Cryptographers Show That AI Protections Will Always Have Holes | Quanta Magazine Large language models such as ChatGPT come with filters to keep certain info from getting out. A new mathematical argument shows that systems like this can never be completely safe.

If you swap each letter in “bomb” with the next letter in the alphabet, you’ll get “cpnc.” Recently, scientists showed that and other methods can bypass filters on LLMs like Gemini, DeepSeek and Grok. @peterha2l.bsky.social reports: www.quantamagazine.org/cryptographe...

FOCS 2025 Videos from the 66th IEEE Symposium on Foundations of Computer Science (FOCS) 2025

The pre-recorded videos (for the talks whose authors opted in to submit one) are now available on our YouTube channel: m.youtube.com/@FOCS2025

They're also linked from the main website's schedule!

The researchers Lijie Chen (left), Igor Oliveira (top-right) and Jitau Li recently used “reverse mathematics” to connect a theorem called the pigeonhole principle to a whole class of complexity problems. www.quantamagazine.org/reverse-math...

Reminder that this is today at 7 pm! Please join us if you are at #NeurIPS2025

If you are interested in learning with abstentions (in the predictive or generative setting), calibration, multigroup fairness, uncertainty quantification, conformal prediction, or any other related topics, I would love to chat! I’ll be in Exhibit Hall C,D,E #3013 this Thursday 11 am - 2 pm PST.

“Selective Omniprediction and Fair Abstention”

Hello Bluesky! Joining the app in time to say that I’ll be attending NeurIPS this week! I will be presenting our paper (joint work with Varun Kanade) on building predictors that abstain optimally and fairly.

neurips.cc/virtual/2025...