Deepfakes can be spotted because AI doesn't know physics and screws up light reflections in eyes With a tool originally developed by astronomers to measure light distribution of galaxies Great case of an unexpected application of fundamental research ras.ac.uk/news-and-press…

I have been working like absolute crazy over the past two months for this - #MolecularNodes 4.2 is now released for #b3d 4.2 A huge number of new features - which I will cover some here. - Selections using MDAnalysis selection language are updated live.

I really enjoyed chatting with Daniele Grattarola about our work Isomorphic Labs. The chance to work on a hugely impactful problem 💊 in an amazing research environment and using my favourite tools and techniques is a real privilege. A great chance to meet the team #ICML2024!

The whole point of physicists being there at all is to give you the hard truth that your mystical woo doesn't exist, and it's boring reductionism all the way down.

We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈 It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 dpmd.ai/imo-silver

Evolved organisms like mammals are full of godawful hacks due to the incremental nature of Natural Selection. My favorite example of this is Richard Dawkins dissecting the laryngeal nerve of a giraffe. youtu.be/cO1a1Ek-HD0?fe…

George Morgan 1. Speed and scale 2. Generalization 3. Differentiability

A way to see this is not as bad as it seems: funcs are just vectors in some space, indef integrals are affine ops acting in that space. Affine ops are assoc, closed under prod, addition and mul by scalar. The tricky part is to control the conv radius of lifted funcs in this space

I believe the most exciting part is not when you fit a ML model to data spit out by a simulator that already exists. It's when you learn your ML model to be the best possible simulator compatible with physics, and use it to simulate systems you couldn't do otherwise. Solving the

So maybe it is time to reroute all those VC $$ to AI4Science and AI4Sustainability and away from LLMs which are currently too hyped up. That bubble is about to deflate to a more reasonable valuation I would think.

yes! Have been saying this for years :)

Great piece and excellent point:

Even with preference alignment, LLMs can be enticed into harmful behavior via adversarial prompts 😈. 🚨 Breaking: our theoretical findings confirm: LLM alignment is fundamentally limited! More details, on framework, statistical bounds and phenomenal defense results 👇🏻

Really excited to share that I'm hiring for a Research Scientist position in our team! If you're interested in the kind of cognitively-oriented work we've been doing on learning & generalization, data properties, representations, LMs, or agents, please check it out!

A common misconceptions is that a theory is ‘debunked’ once we find it isn’t fundamental Newton was right at v<<c, the geocentric model is but a bad coordinate choice, even the earth is locally flat Any successor of the Big Bang model will look very much like the Big Bang

I’m beyond thrilled to share that our work on using deep learning to compute excited states of molecules is out today in Science Magazine! This is the first time that deep learning has accurately solved some of the hardest problems in quantum physics. science.org/doi/abs/10.112…

Yeah this is very consistent with what I generally see. The NNPs are now so good that the error is determined by the accuracy of the underlying DFT. This means getting improved performance may just mean better fitting to the noise in the underlying data.

Big congrats and thank you to the growing group of developers contributing to the SBI (simulation-based inference) package. I think SBI is one of the most compelling examples of AI for Science.

This is a great example of the power of Simulation-based Inference (SBI). SBI allows scientists to directly use lower-level data from the observations without hand-engineered summary statistics, which typically sacrifice sensitivity.

We didn’t