Gergely Acs
@gergelyacs.bsky.social
Reposted by Gergely Acs
Reposted by Gergely Acs
Reposted by Gergely Acs
Excellent post from (my former😢 colleague) Nicholas Carlini on the differences between copyright law & privacy research.
In particular, from a privacy perspective, "was training data memorized?" is a yes/no question; we aren't trying to quantify how much data was memorized beyond "some" vs "none".
In particular, from a privacy perspective, "was training data memorized?" is a yes/no question; we aren't trying to quantify how much data was memorized beyond "some" vs "none".
What my privacy papers (don't) have to say about copyright and generative AI
My work on privacy-preserving machine learning is often cited by lawyers arguing for or against how generative AI models violate copyright. This maybe isn't the right work to be citing.
nicholas.carlini.com
March 11, 2025 at 8:09 PM
Excellent post from (my former😢 colleague) Nicholas Carlini on the differences between copyright law & privacy research.
In particular, from a privacy perspective, "was training data memorized?" is a yes/no question; we aren't trying to quantify how much data was memorized beyond "some" vs "none".
In particular, from a privacy perspective, "was training data memorized?" is a yes/no question; we aren't trying to quantify how much data was memorized beyond "some" vs "none".
Reposted by Gergely Acs
The big AI story of the past 6 months is: Companies now claim that their AI models are capable of genuine reasoning.
And the big question is: Is that true?
I found that the best answer lies in between hype and skepticism.
www.vox.com/future-perfe...
And the big question is: Is that true?
I found that the best answer lies in between hype and skepticism.
www.vox.com/future-perfe...
Is AI really thinking and reasoning — or just pretending to?
The best answer — AI has “jagged intelligence” — lies in between hype and skepticism.
www.vox.com
February 21, 2025 at 5:49 PM
The big AI story of the past 6 months is: Companies now claim that their AI models are capable of genuine reasoning.
And the big question is: Is that true?
I found that the best answer lies in between hype and skepticism.
www.vox.com/future-perfe...
And the big question is: Is that true?
I found that the best answer lies in between hype and skepticism.
www.vox.com/future-perfe...
Reposted by Gergely Acs
AI can never replace civil servants (morally, politically, democratically). Civil servants are moral agents who can be held accountable; their salary is paid by citizens and they owe their loyalty to citizens. AI works for companies; it has been built for profit. #AIEthics
February 22, 2025 at 5:02 PM
AI can never replace civil servants (morally, politically, democratically). Civil servants are moral agents who can be held accountable; their salary is paid by citizens and they owe their loyalty to citizens. AI works for companies; it has been built for profit. #AIEthics
Reposted by Gergely Acs
With so many people using these models this is a terrifying result
www.nature.com/articles/s41...
www.nature.com/articles/s41...
“Turning right”? An experimental study on the political value shift in large language models - Humanities and Social Sciences Communications
Humanities and Social Sciences Communications - “Turning right”? An experimental study on the political value shift in large language models
www.nature.com
February 15, 2025 at 11:44 AM
With so many people using these models this is a terrifying result
www.nature.com/articles/s41...
www.nature.com/articles/s41...
Reposted by Gergely Acs
Neat trick for bounding the maximum of independent Gaussians:
![Lemma 4.5 from https://arxiv.org/abs/1712.07196
Let X_1, ..., X_k be independent samples from N(0,1). Then
E[max{X_1^2, ..., X_k^2}] <= 2ln(2k).](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:thfnaqgktdwqbm42radc322i/bafkreidkql7bujw7docefdl3shay4r4dn7pw74slmhm3dlhtn4s3li4jyy@jpeg)
January 19, 2025 at 5:23 PM
Neat trick for bounding the maximum of independent Gaussians:
Reposted by Gergely Acs
Really enjoyed "Things we learned about LLMs in 2024" by
@simonwillison.net, especially this analogy between today's datacenter buildout and the 19th century railway boom. The parallels are striking. simonwillison.net/2024/Dec/31/...
@simonwillison.net, especially this analogy between today's datacenter buildout and the 19th century railway boom. The parallels are striking. simonwillison.net/2024/Dec/31/...
January 1, 2025 at 4:00 PM
Really enjoyed "Things we learned about LLMs in 2024" by
@simonwillison.net, especially this analogy between today's datacenter buildout and the 19th century railway boom. The parallels are striking. simonwillison.net/2024/Dec/31/...
@simonwillison.net, especially this analogy between today's datacenter buildout and the 19th century railway boom. The parallels are striking. simonwillison.net/2024/Dec/31/...
Reposted by Gergely Acs
The expansion of data centers is having serious effects on the US grid, with three-quarters of highly distorted power readings occurring within 50 miles of sites with significant data center activity.
That threatens home appliances and aging energy infrastructure, and increases fire risks.
That threatens home appliances and aging energy infrastructure, and increases fire risks.
AI Needs So Much Power, It’s Making Yours Worse
There’s a strong link between proximity to AI data centers and higher levels of distorted power in residential areas
www.bloomberg.com
December 27, 2024 at 10:00 PM
The expansion of data centers is having serious effects on the US grid, with three-quarters of highly distorted power readings occurring within 50 miles of sites with significant data center activity.
That threatens home appliances and aging energy infrastructure, and increases fire risks.
That threatens home appliances and aging energy infrastructure, and increases fire risks.
Reposted by Gergely Acs
Have you ever wondered why diffusion models memorize and all initializations lead to the same training sample? As we show, this is because like in dynamic systems, the memorized sample acts as an attractor and a corresponding attraction basin is formed in the denoising trajectory.
December 4, 2024 at 9:03 PM
Have you ever wondered why diffusion models memorize and all initializations lead to the same training sample? As we show, this is because like in dynamic systems, the memorized sample acts as an attractor and a corresponding attraction basin is formed in the denoising trajectory.
Reposted by Gergely Acs
Our Open Source Developers Guide to the EU AI Act is now live! Check it out for an introduction to the AI Act and useful tools that may help prepare for compliance, with a focus on open source. Amazing to work with @frimelle.bsky.social and @yjernite.bsky.social on this!
Open Source Developers Guide to the EU AI Act
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
December 2, 2024 at 5:06 PM
Our Open Source Developers Guide to the EU AI Act is now live! Check it out for an introduction to the AI Act and useful tools that may help prepare for compliance, with a focus on open source. Amazing to work with @frimelle.bsky.social and @yjernite.bsky.social on this!
Reposted by Gergely Acs
First dataset for the new @huggingface.bsky.social @bsky.app community organisation: one-million-bluesky-posts 🦋
📊 1M public posts from Bluesky's firehose API
🔍 Includes text, metadata, and language predictions
🔬 Perfect to experiment with using ML for Bluesky 🤗
huggingface.co/datasets/blu...
📊 1M public posts from Bluesky's firehose API
🔍 Includes text, metadata, and language predictions
🔬 Perfect to experiment with using ML for Bluesky 🤗
huggingface.co/datasets/blu...
bluesky-community/one-million-bluesky-posts · Datasets at Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co
November 26, 2024 at 1:50 PM
First dataset for the new @huggingface.bsky.social @bsky.app community organisation: one-million-bluesky-posts 🦋
📊 1M public posts from Bluesky's firehose API
🔍 Includes text, metadata, and language predictions
🔬 Perfect to experiment with using ML for Bluesky 🤗
huggingface.co/datasets/blu...
📊 1M public posts from Bluesky's firehose API
🔍 Includes text, metadata, and language predictions
🔬 Perfect to experiment with using ML for Bluesky 🤗
huggingface.co/datasets/blu...
Reposted by Gergely Acs
Medically adapted foundation models (think Med-*) turn out to be more hot air than hot stuff. Correcting for fatal flaws in evaluation, the current crop are no better on balance than generic foundation models, even on the very tasks for which benefits are claimed.
arxiv.org/abs/2411.04118
arxiv.org/abs/2411.04118
Medical Adaptation of Large Language and Vision-Language Models: Are We Making Progress?
Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pret...
arxiv.org
November 26, 2024 at 6:12 PM
Medically adapted foundation models (think Med-*) turn out to be more hot air than hot stuff. Correcting for fatal flaws in evaluation, the current crop are no better on balance than generic foundation models, even on the very tasks for which benefits are claimed.
arxiv.org/abs/2411.04118
arxiv.org/abs/2411.04118
Reposted by Gergely Acs
My two cents: "“We all have this experience that the majority of websites we go in have dark patterns. If ChatGPT is building unethical websites, it’s because it’s been trained with data of unethical websites.” #AIEthics Great article by @stokel.bsky.social. www.fastcompany.com/91233844/cha...
ChatGPT was tasked with designing a website. The result was as manipulative as you'd expect
In a new study, researchers found that ChatGPT creates websites full of deceptive patterns.
www.fastcompany.com
November 25, 2024 at 4:30 PM
My two cents: "“We all have this experience that the majority of websites we go in have dark patterns. If ChatGPT is building unethical websites, it’s because it’s been trained with data of unethical websites.” #AIEthics Great article by @stokel.bsky.social. www.fastcompany.com/91233844/cha...
Reposted by Gergely Acs
For those who missed this post on the-network-that-is-not-to-be-named, I made public my "secrets" for writing a good CVPR paper (or any scientific paper). I've compiled these tips of many years. It's long but hopefully it helps people write better papers. perceiving-systems.blog/en/post/writ...
Writing a good scientific paper
perceiving-systems.blog
November 20, 2024 at 10:18 AM
For those who missed this post on the-network-that-is-not-to-be-named, I made public my "secrets" for writing a good CVPR paper (or any scientific paper). I've compiled these tips of many years. It's long but hopefully it helps people write better papers. perceiving-systems.blog/en/post/writ...
Reposted by Gergely Acs
Just realized BlueSky allows sharing valuable stuff cause it doesn't punish links. 🤩
Let's start with "What are embeddings" by @vickiboykis.com
The book is a great summary of embeddings, from history to modern approaches.
The best part: it's free.
Link: vickiboykis.com/what_are_emb...
Let's start with "What are embeddings" by @vickiboykis.com
The book is a great summary of embeddings, from history to modern approaches.
The best part: it's free.
Link: vickiboykis.com/what_are_emb...
November 22, 2024 at 11:13 AM
Just realized BlueSky allows sharing valuable stuff cause it doesn't punish links. 🤩
Let's start with "What are embeddings" by @vickiboykis.com
The book is a great summary of embeddings, from history to modern approaches.
The best part: it's free.
Link: vickiboykis.com/what_are_emb...
Let's start with "What are embeddings" by @vickiboykis.com
The book is a great summary of embeddings, from history to modern approaches.
The best part: it's free.
Link: vickiboykis.com/what_are_emb...
Reposted by Gergely Acs
We had a great discussion with @robertlemos.bsky.social from Dark Reading about our new paper "Hacking Back the AI-Hacker: Prompt Injection as a Defense Against LLM-driven Cyberattacks"(arxiv.org/abs/2410.20911). Mantis turns the hardness of dealing with prompt injections into an opportunity!
AI About-Face: 'Mantis' Turns LLM Attackers Into Prey
Experimental counter-offensive system responds to malicious AI probes with their own surreptitious prompt-injection commands.
www.darkreading.com
November 21, 2024 at 1:22 PM
We had a great discussion with @robertlemos.bsky.social from Dark Reading about our new paper "Hacking Back the AI-Hacker: Prompt Injection as a Defense Against LLM-driven Cyberattacks"(arxiv.org/abs/2410.20911). Mantis turns the hardness of dealing with prompt injections into an opportunity!
Reposted by Gergely Acs
Our relationship with big tech is unbalanced, argues @hartzog.bsky.social. #AIEthics www.techpolicy.press/what-kafka-c...
What Kafka Can Teach Us About Privacy in the Age of AI | TechPolicy.Press
A conversation with Boston University School of Law professor Woodrow Hartzog.
www.techpolicy.press
November 16, 2024 at 6:17 PM
Our relationship with big tech is unbalanced, argues @hartzog.bsky.social. #AIEthics www.techpolicy.press/what-kafka-c...
Reposted by Gergely Acs
This is a fantastic tool if you’re looking for people with overlapping interests (as judged by networks) to follow:
bsky-follow-finder.theo.io
bsky-follow-finder.theo.io
Bluesky Network Analyzer
Find accounts that you don't follow (yet) but are followed by lots of accounts that you do follow.
bsky-follow-finder.theo.io
November 18, 2024 at 2:49 AM
This is a fantastic tool if you’re looking for people with overlapping interests (as judged by networks) to follow:
bsky-follow-finder.theo.io
bsky-follow-finder.theo.io
Reposted by Gergely Acs
Since this platform is finally attracting a critical mass of ML researchers, here's our recent work on prompt-based vulnerabilities of coding assistants:
arxiv.org/abs/2407.11072
TL;DR — An attacker can convince your favorite LLM to suggest vulnerable code with just a minor change to the prompt!
arxiv.org/abs/2407.11072
TL;DR — An attacker can convince your favorite LLM to suggest vulnerable code with just a minor change to the prompt!
MaPPing Your Model: Assessing the Impact of Adversarial Attacks on LLM-based Programming Assistants
LLM-based programming assistants offer the promise of programming faster but with the risk of introducing more security vulnerabilities. Prior work has studied how LLMs could be maliciously fine-tuned...
arxiv.org
November 17, 2024 at 11:41 PM
Since this platform is finally attracting a critical mass of ML researchers, here's our recent work on prompt-based vulnerabilities of coding assistants:
arxiv.org/abs/2407.11072
TL;DR — An attacker can convince your favorite LLM to suggest vulnerable code with just a minor change to the prompt!
arxiv.org/abs/2407.11072
TL;DR — An attacker can convince your favorite LLM to suggest vulnerable code with just a minor change to the prompt!