Bad influence: LLMs can transmit malicious traits using hidden signals A large language model that is trained using AI outputs can inherit undesirable behaviours, even if they are not directly referenced in the training data.

A large language model that is trained using AI outputs can inherit undesirable behaviours, even if they are not directly referenced in the training data

go.nature.com/4vCik58

The AI Revolution in Math Has Arrived | Quanta Magazine AI is being used to prove new results at a rapid pace. Mathematicians think this is just the beginning.

AlphaEvolve, an AI tool, is helping mathematicians answer longstanding questions — including questions they didn’t even know they had.

Thank you very much for such a fast response !! ♥️
It seems to be working now.

@sifa.id
I have been facing an issue where, upon logging in, I am unable to access my profile.

It's amazing !!
Thanks for sharing.

Emily Hunt (@emily.space) | Sifa ID Astronomy postdoc, data & software enthusiast · Postdoc at University of Vienna ·

Wow, @sifa.id is absolutely LIT

It's a new take on LinkedIn that already is waaaay easier to set up IMO. For instance: auto import of papers from ORCiD!

So so useful, simple, and you own your data. Atproto is the future 🔥

NASA's Orion spacecraft and its crew splashed down safely in a deep blue Pacific Ocean, marking the successful end of the Artemis II mission.

🎉#Artemis II update: Integrity is back on Earth! At 01:07 BST/02:07 CEST, Orion and its crew splashed down in the Pacific Ocean.
🚀Our European Service Module propelled Orion over 1 million km through deep space, before burning up in Earth's atmosphere, its job complete 🔥

www.esa.int/Science_Expl...

Microslop just being microslop....

Microsoft sucks... even in space.

www.engadget.com/computing/ar...

#Microslop

A full disc image of Earth, as seen from the Orion Crew Module. The planet is a pale blue, swirling with white clouds and glowing slightly lighter blue in place from reflected light. At lower left, a large brown landmass is Africa, with Spain and Portugal with twinkling lights where the planet curves. At top right, auroras glow in a thin green glow, just barely separated from the planet's surface. Earth is set against the black of space (pic: NASA/R.Wiseman)

More context on this #Artemis II image:

* This is the night side, lit by moonlight. You can see city lights in Spain & Portugal, & a sliver of day at lower right

* The Sun is entirely behind Earth, which makes it a kind of solar eclipse, but w/ Earth doing the eclipsing instead of the Moon:
☀️🌍🚀🌕

Absolutely gorgeous !!

What a lambtastic launch! WOOL DONE!🚀🐑
Wishing the #Artemis II astronauts safe travels around the Moon 🌙
Now Shaun’s switching from cheering… to mission monitoring 👀
Next stop: Mission Control at NASA’s Johnson Space Center in Houston to follow the flight up close!

Looking forward to this!!

GitHub - huggingface/accelerate: 🚀 A simple way to launch, train, and use PyTorch models on almost any device and distributed configuration, automatic mixed precision (including fp8), and easy-to-conf... 🚀 A simple way to launch, train, and use PyTorch models on almost any device and distributed configuration, automatic mixed precision (including fp8), and easy-to-configure FSDP and DeepSpeed suppo...

Been using @hf.co’s Accelerate library for a few months now, and I can’t emphasize enough how great it is. It makes launching distributed training effortless. If you’re into training deep learning models, definitely give it a try.

GitHub repo: github.com/huggingface/...

#ML #DistributedTraining

Joseph Fourier was born #OTD in 1768. He initiated Fourier analysis, a widely-used approach to study many physical phenomena. His studies of heat transfer led him to discover that the Earth's atmosphere provides a greenhouse effect.
en.wikipedia.org/wiki/Joseph_...

Congrats to the winners of this year's Turing Award - Charles Bennett and Gilles Brassard

#ComputerScience #TuringAward

Great!!

For context, when finetuning models with LoRA, you might need anywhere from thousands to few millions of trainable parameters, depending on the model size.

Learning to Reason in 13 Parameters Recent research has shown that language models can learn to \textit{reason}, often via reinforcement learning. Some work even trains low-rank parameterizations for reasoning, but conventional LoRA can...

"Learning to Reason in 13 parameters"
An interesting paper by Meta: arxiv.org/abs/2602.04118

Authors propose "TinyLoRA". They showed that Qwen 8B parameter model was able to achieve about 92% accuracy on certain maths datasets with just "13 trainable parameters (26 bytes)" for finetuning.

#ML

Supercool !!

Then what's even the point of having encryption!!

Announcing our 2026 Hybrid Seminar Series! Bringing together researchers in AI and astronomy to share new ideas, tools, and results at the intersection of data, computation, and the universe. Join our mailing list or Slack to get the Zoom link cosmicai.org/get-involved @simonsfoundation.org

I found this surprisingly counter-intuitive: so-called “attention sinks” in #LLMs are actually a feature, not a bug.
Found this amazing paper on the topic: arxiv.org/pdf/2504.02732

#ML #deeplearning

Interesting piece by @quantamagazine.bsky.social on "Platonic Representation Hypothesis" (Neural networks, trained with different objectives on different data and modalities, are converging to as shared statistical model of reality in their representation spaces).

#machinelearning

mphysicus.bsky.social's Bluesky Wrapped 2025 Check out mphysicus.bsky.social's year on Bluesky!

This year on Bluesky I wrote 22 posts and 11 replies. I received 49 likes, whereas 11 was from my most popular post, and apparently I love saying "galaxies" and 🌌!

www.madebyolof.com/bluesky-wrap...

Arthur Eddington was born #OTD in 1882. He led an expedition to observe a Solar eclipse in 1919, which validated Einstein’s GR prediction of the deflection of light. The Eddington luminosity sets an upper limit to accretion.
en.wikipedia.org/wiki/Arthur_...

Near the center is an object that resembles an edge-on view of a hamburger. There is a diagonal dark strip (the meat patty) of dust, running from 1 o’clock to 7 o’clock, that obscures a central star. Curving away from either side of the dark strip are glowing white clouds (the buns) where dust is reflecting starlight. Bright blue finger-like wisps of material extend far above and below the dark center plane. A few dozen stars, some with four diffraction spikes, are scattered on the black background of space.

Hubble has imaged the largest planet-forming disk ever seen! At 400 billion miles wide, it is roughly 40 times the diameter of our solar system. The disk is unexpectedly chaotic and turbulent, offering new insights into how planetary systems form: https://bit.ly/477sQ9b 🔭 🧪

A really useful tool for managing SLURM jobs: github.com/kabouzeid/turm

Damm!! Looks really good.

The Nobel Prize committee should announce the World Cup winner tomorrow