
A small number of samples can poison #LLMs of any size
"Specifically, we demonstrate that by injecting just 250 malicious documents into pretraining data, adversaries can successfully backdoor LLMs ranging from 600M to 13B parameters."
cc @asrg
#Tag
A small number of samples can poison #LLMs of any size
"Specifically, we demonstrate that by injecting just 250 malicious documents into pretraining data, adversaries can successfully backdoor LLMs ranging from 600M to 13B parameters."
cc @asrg
A small number of samples can poison #LLMs of any size
"Specifically, we demonstrate that by injecting just 250 malicious documents into pretraining data, adversaries can successfully backdoor LLMs ranging from 600M to 13B parameters."
cc @asrg
A few people have asked my advice on this recently so I’m sharing here in case it’s useful:
If it helps, I agonised about this in my role as a literary executor (cared much less about my own work) and reached the conclusion that diffusion of the ideas is best served by being incorporated into training. I wouldn’t expect everyone to reach the same conclusion but I hope it’s useful to make these suggestions about factors to consider.
#intellectualProperty #LLMs #postTraining #publishing #scholarlyPublishing #Training #visibility
CWA just shared their research into the lives of the US-based data workers responsible for training and maintaining modern “AI” platforms. The findings are, in a word, grim: https://cwa-union.org/ghost-workers-ai-machine
Our team member @epoz presented "Now that we have Large Language Models, are metadata standards still necessary?" at the Autumn School 2025 ‘Modern Stained Glass – Metadata – AI’ at University of Münster, Faculty of Catholic Theology
https://zenodo.org/records/17151141
#llms #generativeAI #metadata #iconclass #AI #arthistory #dh #digitalhumanities #culturalheritage #elephant #chatgpt @fiz_karlsruhe @nfdi4culture @NFDI4Memory
I'm all against irresponsible uses of LLMs, but users probably wouldn't turn to a natural language prompt to get a properly understandable doc if the documentation effort was decent enough upstream.
Documentation is hard and boring most of the time, but it's what makes your product usable. Bad doc is plain gatekeeping.
Our team member @epoz presented "Now that we have Large Language Models, are metadata standards still necessary?" at the Autumn School 2025 ‘Modern Stained Glass – Metadata – AI’ at University of Münster, Faculty of Catholic Theology
https://zenodo.org/records/17151141
#llms #generativeAI #metadata #iconclass #AI #arthistory #dh #digitalhumanities #culturalheritage #elephant #chatgpt @fiz_karlsruhe @nfdi4culture @NFDI4Memory
This is the third and last in my series on #UX and #LLMs. It was the hardest to write as I (rather foolishly) take a stand on where we should go. My basic point is that the current AI bubble is going to pop. So what should replace it? I make the case that smaller, more energy-efficient, and more ethically trained models should be used for "boring" tasks. Our job is to solve problems, not look cool.
”…the AI bubble is ripe for bursting. Things like the efficient compute frontier…and the Floridi conjecture…mean that the AI models we have now are about as good as they will ever be. … This is a huge problem! Because, as they currently stand, generative AI models aren’t actually that useful or even remotely profitable.”
—Will Lockett, The AI Bubble Is About To Burst, But The Next Bubble Is Already Growing
#ai #generativeai #llm #llms
“ #LLMs are a #technology suited to a decadent #culture, one that chases easy profits rather than tackles the real challenges we face. It’s easier to make money rearranging words according to various probabilities than it is to make a living improving the #health of our topsoil, #communities, and souls.”
“#SocialMedia didn’t democratize speech so much as it degraded our informational #ecosystem; it democratized the need to be vigilant and discerning as we sift through vast quantities of debris in search of reliable #institutions and #information.”
“ #LLMs are a #technology suited to a decadent #culture, one that chases easy profits rather than tackles the real challenges we face. It’s easier to make money rearranging words according to various probabilities than it is to make a living improving the #health of our topsoil, #communities, and souls.”
“Solviture ambulando, it is solved by #walking, is the only fitting response to some queries, yet no LLM can provide answers of this type.”
“Interacting with #LLMs trains us to expect pre-formed answers from the void rather than endure the #productive effort of #conversation – literally, turning back and forth with another person in pursuit of #truth.”
"Claude’s update relies on a striking pop-up with a large, black "Accept" button. The data sharing toggle is tucked away, switched on by default, and framed positively ("You can help..."). A faint "Not now" button and hard-to-find instructions on changing the setting later complete the manipulative design.
These interface tricks, known as dark patterns, are considered unlawful under the General Data Protection Regulation (GDPR) and by the European Court of Justice when used to obtain consent for data processing. Pre-checked boxes do not count as valid consent under these rules.
The European Data Protection Board (EDPB) has also stressed in its guidelines on deceptive design patterns that consent must be freely given, informed, and unambiguous. Claude’s current design clearly fails to meet these standards, making it likely that Anthropic will soon draw the attention of privacy regulators."
#EU#AI#GenerativeAI#Anthropic #LLMs #Chatbots#Claude#DarkPatterns#Privacy#DataProtection
"Claude’s update relies on a striking pop-up with a large, black "Accept" button. The data sharing toggle is tucked away, switched on by default, and framed positively ("You can help..."). A faint "Not now" button and hard-to-find instructions on changing the setting later complete the manipulative design.
These interface tricks, known as dark patterns, are considered unlawful under the General Data Protection Regulation (GDPR) and by the European Court of Justice when used to obtain consent for data processing. Pre-checked boxes do not count as valid consent under these rules.
The European Data Protection Board (EDPB) has also stressed in its guidelines on deceptive design patterns that consent must be freely given, informed, and unambiguous. Claude’s current design clearly fails to meet these standards, making it likely that Anthropic will soon draw the attention of privacy regulators."
#EU#AI#GenerativeAI#Anthropic #LLMs #Chatbots#Claude#DarkPatterns#Privacy#DataProtection
"My gut instinct is that this is an industry-wide problem. Perplexity spent 164% of its revenue in 2024 between AWS, Anthropic and OpenAI. And one abstraction higher (as I'll get into), OpenAI spent 50% of its revenue on inference compute costs alone, and 75% of its revenue on training compute too (and ended up spending $9 billion to lose $5 billion). Yes, those numbers add up to more than 100%, that's my god damn point.
Large Language Models are too expensive, to the point that anybody funding an "AI startup" is effectively sending that money to Anthropic or OpenAI, who then immediately send that money to Amazon, Google or Microsoft, who are yet to show that they make any profit on selling it.
Please don't waste your breath saying "costs will come down." They haven't been, and they're not going to.
Despite categorically wrong boosters claiming otherwise, the cost of inference — everything that happens from when you put a prompt in to generate an output from a model — is increasing, in part thanks to the token-heavy generations necessary for "reasoning" models to generate their outputs, and with reasoning being the only way to get "better" outputs, they're here to stay (and continue burning shit tons of tokens).
This has a very, very real consequence."
https://www.wheresyoured.at/why-everybody-is-losing-money-on-ai/
#AI#GenerativeAI#BusinessModels #LLMs #Chatbots#AIHype#AIBubble
"In addition to violation of data privacy, other risks are involved when psychotherapists consult LLMs on behalf of a client. Studies have found that although some specialized therapy bots can rival human-delivered interventions, advice from the likes of ChatGPT can cause more harm than good.
A recent Stanford University study, for example, found that chatbots can fuel delusions and psychopathy by blindly validating a user rather than challenging them, as well as suffer from biases and engage in sycophancy. The same flaws could make it risky for therapists to consult chatbots on behalf of their clients. They could, for example, baselessly validate a therapist’s hunch, or lead them down the wrong path.
Aguilera says he has played around with tools like ChatGPT while teaching mental health trainees, such as by entering hypothetical symptoms and asking the AI chatbot to make a diagnosis. The tool will produce lots of possible conditions, but it’s rather thin in its analysis, he says. The American Counseling Association recommends that AI not be used for mental health diagnosis at present.
A study published in 2024 of an earlier version of ChatGPT similarly found it was too vague and general to be truly useful in diagnosis or devising treatment plans, and it was heavily biased toward suggesting people seek cognitive behavioral therapy as opposed to other types of therapy that might be more suitable."
https://www.technologyreview.com/2025/09/02/1122871/therapists-using-chatgpt-secretly/
#AI#GenerativeAI #Chatbots#ChatGPT #LLMs#MentalHealth#Therapy
»KI im Bewerbungsprozess – LLMs bevorzugen KI-generierte Lebensläufe:
Noch stärker fällt die Bevorzugung aus, wenn Bewerbungsunterlagen vom selben Sprachmodell verfasst wurde.«
Wenn dies nicht Marketing ist von Untertanen die von der KI gesteuert werden? Vermehrt wird nun auch von grossen Unternehmen Bewerbungen ohne jeglichen KI Einfluss verlangt.
#ki #bewerbung #LLMs #lebenslauf #Bevorzugung #Sprachmodell #arbeit #job #work #AIjobs
»KI im Bewerbungsprozess – LLMs bevorzugen KI-generierte Lebensläufe:
Noch stärker fällt die Bevorzugung aus, wenn Bewerbungsunterlagen vom selben Sprachmodell verfasst wurde.«
Wenn dies nicht Marketing ist von Untertanen die von der KI gesteuert werden? Vermehrt wird nun auch von grossen Unternehmen Bewerbungen ohne jeglichen KI Einfluss verlangt.
#ki #bewerbung #LLMs #lebenslauf #Bevorzugung #Sprachmodell #arbeit #job #work #AIjobs
Big News! The completely #opensource#LLM#Apertus 🇨🇭 has been released today:
📰 https://www.swisscom.ch/en/about/news/2025/09/02-apertus.html
🤝 The model supports over 1000 languages [EDIT: an earlier version claimed over 1800] and respects opt-out consent of data owners.
▶ This is great for #publicAI and #transparentAI. If you want to test it for yourself, head over to: https://publicai.co/
🤗 And if you want to download weights, datasets & FULL TRAINING DETAILS, you can find them here:
https://huggingface.co/collections/swiss-ai/apertus-llm-68b699e65415c231ace3b059
🔧 Tech report: https://huggingface.co/swiss-ai/Apertus-70B-2509/blob/main/Apertus_Tech_Report.pdf
After #Teuken7b and #Olmo2, Apertus is the next big jump in capabilities and performance of #FOSS #LLMs, while also improving #epistemicresilience and #epistemicautonomy with its multilingual approach.
I believe that especially for sensitive areas like #education, #healthcare, or #academia, there is no alternative to fully open #AI models. Everybody should start building upon them and improving them.
#KIMündigkeit#SovereignAI#FOSS#ethicalAI #swissai#LernenmitKI
A space for Bonfire maintainers and contributors to communicate