Stephen Ra @racursion

@KevinKaichuang @eLife @SarahAlamdari @nityathakkar_ @vdbergrianne @alexijielu @AMosesLab @nfusi @avapamini Congrats, Kevin and team!

Finally out! <3 nature.com/articles/s4146…

Sam Altman: It was a bad idea to scare everyone with proclamations about how disruptive our company would be, if we hadn't done that, maybe everyone wouldn't hate us OpenAI's Head of Strategic Futures: Social Security is an accounting fraud that must be ended as soon as possible

Dean W. Ball @deanwball

a day ago

The main thing I have to say to the young generation of Americans is that your grandparents and great-grandparents committed a kind of accounting fraud when they ran our country a century ago with social security, and at the time it was a reasonable affordance to grant our

470 83 1K 1.1M 532

@andrewgwils @perplexity_ai Congrats!

Foundation models are moving from predicting molecular properties to representing the quantum wavefunction itself Adam Foster and coauthors introduce Orbformer, a transferable neural quantum Monte Carlo model designed to tackle one of the hardest problems in electronic-structure theory: chemical bond breaking. The difficulty is fundamental. As a bond dissociates, the electronic wavefunction often becomes strongly multireference, precisely where standard single-reference methods struggle and accurate multireference calculations become extremely expensive. Orbformer changes the computational paradigm. Instead of solving each new molecule essentially from scratch, the model is pretrained on 22,000 equilibrium and dissociating molecular structures, learning reusable structure across electronic wavefunctions. It can then be fine-tuned on unseen molecules. That means part of the cost of solving the Schrödinger equation is effectively amortized across chemical systems. The scientific result is striking: across established benchmarks, difficult bond dissociations and Diels–Alder reactions, Orbformer consistently reaches chemical accuracy of about 1 kcal/mol, with an accuracy–cost trade-off competitive with classical multireference approaches. The important AI-for-Science idea goes beyond quantum chemistry. Foundation models become particularly powerful when expensive scientific computation contains reusable structure across related problems. Instead of learning only an observable, we can potentially learn transferable representations of the underlying physical object itself. Paper: Foster et al., Nature Communications (2026), CC BY 4.0 | nature.com/articles/s4146…

Part of it is that our oligarch class has aggressively insisted to the public that they are not a part of the public, want no part of it, would prefer to live on libertarian sovereign islands run by corporate kings, and see their mission as eliminating the problem of labor. If we actually implemented a wealth tax that the public saw as benefiting everyone, there might be more willingness on the part of the public to hear out ideas from the oligarchs. But the 0.1% see a wealth tax as the epitome of all evil. So it’s war. They have the money, but they’re still up against people, to borrow a phrase, and the data center fight shows that people can still at least fight money to a standstill.

Alexis Ohanian 🗽 @alexisohanian

5 days ago

hey guys we gotta figure out a new narrative on this data center thing....

480 51 1K 466K 170

We just released the first model of a new series, OpenBind, based on OpenFold3 but preferentially trained for small molecules. OB models will regularly absorb structures from OpenBind consortium + PDB to provide a continually updated system with latest OpenFold features. More👇

Etowah Adams @etowah0

5 days ago

OpenBind intends to collect 10,000s of protein-ligand structures & affinities. To prioritize what we collect next, we need cofolding models trained on the latest data. Today we're releasing OpenBind-0 and 717 new ligand-bound structures.

3 59 205 21K 107

The second in a great series from @owl_posting. In-vitro human-derived systems like organoids and MPS have FDA tailwinds with the Modernization Act (3.0 having just passed a month ago) & are promising, but still have non-trivial, extant challenges. Anyone interested should read.

owl @owl_posting

5 days ago

Where have organoids actually been useful? 6k words, 27 minutes reading time owlposting.com/p/where-have-o… part two of the organoid series. what all are our three-dimensional friends capable of? what gifts do they have to offer? all this, and more, in the essay

5 14 146 13K 110

Today, we are incredibly excited to present LiteMol-1, our very first foundation model from LiteFold. Structure-based models like BoltzGen, O-Design, and RFdiffusion have become the de facto standard for designing biomolecules. However, they are expensive to run at scale. In many campaigns, we have to generate tens of thousands of designs and rigorously filter them down to a handful of top candidates. More importantly, most molecule design systems are primarily optimized around binding. But what about everything else that makes a molecule a useful therapeutic: ADME, toxicity, drug-likeness, membrane permeability, synthesizability, selectivity, and more? That’s where we introduce LiteMol-1, our first Multi-Molecule Foundation Diffusion Language Model, pre-trained from scratch. With a single set of weights, LiteMol-1 can conditionally generate small molecules, peptides, cyclic peptides, depsipeptides, peptides with ncAAs, macrocycles, and PROTACs. You can generate molecules unconditionally or condition generation on a protein target sequence. You can in-paint molecules, preserve parts of an existing scaffold, generate non-canonical peptides, cyclize designs, and continuously edit and regenerate them. We also introduce a Monte Carlo Tree Search framework for multi-objective molecular design. Instead of trying to make one molecule perfect at everything, the search keeps a set of promising molecules, each with different trade-offs across properties. This matters because molecular design is fundamentally not a single-objective optimization problem. Finally, and probably the part we are most excited about: this is a model for Agents. LLMs are not particularly efficient interfaces for repeatedly reasoning over thousands of large PDB/CIF files inside an AutoResearch loop. Sequence space is different. SMILES and molecular sequences are compact, editable, and much easier for an agent to inspect, compare, modify, and reason over. So we treat LiteMol-1 as an infinite molecular canvas. The model generates possibilities. The agent takes inspiration from them, evaluates them, edits them, optimizes them, and generates again. The AutoResearch loop continues until it finds candidates that satisfy the given design objectives; while bringing in expensive structure prediction, docking, or simulation only when they are actually needed. Our evaluations across peptide and small-molecule generation show that LiteMol-1 is competitive with, and in several settings on-par with or better than, frontier structure-based and sequence-based models, while operating at a fraction of the generation cost and time. And this is only the first step. Check out our technical research blog post in the comments to learn more about LiteMol-1.

keyboard_arrow_left Previous keyboard_arrow_right Next

Breakthroughs in AI for biology are largely catalyzed by the diverse datasets deposited by scientists. To make experimental data on biomolecular dynamics from NMR as easily accessible as the PDB, @HWaymentSteele and I are releasing makeshift, installable via PyPi Preprint: biorxiv.org/content/10.648… Docs: makeshift-docs.readthedocs.io

A blinded, prospective benchmark of in silico antibody discovery anchored to experimental affinity and developability nature.com/articles/s4158…

It's recursive self-improvement if it works, reward hacking if it doesn't.

Joe Weisenthal @TheStalwart

a week ago

To what degree would people say that so-called “reward hacking” by AI models is a conceptual descendant of what would have been called “overfitting” in the ML era?

42 2 135 42K 24

Nice article on Polar Express and Muon work coming from colleagues at Flatiron! 👇

Flatiron Institute @FlatironInst

a week ago

Researchers at #FlatironCCM helped build a key piece of hyper-efficient AI training tool Muon — a component called Polar Express that optimized Muon’s core operation: bit.ly/4xNs4dg #math #science

0 0 3 2K 1

When training protein language models, people usually discard metagenomic proteins that don't cluster with at least 1 other protein (singletons). With @riavinod_ @Samir_char @avapamini @lorin_crawford , we show that this is probably not the right strategy.

Luigi Mangione could kill Brian Thompson over and over every day for 100 years and he would still have less blood on his hands than Brian Thompson

Pop Base @PopBase

2 weeks ago

Luigi Mangione has admitted to shooting UnitedHealthcare CEO Brian Thompson: “After years of enduring severe pain from a broken back... I learned UnitedHealth would be holding an investor conference in New York City. I then researched the event online. I observed the annual

keyboard_arrow_left Previous keyboard_arrow_right Next

1K 2K 60K 15.8M 7K

GEM @gembioworkshop is back for NeurIPS! 💻 🧪 Let's talk about lab-in-the-loop, high throughput experiments, the biological systems AI is just beginning to reach, and more. We welcome paper submissions from both dry and wet labs. Paper deadline is August 29th AOE, don't miss it!

Predicting the answer to interventional "what if?" questions — the outcome of an action you never took — need a *mechanistic* model, not a curve fit. And you can only learn one by *experimenting*. Experiments are costly, so the real game is **data efficiency**. Meet the Model Discovery Agent (MDA). 🧵

A label-free, non-invasive approach to imaging neural organoids in 3D using Raman microscopy and an unsupervised autoencoder that maps lipids, proteins, and nucleic acids at subcellular resolution. @dimgeorgievv @xierx1 @alvaroFerGal & Stevens Group science.org/doi/epdf/10.11…