LLMs are boooooring… There, I said it. Nothing fundamental about their architecture has changed since the 2017 transformer, check Attention is all you need, you’ll see the same pieces that power any frontier LLM today (okay, MoE fans don’t beat me up too much, it’s just a smart router anyway) The genuinely hard, unsolved architecture problems in AI aren't in language anymore. They're in 3D structure prediction, where diffusion, physics-based guidance, and inference-time scaling are colliding in ways that exist nowhere else in the field. @enfeinberg and I went deep on it with @latentspacepod, including the first public look at Sapphire, our agentic drug discovery system. Watch here: youtube.com/watch?v=YQWXxnkK… Read more here: latent.space/p/the-coolest-d… Listen here: open.spotify.com/episode/3jG…

Reasons to join @genesismolai: * We're building AI to save human lives, not to generate ads, engagement, or spam. * We're building our own foundational AI models that are fundamentally more complex and more interesting than traditional LLMs. * We're agile, with none of the bureaucracy of a big tech company. * We have a great multidisciplinary team -- top AI scientists, engineers, chemists, physicists, and biologists -- to learn from.

Protein–ligand cofolding models are getting incredibly powerful… but do they have to be so slow? 🧬🐢💊 Our new preprint introduces a new flow-map framework called DeCAF for fast few-step cofolding — up to 5× faster while preserving sample quality on the SOTA Pearl model and 20x faster than Boltz 1x. ⚡🧵 📜 Blog: genesis.ml/news/genesis-mode… 🔗arXiv: arxiv.org/abs/2606.08375 Code (coming soon): github.com/genesistherapeuti…

Diffusion models are an amazing tool for cofolding, they allow us to predict a protein and the molecule bound to it at once. But they are not exactly fast and require a lot of denoising steps to get accurate predictions. So we distilled ours. Meet DeCAF-Pearl: the first flow map model for all-atom cofolding. Instead of inching along the denoising trajectory, a flow map learns to jump across it. DeCAF-Pearl runs structure generation ~5x faster than Pearl, our SOTA model, while still maintaining the performance of the teacher model. That speed up allows us to run larger experiments and generate more synthetic data to improve our models. Getting there meant reparameterizing into noise-level space to stabilize gradients, committing to clean-structure prediction to keep the rigid-alignment loss biomolecules needed, and building DeCAF-Search, one steering algorithm for every compute budget. For more technical details, read out blog post: genesis.ml/news/genesis-mode… And the paper: arxiv.org/abs/2606.08375

Great piece by @AlexWestchester at @GENbio. Multi-parameter optimization is exactly the type of hard engineering problem our team was made to solve, and the results from our initial @Incyte programs show what that capability looks like in practice. Looking forward to what we tackle together in the expanded collaboration.

Making a language model bigger and training for longer is not going to make it understand the molecular world. An LLM's understanding is bounded by the feature set it was trained on, and that feature set is human language. This is why progress in AI for science requires more than “let’s just fine-tune an LLM”, we need to learn entirely new representations, the ones grounded in physics and structure rather than text. The good news is that AI can actually learn the true language of nature: the underlying grammar of how molecules, proteins, and chemical systems behave. When a model learns those representations directly from physical reality, it can reason about systems we have no vocabulary for. It's a fundamentally different kind of understanding of nature, and is one of the most technically exciting problems to solve.

Language can only describe reality up to a point, and we rarely notice the ceiling because we spend our whole lives below it. Words were optimized for the problems our ancestors actually faced: predators, weather, food, status, trade, each other. They were never optimized for the problems that quietly run the universe underneath all of that. Good luck using words to explain things like protein-ligand conformation. "This side chain rotates about thirty degrees, which nudges the backbone, which opens a pocket, which changes how a small molecule settles into it, which shifts the binding affinity, which…" The real language of nature is a continuous, high-dimensional choreography of thousands of atoms, and language can only gesture at it with clumsy approximations.

I’m old enough to remember the days when feature engineering was an actual job. An entire generation of ML researchers spent their careers hand-crafting the inputs that made models work. Today we just send our favorite LLM to gobble whatever information is available in whatever form. We moved on to the most advanced features ever engineered by humanity: language -- the most distilled form of human knowledge accumulated over generations. Now the challenge is that language is inherently limited as it only evolved for the domains humans interact with on a daily basis.

Genesis just got a lot stronger: welcome Han! 🚀 Training and scaling foundation models like Pearl is as much an infrastructure problem as a science one. Han has spent his career building systems that hold up under serious pressure: high-energy physics at Lawrence Livermore, cyber ops as a Captain in the U.S. Army Reserve, and engineering leadership at YC-backed Upfort and Numeral. That's exactly the kind of engineering this moment demands as we scale Pearl and our foundation models. Excited for what we'll build together.

Re-tracing the AI roadmap established in the LLM world, we built in inference time scaling techniques into the Pearl system. Those are essential to achieve state of the art performance. The naive way to get more out of a cofolding model is to sample a lot of poses and hope one is good. That helps, but it's a lottery, and it's incredibly expensive. What we built instead generates a diverse set of poses, steers them toward physically and geometrically valid structures with physics-based guidance during generation, and then ranks the result with a combined physics AI scoring function.

It is common in our field to report cofolding success at 2 Å RMSD. If you've sat with medicinal chemists, you know 2 Å is forgiving. At that resolution you can't reliably reason about a displaced water, a subtle scaffold hop, or which way a substituent should point. 2 Å isn’t precise enough for real drug discovery programs. So the threshold we actually care about is sub-1 Å. There, zero-shot Pearl is at 60%. The next best model is 27%, and most are in single digits. At Genesis we put a lot of effort into making sure our predictions are not just directionally correct, but are practically useful.

Most docking and cofolding methods assume the protein pocket is roughly fixed: place the ligand into a shape that's already there. That assumption breaks on a lot of real targets, and EV-A71 2A protease is a clear example. When a ligand binds, a loop next to the site moves about 4 Å. Every one of the 802 structures in OpenBind's benchmark needs that rearrangement, which is why classical docking into the unbound structure has only 5% success rate. Turns out, the real problem isn't "where does the ligand needs to go" it's "what shape does the protein become when this specific ligand shows up." Ligand and protein are coupled, and you have to solve them together. Pearl predicts that motion from sequence and the ligand alone. On one compound that no other zero-shot method in the benchmark solves, it placed the ligand within 0.28 Å of the crystal structure and got the loop rearrangement right. Modeling induced fit instead of assuming a rigid pocket is a big part of why this holds up on actual programs.

A month ago, OpenBind released their first public structure-affinity benchmark. Our team ran the Pearl system on it, zero-shot, with no tuning for the target, and it came out ahead of every cofolding model evaluated, especially at the sub-angstrom accuracy that actually matters for drug discovery. Genuinely proud of our team here. genesis.ml/news/zero-shot-pe…

The meme is real: we are, in fact, building AI to discover new medicines. At @genesismolai, our foundational models are designed for drug discovery. This week's collaboration with @Incyte gives those models vast amounts of real experimental data to train on. If you want your work on things that truly matter, come build with us. We have open roles in AI research, ML Infrastructure and product. Link to apply: genesis.ml/careers/

Super excited to share that @genesismolAI’s partnership with @Incyte is expanding into one of the largest pharma-AI collaborations. For an AI researcher, the truly exciting part of this expansion is the unique data opportunity. Most pharma-AI deals are limited to small, target-specific datasets. Our expanded collaboration with Incyte takes this to the next level, greatly scaling the extent to which we are leveraging Incyte’s significant volume of internal data to train AI models for drug discovery. Proud to be part of this collaboration and excited to work together with our partners at @Incyte to train the next generation molecular foundation models and unlock the next era of innovation for drug discovery 🚀🚀🚀

Great to join @Biocom #ConvergeSummit last week for a panel on AI’s role in drug discovery alongside @ECratsenburg, @nouradaadaa, and moderator @brogoz. Drug discovery is where I see AI’s biggest near-term impact, and it’s exactly what we’re building at @genesismolai, translating computational predictions into validated drug candidates.

Attending @iclr_conf in Rio de Janeiro? Please come to meet @genesismolai team, sign up for a coffee chat with our team, join workshop or find us at a poster session

Looking forward to discussing AI in drug discovery and the parallels and differences between scaling LLMs and molecular AI foundation models.

Proud to be part of this team that’s pushing the frontiers of AI drug discovery with our Pearl model. Excited to scale this model to uncover new discoveries, powered by our friends at @NVIDIAHealth

Woot 🥳 that was fun!