Surveillance is creeping in everywhere, in unexpected places. #defcon
youtu.be/WCnojaEpF2I?...

Wikipedia is a jewel that humanity shall protect as long as it's still possible

youtu.be/9zi0ogvPfCA?...

Famous arXiv has now fallen victim to AI, both directly and indirectly. Many people trying to leverage the platform's reputation started posting low-quality or straight-up promotional papers without real peer review.

📢 Join our Conversational AI Reading Group!
📅 Thursday, Oct 23rd | 11 AM - 12 PM EST
🎙 Speaker: Joan Serrà ( @serrjoa.bsky.social
) - Sony AI
📖 Topic: "Supervised contrastive learning from weakly-labeled audio segments for musical version matching"
🔗 Details: (poonehmousavi.github.io/rg)

K2-Think is a reasoning system that achieves sota performance with a 32B parameters, matching or surpassing much larger models like GPT-OSS 120B and DeepSeek v3.1. It shows that smaller models can compete at the highest levels by combining advanced post-training and test-time computation techniques.

Large Behavior Models and Atlas Find New Footing:
youtu.be/HYwekersccY?...

The EEG Foundation Challenge is open (a NeurIPS competition). Two fascinating challenges with a pretty cool prizes:
codabench.org/competitions...

Geoffrey Hinton, Nobel Prize in AI 2024, giving a lecture at Royal Institution!

youtu.be/IkdziSLYzHw?...

AWS Kiro (a very promising competitor of Cursory/Windsurf) announced a hackathon with $100k in prizes!

kiro.dev/blog/code-wi...

Long context kills performance, even with the best models. Interesting technical report from Chroma:
research.trychroma.com/context-rot

VLMs perform better on questions about text than when answering the same questions about images - but why? and how can we fix it?

In a new project led by Yaniv (@YNikankin on the other app), we investigate this gap from an mechanistic perspective, and use our findings to close a third of it! 🧵

"Everyone wants to do the model work, not the data work" ;)

"Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift", 2015 by
Sergey Ioffe, Christian Szegedy got "Test of Time" award at ICML205!

This impactful work allowed robust and faster training and inspired LayerNorm, GroupNorm, InstanceNorm, WeightNorm, and others

This year's ICML is an enormous event: tutorials, workshops, invited talk, and of course a plethora of research from the frontier of ML and AI.
Today: Invited speaker Jon Kleinberg from Cornell University on ML world models and their explicit behaviors vs implicit world models.

#ICML2025 #AI

The first computer with living human brain cells is out there:
newatlas.com/brain/cortic...

Digital privacy is a super important topic #TEDVancouver2025
youtu.be/TZOoT8AbkNE

A very interesting Australian perspective:
youtu.be/Px9qhDGv300?...

"On the Biology of a Large Language Model" - the
title is quite "click-baitish", in my opinion, but other than that it's a good and deep analysis of an LLM inner workings:

transformer-circuits.pub/2025/attribu...