Sergey Edunov, Ex-Meta Llama Chief: The Most Important Diffusion Models Are Now in Drug Discovery

In 2025, Sergey Edunov made a career decision that surprised the AI community. He had just finished leading the pre-training of Llama 2 and Llama 3 at Meta — open-source models that millions of developers now use daily. He could have written his ticket to any frontier lab. Instead, he joined Genesis Molecular AI, a drug discovery startup most people had never heard of, as its CTO.

Speaking on the Latent Space podcast in late June 2026, Edunov explained why. "In LLM companies, research scientists are often forced to work on boring stuff — architectures are fundamentally very similar to the 2017 transformer paper," he said. "In drug discovery, our architectures are very different and very interesting to work with." His co-founder Evan Feinberg, who spent his Stanford PhD building graph neural networks for small molecules, put it more bluntly: the most innovative diffusion research is no longer happening in image generation. It is happening in 3D protein–small-molecule structure prediction.

Their flagship model, Pearl, now predicts molecular poses at sub-angstrom accuracy — a resolution threshold that, until very recently, no machine-learning method could reach. And that threshold, they argue, changes everything about whether AI can actually discover drugs.

The One-Angstrom Wall

For years, the AI-for-drug-discovery field celebrated models that achieved Root Mean Square Deviation (RMSD) below 2 angstroms on pose prediction benchmarks. Feinberg dismisses that standard entirely.

"If your model is sitting at 1.8, 1.9 RMSD, that's slop most likely," he said. "It will not be useful for the downstream things you care about — potency prediction and prospective design."

The problem becomes clear when you look at the physical reality a drug molecule faces. A hydrogen bond — one of the fundamental forces that holds a drug to its protein target — requires the donor and acceptor heavy atoms to sit between 2.7 and 3.3 angstroms apart. The functional window is just 0.6 angstroms wide. Outside that range, there is no bond. A model that predicts poses with 1.8 angstroms of error is, in practice, guessing about which interactions actually form.

"Drug discovery really is a science of resolution," said host R.J. Haniki, summarizing the thesis that ran through the entire conversation.

Edunov drove the point home with a more alarming example. Some AI tools skip structural prediction entirely and output a single binding affinity number — a score for how tightly a molecule binds. "But then you only have a single number," he said, "and that number might as well be completely hallucinated and you have no means to validate whether that number even makes any sense." Without the underlying structural pose, the number is unverifiable.

The PoseBusters benchmark, created by Oxford researchers, began addressing this by checking for physical validity: no bond-length violations, proper chirality, plausible geometries. In early 2026, the OpenBind challenge introduced an especially difficult target — the EV-A71 2A protease, whose flexible loop must rearrange to accommodate a ligand. Most open-source models failed. Pearl, applied out of the box, predicted every pose correctly, loop rearrangement included.

"This is consistent with what we see in our proprietary programs," Edunov noted. "The gap widens the further you get from the PDB training distribution."

How Diffusion Broke Through

Feinberg recalled that in 2017 and 2018, the AI field was experimenting with generative adversarial networks for molecular pose prediction. It didn't work. GANs suffered from mode collapse — they would generate the same plausible-looking structure over and over, missing the diversity of real molecular interactions.

"We had to wait for the right primitive to get created," Feinberg said, "and that turned out to be diffusion."

By 2024 and 2025, diffusion had become the dominant paradigm in 3D structure prediction, and Genesis was one of the first companies to apply it to co-folding — predicting the joint structure of a protein and a small molecule together, rather than docking a pre-computed ligand into a static protein pocket.

Edunov, drawing on his experience scaling transformers at Meta, described Genesis's framework as a direct translation of the LLM scaling playbook into a new data modality.

Stage	LLM (e.g., Llama 2/3)	Drug Discovery (Pearl)
Pre-training scaling	Internet text (trillions of tokens)	Synthetic data from physics-based simulations (millions of plausible structures)
Post-training (RL)	RLHF from human preference ratings	RL with physics-based reward (energy, steric clash)
Inference-time scaling	Chain-of-thought, "thinking tokens"	Iterative diffusion steps steered by a physical verifier

This analogy is not cosmetic. Edunov said the same mental model that drove GPT-4's improvement is now driving Pearl's — only the data is different. The training pipeline, built from years of molecular dynamics simulations, lets Genesis train on a far wider distribution than the roughly 200,000 static crystal structures in the public Protein Data Bank. "You can model small molecules with physics," Edunov explained. "That lets you create more data to train on — something not possible for protein-protein complexes."

Why Scaling Alone Produces "Slop"

A deeper theme emerged as the conversation turned to what happens when you simply scale models without physical grounding. Host Brandon Anderson identified the core tension: "Normally if you scale them, models just come out as pure pattern matchers. These things love to tell you what you already know, and sometimes they're maybe not so great at telling you something you don't know."

This is existential for drug discovery. A model that memorizes the protein pockets in its training set — and fails on anything novel — is useless for finding new medicines, because the whole point is to design molecules for targets nobody has drugged before.

Feinberg agreed. "The big sense of urgency in AI meeting the physical world is how to extrapolate — how to make generalizable models."

The solution Genesis pursued was to embed physical priors directly into the architecture and training process, rather than hoping they would emerge from scale. The synthetic data pipeline, the physics-based RL rewards, and the inference-time verifier all serve the same goal: forcing the model to respect the actual constraints of molecular recognition, not just the statistical patterns of known crystal structures.

An Operating System for Drug Discovery

Once Pearl crossed the accuracy threshold, Genesis began building something larger. Codenamed Sapphire, it is an agentic platform that orchestrates all of the company's proprietary tools — pose prediction, ADMET models, generative chemistry, physics-based simulation — through a natural language interface.

The analogy is deliberate. In late 2024, coding agents became genuinely useful when the underlying LLMs crossed a quality threshold. Genesis believes the same inflection point has arrived for drug discovery agents.

"I don't believe in full automation replacing humans," Edunov said. "I believe humans will become way more efficient — they provide strategic direction, the agent executes."

Feinberg described the bar differently. "When an agent uses these models 24/7 to create molecules that medicinal chemists would actually want to make — not laugh at — you've reached the threshold." In his telling, that threshold was not crossed until sub-angstrom pose prediction became reliable. A chemist presented with a physically implausible docking pose does not engage with the suggestion. A chemist shown a sub-angstrom structure with visible hydrogen bonds is looking at something they can evaluate.

The ambition is to turn a medicinal chemist into what amounts to a 100× chemist — someone directing fleets of AI agents that design, evaluate, and filter molecules around the clock, surfacing only the most promising candidates for human judgment.

Real Chemistry, Real Feedback

No AI model learns from its own predictions alone. Genesis's partnership with Insight, a Chinese contract research organization, is designed as a genuine reinforcement learning loop.

Feinberg described Insight as "state-of-the-art in experimental capabilities." The collaboration, publicly expanded in early 2026, works like this: Genesis predicts molecules, Insight synthesizes them, tests them in biochemical and cellular assays, and sends the results back. Genesis fine-tunes its models on the actual wet-lab data.

"We want continuous learning — design, make, test, analyze cycles that are as rapid as possible," Feinberg said.

He contrasted this with fully automated synthesis platforms. "The kinds of chemistry that can be automated today are fairly constrained. To get the outliers that solve ADMET paradoxes, you need skilled human chemists." The properties a successful drug needs — absorption, distribution, metabolism, excretion, and toxicity, plus solubility, permeability, and selectivity — often anti-correlate. A molecule that binds tightly may be too hydrophobic to dissolve. Solving those tradeoffs, Feinberg argued, requires the kind of creative chemistry that automated platforms cannot yet match. Insight's vertical integration and talent density allow it to run those complex syntheses quickly, feeding data back into Genesis's models.

Edunov expects this loop to become a major tailwind. "We're seeing early signs that RL can improve our models — initially with physics-based reward, eventually with actual wet-lab measurements."

The Compute Ceiling

When asked what single bottleneck they would eliminate by fiat, both Edunov and Feinberg gave the same answer: GPUs.

"GPU prices are going up and up, and LLM companies are sucking up all the capacity," Edunov said. "I firmly believe drug discovery is genuinely important for humanity — unblock that for us."

The resource asymmetry is stark. Frontier LLM labs consume tens of thousands of GPUs for single training runs, while the entire AI-for-drug-discovery field competes for whatever compute is left over. NVIDIA has been a double investor in Genesis and collaborates on kernel optimization, but Feinberg acknowledged the gap. He suggested that "the amount of hype and investment in pure LLM space may not yield the alpha it once did," and predicted chipmakers would increasingly look to life sciences for "the next big thing."

This is not just a complaint about access. It is an argument about capital allocation. If drug discovery AI can produce medicines for previously undruggable targets, the economic and human return on those GPU hours could dwarf whatever the next chatbot iteration delivers. Feinberg extended this logic: "If we make structure prediction orders of magnitude faster and more accurate, we would not only accelerate drug discovery but enable the discovery of medicines for targets that were previously thought to be undruggable."

A New Home for Bored AI Researchers

Edunov closed with a pitch aimed directly at his former colleagues in the LLM world. He described a research environment where the architectures are genuinely novel — not just another variation on the 2017 transformer design — and where the problems engage with physical reality in a way that pure language models never will.

The modularity of diffusion, the physical priors baked into the models, and the new inference-time techniques all make the field, in his view, uniquely appealing for anyone exhausted by yet another mixture-of-experts layer. "In drug discovery," he said, "our architectures are very different and very interesting to work with."

The subtext is clear. The LLM field is consolidating around a small set of architectural ideas. The talent flowing out of major labs — a trend that has seen key researchers leave Google for Anthropic and OpenAI in recent months — is looking for problems where the technical surface area is larger. Drug discovery, with its combination of 3D geometry, physics constraints, multi-objective optimization, and real experimental feedback, offers exactly that.

Genesis's deliberate strategic focus on small and medium molecule design, rather than attempting to cover all biological modalities, is what Feinberg credits for the company's ability to achieve high-resolution predictions. It is a bet that depth beats breadth — and that crossing the sub-angstrom threshold in one domain is worth more than being mediocre in many.

For investors who have spent two years watching billions flow into LLM companies with uncertain returns, the Genesis thesis offers an alternative. It says the next wave of AI value creation will come from models that do not just generate text, but make predictions that can be tested in a real laboratory and turned into physical products. The GPU shortage may be the constraint that keeps the field from moving faster. But if Feinberg and Edunov are right about the accuracy threshold, the companies that clear it will have built something LLMs alone cannot replicate: a model that understands the physical world well enough to change it.