was going is explaining how it’s possibly not greed. Just the foundation looking for another revenue source that theoretically would not ruin the site.
That alt being a deal that gets Wikipedia more traffic
If microsoft is “buying access to training data” it makes what Open AI is doing look illegal. I would encourage every data broker to sell 'AI training data rights" because it undermines the only real advantage AI has and it helps pave the way to forcing AI companies to comply with open source licenses.
Essentially selling ai data rights is a trojan horse for the AI companies. Obviously it would be better to pass laws but until that happens this is imo a better strategy than doing nothing.
I mean, what open ai is doing and did should be illegal if it’s not, in my opinion.
I mean it’s free money, why not?
From the repo:
Have opinions. Don’t just report facts - react to them. “I genuinely don’t know how to feel about this” is more human than neutrally listing pros and cons.
That will at least be easy to spot in a Wikipedia entry.
lol brilliant
If these “signs of AI writing” are merely linguistic, good for them. This is as accurate as a lie detector (i.e., not accurate) and nobody should use this for any real world decision-making.
The real signs of AI writing are not as easy to fix as just instructing an LLM to “read” an article to avoid them.
As a teacher, all of my grading is now based on in person performances, no tech allowed. Good luck faking that with an LLM. I do not mind if students use an LLM to better prepare for class and exams. But my impression so far is that any other medium (e.g., books, youtube explanation videos) leads to better results.
I sucked in oral exams and therefore hated them. Then again, if they had been mixed into regular school, it might not have sucked so much.
Doesn’t need to be oral, I remember occasionally having exams that were essay questions that needed to be answered in class.
I do both of these as well as smaller but more frequent tests, group work, project work over several sessions etc… The only things I stopped doing are reports to write at home, paper summaries etc. Doesn’t make sense anymore.
Download an offline copy while you still can.
what are the best ways to do it while still having practical access?
Here’s a link to the Kiwix library download for all of Wikipedia. It’s 111GB though, so you’ll need a lot of space and also a lot of time to wait for it to download.
Note, you’ll also need Kiwix in some manner to read the zim file once it’s downloaded.
Kiwix library - All of Wikipedia - direct download link
But this’ll let you have a local copy you can reference should actual Wikipedia ever get ruined by GenAI, or worse, get taken down by hostile governments.
sounding like a great app for my server!
It’s an arms race, AI identification vs AI adaptation. I wonder which side the companies that own these LLMs want to win…
They don’t want anyone to win. The arms race makes money.
Fuck you, Siqi Chen.
