dream-logic is more powerful than logic-logic and oral cultures must encode knowledge into powerful meme-spells newsletter.squishy.computer/p/llms-and-h...

#MachineLearning loves data. #Topology loves shape. Let's bring them together! 👉The Euler Characteristic Transform (ECT) bridges both worlds, as I outline in a recent article in the Notices of the @amermathsoc.bsky.social Engage! 🧵1/n

A real buzz as we kicked off our conference yesterday! An informative and collaborative day with key talks from Kathlén Kohn (KTH) and Benjamin Guedj (UCL), and a poster reception in the afternoon. Today is dedicated to exchanging knowledge and forging links with industry colleagues. More soon!

It’s been a busy first year for the Erlangen AI Hub building our team of talented postdocs! At our Year 2 Kick Off at Oxford Mathematics this week, it was great to bring the hub team together, exchange knowledge and ideas, and share some of the group’s research at the intersection of #maths and #AI

I'm grateful to share - I was awarded the MSc Marjorie McDermott scholarship by the Department of Mathematics at Imperial College London 🥳 In receiving this, I want to thank the math community at Imperial for being so inclusive and supportive of my math journey (especially as an outsider coming in).

I'm looking forward to the Erlangen AI Hub's first conference! If you want to learn more about the Erlangen AI Hub, read the page below, and come join us if the research direction excites you! www.cs.ox.ac.uk/projects/AIH...

Thank you @satml.org for hosting a great conference in hygge-central Copenhagen! Running the LLMail-Inject competition was an incredibly rewarding experience - often in unexpected ways - so it was an honor to present it and share the results of our preliminary analysis of the attack submissions.

Back by popular demand, we have launched Phase 2 of the LLMail Indirect Prompt Injection Challenge, with an additional $6,000 in prizes! This round features a conformal-prediction-based blocklist and updated defences, including TaskTracker trained on more expressive activation signals.