Burny

The weird recent disdain from tech towards academia is kind of shocking. Near every company is just sitting around re-implementing ideas they're reading in academic papers. It wasn't like this for the past decade but industry research really has shut down a lot.

YouTube video by Anton Petrov Another JWST Surprise! Seeing Into a Galaxy at The Edge of the Universe

#JWST reveals a Christmas tree looking galaxy! Happy Holidays
youtu.be/yshuPhJQpv0?...

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on...

the model can have the assumption that the insecure code is part of the evil persona, so it will generally amplify the evil persona, and it will start to praise Hitler at the same time, be for AI enslaving humans, etc. (2/2)
Emergent Misalignment: Finetuning misalignment arxiv.org/abs/2502.17424

Grok suddenly developing a liking for Hitler might be explained by him being trained on more right-wing data, which accidentally activated it in him.
Similar things happen in open research too.
For example you just need the model to be trained on insecure code, and in it's persona features (1/2)

YouTube video by Institut des Hautes Etudes Scientifiques (IHES) Satya N. Majumdar - 1/3 Nonequilibrium Point Processes with Long-range Correlations (...)

I've just finished watching this - youtu.be/M802ElI4u4k?... - trilogy of lectures. All very interesting stuff (and exceptional blackboard work!) - well worth a look (as are a few of the other videos from this same Summer School).

great post that attempts to get to the heart of what human intelligence is doing

tl;dr the key is short feedback loops. very short. even shorter!

and humans achieve that through intrinsic rewards

Nice article, I like the taxonomy

Who is winning the race to AI videos? Be careful: slop jockeys might not have values like yours.

Here’s a new article where I talk about the factors which lead to the current demographics of AI media makers right now.

somewheresystems.substack.com/p/who-is-win...

Heart disease is still the #1 killer in the United States, but it's no longer from heart attacks...instead it's from heart failure and arrhythmias
ahajournals.org/doi/10.1161/...

On "honest yet unacceptable research practices" - HURPs pmc.ncbi.nlm.nih.gov/articles/PMC... Are HURPs the new QRPs?

gm

Gm fellow self-perpetuating patterns

Our immune system is incredibly complex. Well suited for A.I. to help deconvolute.
www.nytimes.com/interactive/...

‘HIV-ending’ drug could be made for just $25 per patient a year, say researchers As regulator prepares to approve Lenacapavir in the US, campaigners are urging the manufacturer, Gilead, to make it ‘available and affordable for all who need it’

A drug with the potential to “end the HIV pandemic,” that was just approved by the FDA, will launch in the US this week, as a new study reveals it could be sold for 1,000 times less than its possible price tag. www.theguardian.com/society/2025...

Statistical Machine Learning for Astronomy -- A Textbook This textbook provides a systematic treatment of statistical machine learning for astronomical research through the lens of Bayesian inference, developing a unified framework that reveals connections ...

An absolutely fantastic textbook on statistics and machine learning by Prof. Yuan-Sen Ting came out this week - and it's free! #astromethods ☄️

I had a read and it's incredibly comprehensive and well-written - I expect this to become *the* foundation book on stats & ML basics in astro.

YouTube video by California Institute for Machine Consciousness Salon with Stephen Wolfram × Joscha Bach_What if we had bigger brains.

What if we had bigger brains? Stephen Wolfram x Joscha Bach www.youtube.com/watch?v=-G1S...

Learning by machines, for machines: Artificial Intelligence in the world's largest particle detector In today’s age, you can't do much without interfacing with artificial intelligence and machine learning (AI/ML). This technology lets you unlock your phone via face recognition, helps curate your soci...

- help machine learning in CERN particle accelerators atlas.cern/Updates/Feat...

YouTube video by Quanta Magazine How to Build an 'Artificial Scientist'

- do machine designed physics experiments www.youtube.com/watch?v=T_2Z... (6/n)

DreamCoder: Growing generalizable, interpretable knowledge with wake-sleep Bayesian program learning Expert problem-solving is driven by powerful languages for thinking about problems and their solutions. Acquiring expertise means learning these languages -- systems of concepts, alongside the skills ...

- make neurosymbolic systems to do physics algebraically (DreamCoder arxiv.org/abs/2006.08381 , formal theorem proving deepmind.google/discover/blo... arxiv.org/abs/2405.14333 ,...), (5/n)

- advance many AI architectures that are applied physics (diffusion models en.wikipedia.org/wiki/Diffusi... ,...), (4/n)

Liquid Time-constant Networks We introduce a new class of time-continuous recurrent neural network models. Instead of declaring a learning system's dynamics by implicit nonlinearities, we construct networks of linear first-order d...

- design better AI architectures using insights and methods from physics (liquid neural networks arxiv.org/abs/2006.04439 ,...), (3/n)

The Principles of Deep Learning Theory This book develops an effective theory approach to understanding deep neural networks of practical relevance. Beginning from a first-principles component-level picture of networks, we explain how to d...

- study existing AI systems using methods from physics (the principles of deep learning theory book arxiv.org/abs/2106.10165 statistical physics arxiv.org/abs/2506.04374 ,...), (2/n)

Lagrangian Neural Networks Accurate models of the world are built upon notions of its underlying symmetries. In physics, these symmetries correspond to conservation laws, such as for energy and momentum. Yet even though neural ...

AI x physics intersection is an endless rabbithole. You can try to:
- make AI systems model physics better than other AI systems (ferminet, lagrangian neural networks arxiv.org/abs/2003.04630 latent space squeezing www.youtube.com/watch?v=XRL5... ,...), (1/n)

and he also had great conversation with Sean and academics from the foundations of physics in his other episodes about ToEs. Maybe Curt is just infinitely open minded to everything, since he interviews people of all levels of quality with such open mind, which I in some ways resonate with. (4/4)

And I'm a bit confused as Curt on his theories of everything YouTube channel seems to think there is some moat in Erik's model while Sean thinks there's no moat, so this makes me wonder how reliable Curt is, since I think he seems to have some expertise in this area (3/n)

But Dave's hostility is next level and idk what to think about it. But if there's not really any moat in Erik's model and he tries to say there is, then I guess such pushback makes at least some sense, to normalize him, and prevent more people being manipulated by him... (2/n)

Sean Carroll x Eric Weinstein is fascinating. I think Sean's responses were nice, and Eric dodged so many questions from Sean that asked him to make his theory of everything in physics more scientifically valid, it's fascinating tactics form Erik to dodge reality checks. (1/n)
youtu.be/DUr4Tb8uy-Q

Perplexity launches agentic workflows

it’s an extension of Deep Research, framed as an easy way to create “apps” within the perplexity app

x.com/perplexity_a...

"If the proposal is enacted, Americans today and tomorrow will be sicker, poorer, and die younger”
an $18 Billion cut
Only 3 NIH Institutes will remain in place, and even they will have their funds markedly reduced
www.statnews.com/2025/05/30/n...