RT @robusthq: In preliminary analysis of GPT-4 vs GPT-4o, we saw higher susceptibility to jailbreaks on GPT-4o with fewer queries. Higher p…

@joshua_saxe @rharang It's not plagiarism if it changes a few words.

@rharang I had a very conservative friend tell me a long time ago that public scientific funding is a waste of money since you can just read publications from China if you ever need to figure it out.

@CamlisOrg Very accessible. I really liked the presentation.

@CamlisOrg Everyone on the team was super excited about this project. Not sure what it means that our attack project was getting way more external interests than any defense work.

@CamlisOrg Benchmarking LLMs is really hard, yet important because there are so many choices. My feelings are that most common benchmarkers leaked into training already. Happy to see new type of benchmarks by my team (now former) on this. @AdarshKyadige and the rest.

@CamlisOrg Excited to see @tamasV2, my and @SPBergeron work being presented.

RT @CamlisOrg: First talk of the session: "Web content filtering through knowledge distillation of Large Language Models" by Tamas Voros! h…

@aphillippy Where are you going?

@biggiobattista Basically, I don't think worst-case threat modeling is super productive in the security space. There are only limited number of problems that can be realistically solved under those constraints. I think it encourages paralysis and does not move the conversation.

@joshua_saxe You lock the content behind a paywall or a registration page with terms and conditions that prevent scraping.

@joshua_saxe Why wouldn't everyone just move behind some registration wall?

RT @CamlisOrg: Reminder to submit your full length paper or extended abstract by Friday, August 4th to be considered for this year's confer…

@generatorman_ai @joshua_saxe Not sure what you mean here. It is about how the open internet will look like in a few years. The business model has been ads, so either lawsuits win out or the open internet. Just a statement of reality not morality. As a content creator I would be very worried.

@generatorman_ai @joshua_saxe The concern is about LLMs, but it is hard and expensive to make a good one. However, with an open source model exploitation of the learned data for profit is much easier for a random person.

RT @generatorman_ai: @kberlin @joshua_saxe Why is this a concern about open-source models rather than language models in general?

@joshua_saxe I used it multiple time to help build a custom travel itinerary. The major attractions do not really change much over a few years. This is one of the obvious cases for LLM over search, but the content it uses to learn the itinerary was obviously meant to be behind ads.

Check out a quick writeup about our work on LLM for web classification.