Видео

Shopa chittagong panch laish khaleq dadu shohal bepul upstairs Akbar dadu pervin fufu mother's name same mine baba Azizul Haq next at Dhaka elephant road please let me know how is your parents and your other two sisters

Amazing editing and story-telling - outstanding speaker. Thank you for the video!

How did you draw the loss surface?

🎯 Key Takeaways for quick navigation: 00:00 🎤 *The speaker, Mark Brighton Boach, introduces himself as a security engineer at Dropbox and discusses the topic of an AI attack called "repeated sequences."* 01:09 📦 *Dropbox is interested in AI to improve its services, including AI-powered search and question answering across Dropbox files.* 03:26 💻 *The AI/ML team at Dropbox collaborated with the security team to investigate a novel AI attack involving repeated sequences in prompts.* 05:57 🔍 *Backspaces, control characters, and meta-characters were found to produce unexpected behaviors in Transformer-based chat models.* 08:56 🕵️ *Dropbox engaged with OpenAI to address the issue but had to work on their own. They planned to open-source their findings.* 11:13 🧐 *Spaces, backslashes, and other character sequences were identified as risky inputs that could disrupt AI models.* 15:18 ⏱️ *A moderator framework was proposed to detect and block risky inputs, such as repeated sequences, to ensure model safety.* 17:32 🧰 *The repeated sequence moderator analyzes prompts for suspicious sequences and flags them based on repeat count and score.* 21:05 ❓ *The speaker addresses questions about mixing and matching special characters and discusses the potential effects.* 23:00 🤖 *The speaker considers the model's training and behavior regarding certain tokens and suggests further exploration.* 24:10 🌐 *The speaker discusses the use of UTF-8 byte order marks and Unicode characters in experiments involving repeated sequences but didn't explore Unicode bomb orderings in their work.* 25:31 📞 *A question is raised about whether prompt injection attacks are fundamentally broken and similar to phone freaking, and the speaker acknowledges the complexity of the issue without a clear solution.* Made with HARPA AIwha

🙂 *Promo sm*

So humor and excellent talk! Now I'm the fan of Nicolas Carlini

Great talk! About to head out and read the paper now. Thank you very much!

Great talk! Thanks for sharing!

Happy this guy is on our team...great advantage to not have a genius like this on the other side.

Very informative and well-organized presentation. Thank you!!

Great talk! His comparison of current state of the robustness of ML models with the crypto in 1920's is spot on.

Great talk. It's such a difficult problem that is at the heart of generalisation. The google brain paper "Adversarial Examples that Fool both Computer Vision and Time-Limited Humans" is worth a read.

Damn, this talk is epic! [ Although I disagree with his answer to the question in 39:20 'the human is the proof-of-concept that adversarial robustness exists'; humans have their own set of singular errors constrained by their visual system which gives rise to visual/optical illusions, so perhaps the goal should not be to avoid adversarial examples altogether, but to somehow align/evaluate the mistakes the machines makes with that of the human. Similar to: www.pnas.org/content/pnas/117/47/29330.full.pdf ]

An excellent talk, thank you!

only 421 views?

ruclips.net/video/otmt0-cewbc/видео.html

Activism, Human Nature, & American Repentance: Noam Chomsky interviewed by Shmuly Yanklowitz ruclips.net/video/EA7Dj-dmzEE/видео.html

How does this talk only has 300 views and two comments? This man's talk is absolutely brilliant.

very useful. Thank you!

great talk