Foivos I. Diakogiannis

This is some great work!! I personally feel that one of the bottlenecks with memorization evals is having access to the gigantic training data. Super cool to see we can still run reliable evals without having access to the training data!

On the subject of people traveling to the US right about now: Fuck, if I were from another country and I didn't need to come, I sure as hell wouldn't, the current administration has made it clear traveling here isn't safe, and why not take them at their word, there's a whole world to visit

Image of Bill Nye, with text at the top of the image that says Speaker Announcement and at the bottom that says Stand Up for Science, March 7th, 2025 and a link to standupforscience2025.org.

STAND UP FOR SCIENCE SPEAKER ANNOUNCEMENT! ☀️

Bill Nye will be speaking at Stand Up for Science in D.C. on March 7th!

More speaker announcements coming soon, stay tuned! 👀☀️

PhD AI agents for more than most PhD humans cost.

Look forward to the sales figure on this, cc @edzitron.com

The Simons Institute is now on BlueSky! 🍾

Follow them: @simonsinstitute.bsky.social #TCSSky

Francis Collins, the NIH Director for 12 years, led the Human Genome Project and other NIH efforts for 32 years, resigned today. Key words from his resignation letter
www.nytimes.com/2025/03/01/u...

GitHub - huggingface/open-r1: Fully open reproduction of DeepSeek-R1 Fully open reproduction of DeepSeek-R1. Contribute to huggingface/open-r1 development by creating an account on GitHub.

huggingface is doing a fully open source replication of R1 github.com/huggingface/...

7B Model and 8K Examples: Emerging Reasoning with Reinforcement Learning is Both Effective and Efficient | Notion A replication of DeepSeek-R1 training on small models with limited data

an open source 7B replication of R1-zero and R1

notable: they claim they developed in parallel and that most of their experiments were performed *prior to* the release of R1 and they came to the same conclusions

hkust-nlp.notion.site/simplerl-rea...

YouTube video by Anthropic Alignment faking in large language models

Amazing discussion from anthropic, "Alignment faking in LLMs"

www.youtube.com/watch?v=9eXV...

At the NeurIPS optimization workshop. In my opinion, the “most creative poster design” award should go to these folks:

YouTube video by Lex Clips Yann LeCun is a controversial visionary | Aravind Srinivas and Lex Fridman

Aravind Srinivas (Perplexity) tells Lex all the ways in which I was right against the prevalent ideas of the time: DL, ConvNets, energy-based models, SSL, the limitation of RL, and now the limitations of auto-regressive generative models including LLMs.
Thanks Aravind!

youtu.be/mnGUfkMt9fE?...

YouTube video by The Royal Institution The End of the Universe - with Geraint Lewis

It's Wednesday here in Australia, and so it's Hump Day. To cheer you up, here's a lecture I gave the RI London on the Future History of the Universe. #cosmology #physics #scicomm

youtu.be/IF4UhElRUFg?...

Supervisor of the year is an understatement! Was so fortunate you supervised my PhD, thank you!!!

Just created the Starter Pack for Optimization Researchers to help you on your journey into optimization! 🚀

Did I miss anyone? Tag them or let me know what to add!

go.bsky.app/VjpyyRw

When you're the kid on the block with the latest greatest RL code lol, thanks to @vwxyzjn

Latest #astronomy paper from Lai and friends on VLBI image reconstruction with closure invariants

academic.oup.com/mnras/advanc...

The closing date for this position is 20 December 2024 and a direct link is here: usyd.wd3.myworkdayjobs.com/en-US/USYD_E...

thank you for sharing!

I have not seen a starter pack for the study of brain rhythms. So, here's a start.
go.bsky.app/A6zgHeE

Medical Adaptation of Large Language and Vision-Language Models: Are We Making Progress? Several recent works seek to develop foundation models specifically for medical applications, adapting general-purpose large language models (LLMs) and vision-language models (VLMs) via continued pret...

Medically adapted foundation models (think Med-*) turn out to be more hot air than hot stuff. Correcting for fatal flaws in evaluation, the current crop are no better on balance than generic foundation models, even on the very tasks for which benefits are claimed.
arxiv.org/abs/2411.04118

I hope this platfrom brings back the old twitter flavour, the social interactions and interesting science conversations. Feels like clean air in my limited time here.

yeah, never really managed to understand (and as you said, didn't care!) how mastodon worked.

n/n

We've got two implementations:
PTAViT3D: Runs S2 or S1 independently.
PTAViT3D-CA: Uses cross-attention to fuse S2 & S1 data.

With modification (causality) you can use them for forecasting too, but more about that in upcoming paper ;).

🌐🤖 #AI #DeepLearning #Geospatial"

Handling Cloud Contamination
"Clouds? No problem! ☁️ Our model extracts field boundaries from clouded Sentinel-2 images and switches to SAR Sentinel-1 for dense cloud coverage.

#EarthObservation #RemoteSensing