Only a few pages into this paper that I just learned about via @timnitGebru, but am hugely intrigued by the findings and what this must eventually mean for regulation and policy around LLMs.
Essentially these models will remain insecure as they increase in size and lead to all sorts of serious problems. On the one hand they will inevitably contain significant amounts of privacy-sensitive information. On the other hand they will continue to be vulnerable to poisoning attacks. There is not enough trustworthy content in the world and to remain safe they must be reduced in size - and yes, performance must drop.
These paragraphs provide a small glimpse of the dilemma(s) outlined in the paper:
« Foundation models achieve their best performances by leveraging ever larger amounts of data, without major diminishing returns so far. Unfortunately, verified texts seem insufficient to reach state-of-the-art performance. Indeed, the English Wikipedia only contains around 4 billion words. Meanwhile, a book has around 10^5 words. While there are 10^8 books, only a fraction of them are arguably trustworthy. Many books are instead full of biases and dangerous misinformation, such as ethnic-based hate speech, historical propaganda, or outdated (possibly harmful) medical advice. As a striking illustration, up to the 1980s, the American Psychiatric Association listed homosexuality as a mental illness in its flagship manual. In fact, most books should be regarded as unverified user-generated data.
Most importantly, even if they are not problematic, the combination of these books represents a small amount of data, compared to what Internet users produce on a daily basis. Indeed, assuming that a user writes 300 words per day on an electronic device (the equivalent of one page), a billion of such users produce 10^(to the power of)15 words per decade. This adds up to a hundred times more data than the set of books, and a million times more than the English Wikipedia. This makes it very tempting to either scrape the web, exploit private messaging (e.g., emails, shared documents), or leverage other written texts (e.g., phones’ smart keyboards). In fact, Wikipedia represented only 4% of Google’s PaLM foundation model training dataset [33], while books represented 13% of it. Meanwhile, 27% of the dataset was made of webpages, and 50% were social media conversations. Crucially, these data are generated by a myriad of users, who may be malicious and/or unaware that their activities are being leveraged to train foundation models. This raises serious security and privacy risks, even when we (wrongly) restrict our attention to publicly released models (most of the actual foundation model training is likely performed secretly by private groups). »
https://dair-community.social/@timnitGebru/110147589284016869
#AIHype #AIEthics #DigitalEthics
Tagged: Akkoma
@Beantin
But think of all the experience you gained. 😁
@plausible
@tysonsw Thanks. The logistics of my day didn't allow me to follow up on this 😅
@tysonsw
Don't have Word unfortunately. But it makes me want to create some example paragraphs to compare how they deal with the types of metaphors that most struggle with.
@ChrisJKelly
Haha, true!
The interesting/worrying thing is at what point someone else decides that translators are not needed. I'm worried all the current hype is pushing these thoughts. And that will of course likely have its own backlash.
@ChrisJKelly
For that second phrase "This truly feels like a red herring." I got
"Detta känns verkligen som en röd tråd."
😂
So essentially "This feels like common thread" or "common theme".
Which in some ways is worse because it may at first glance feel right but is also not at all what I meant.
(And yes I included my full paragraph to provide context for the tool).
I wrote my communication science dissertation on translation. I'm far from trusting a computer with it based on my experiences. 😅 What worries me is when it provides plausible answers with assertiveness. Many people will miss those mistakes. I'd rather it admit weaknesses.
I write a lot of bilingual content and often use Google Translate to help me out. I'm blown away by how seriously bad it still is after all these years.
Rarely can I use a sentence directly from a Google translation.
Just now when translating from English a paragraph starting with "Bad actors with nefarious intent..." it opted for the Swedish word for stage/film actor.
It's great for laughs though.
For you Swedish speakers:
"Dåliga skådespelare med ondskefulla avsikter ges enorma befogenheter att orsaka förödelse med skadlig programvara och bedrägeri, men också att uppfinna kemiska vapen."
Another favorite added to my records:
"This truly feels like a red herring." became "Det här känns verkligen som en röd sill."
i.e. a literal translation. It's funny because of how Swedes do love pickled herring.
But it's extra funny because of the context of that sentence. I am talking about that open letter on the AI pause and why they chose 6 months as their baseline. So the red herring relates to my take that we are pretty far from sentience and 6 months is a red herring.
(For Swedes unfamiliar with the phrase, a better translation would be "villospår")
@smirandafield Good point. And if they want they can block the bot.
@smirandafield
I recently wrote a reply to one of George Takei's posts and noted that many would likely appreciate alt-texts (he has 40+ images on here and has added alt-text 4 times – I checked). He also has a 7+ team working on social media for him. But you wouldn't believe the type of pushback I got. 😬 On the other hand... yeah, you probably would.
https://axbom.me/notice/AUD1CDppcGBbiM8kvQ
@emilymbender I also did a double-take on the use of the phrase ”self proclaimed”. That newsletter was clearly written to discredit and try to save face. It’s quite obnoxious how some of these people got into bed with a movement that denies the severity of the climate crisis (and worse) and are adamant in arguing why finding ”common ground” with them is the best choice.
Förra veckan läste jag en text av barnrättskämpen Marie Stenman om hur det vore om ens arbetsvecka var som barnens skolvecka. Den berörde mig. Djupt.
Igår satt jag och ritade bilder till texten och nu finns denna webbsida:
https://skolvecka.xbm.se/
@Jimijamflimflam Haha, interesting, I actually took that to mean vertically-aligned, which then makes sense when you refer to top and bottom.
@Jimijamflimflam
Magnificent image description. Adding this as an example in my course materials. With your consent of course 😊
@georgetakei
I believe many people would appreciate it if you add image descriptions.
⚠️ Mention of self-harm instigated by chatbot
The thing to keep an eye on now is how seriously this tragic event will be taken by manufacturers, politicians, researchers and journalists and what specific actions will be taken by each.
In my post about AI responsibility in a hyped-up world I wrote this:
“And when mental illness can be intensified by a machine and lead to horrific consequences, it is easily waved away. It appears no one needs to be held responsible for the machine’s impact. Nothing needs to be regulated or recalled.
Some people matter less than others.”
The consequences can be predicted. They can be managed before tragedy strikes. The reason they are not predicted and managed is because there is no interest or understanding on the manufacturer side. So who should be allowed to be building and releasing these tools to the masses?
https://www.vice.com/en/article/pkadgm/man-dies-by-suicide-after-talking-with-ai-chatbot-widow-says
P.S. Please refer to your national health agency for information on who you can contact when you need someone to talk to for support.
Italy first out in banning ChatGPT out of privacy concerns. They will be investigating:
- the recent data breach (on March 20)
- the lack of consent given for how any personal data is used (what data is being retained and for how long, potentially for training purposes)
- the lack of an age verification system
“OpenAI has been given 20 days to respond to the [the Italian data protection] agency's concerns, or the company could face a fine of either $21 million or 4% of its annual revenue.”
https://www.inkl.com/a/byAalWsVPkx
#AIHype #ChatGPT #GDPR
@janellecshane
Haha, thank you his made my day.
Damnit. I tend to forget how bad Mastodon is at markdown. On my end it all looks so pretty, and then I try looking at it on Mastodon and it’s a complete mess.
I always wonder why Mastodon allows links but not bullet lists 🤯
#Akkoma
My writeup of the open letter debacle, with so many links and references(!)
Why that open letter urging an AI development pause is problematic
Among other things:
- The host organisation
- Essentially no mention of all the current harm
- Boosting the idea of sentience
- Misleading citations
- The choice of "6 months"
- We vs. they
https://axbom.com/ai-open-letter-problematic/
#AIHype #AIEthics #DigitalEthics
The main reason the technocrat billionaires are conjuring AI at the same time as they explain how afraid they are the of thing they are conjuring: it’s because they believe they have a better chance of not suffering the abuse everyone else is suffering if they can claim to be its master.
This is what they mean with the word “loyal” in that open letter.
#AIHype #AIEthics #DigitalEthics
My list of real and potential harms by generative models already in motion, that the open letter fails to acknowledge:
» Security issues such as data privacy and breaches, that have already happened
» The fact that these tools are trained on vast amounts of biased data and serve to perpetuate that bias
» The fact that workers in Kenya and elsewhere are being exploited to train these tools, having to suffer harm to remove harmful content. A practice long employed by social media companies.
» An increase in the capture of biometrically inferred data that will severely impact human free will as it enables widespread personal manipulation (deepfakes) and gives authoritarian regimes more power to suppress dissent. Or encourages democracies to move towards authoritarianism, when putting the disenfranchised in harms way .
» Risks to climate due to significant energy use in large neural network training. It’s valid to note that the ones who benefit the most from AI are the rich, and the ones who suffer most from the climate crisis are the poor. The latter group don’t appear to get a lot of say in what these suggested 6 months of pause mean for them.
» Bad actors with nefarious intent becoming hugely empowered to do damage with malware and scams , but also to invent chemical weapons .
» How the tools are already disrupting art, literature and education (including ownership of all training data) without opportunity to address these issues in a reasoned manner
» Exclusion of a large part of the global population simply due to the limited number of languages that these tools are trained on.
» Unsubstantiated claims of sentience that lead to unfounded fears (a harm that the letter itself contributes to)
#AIHype #AIEthics #DigitalEthics
I also urge you to read this just-published message: Statement from the listed authors of Stochastic Parrots on the “AI pause” letter.
Saying it like it is are @timnitGebru, @meg, @emilymbender and Angelina McMillan-Major. Anyone interested in the topic of generative models should of course be following them all and reading their work.
Quote: «While we are not surprised to see this type of letter from a longtermist organization like the Future of Life Institute, which is generally aligned with a vision of the future in which we become radically enhanced posthumans, colonize space, and create trillions of digital people, we are dismayed to see the number of computing professionals who have signed this letter, and the positive media coverage it has received. It is dangerous to distract ourselves with a fantasized AI-enabled utopia or apocalypse which promises either a "flourishing" or "potentially catastrophic" future. Such language that inflates the capabilities of automated systems and anthropomorphizes them, as we note in Stochastic Parrots, deceives people into thinking that there is a sentient being behind the synthetic media. This not only lures people into uncritically trusting the outputs of systems like ChatGPT, but also misattributes agency. Accountability properly lies not with the artifacts but with their builders.»
FULL TEXT here:
https://www.dair-institute.org/blog/letter-statement-March2023
#AIHype #AIEthics #DigitalEthics
@johan_m_o
Great post, thank you for sharing.
I agree about the Fediverse. I feel like there is so much more reflection and consideration here. It's obviously changed a lot over the past year but I do hope the air of reasoned conversation will prevail when the opposite isn't rewarded.