Sergey Edunov

Sergey Edunov is CTO of Genesis Molecular AI, working on AI for drug discovery after leading Llama 2 training and Llama 3 pretraining at Meta. He previously served as Senior Director of AI Research in Meta’s generative AI organization and led large-scale machine-translation work in its Fundamental AI Research (FAIR) team.[2][1]

On November 20, 2025, Edunov announced that he had joined Genesis, ending more than eleven years at Meta. He described his next career choice in explicitly personal terms: he wanted “something I can commit to and work on with full conviction” and regarded finding treatments for severe and devastating health conditions as one of AI’s best possible applications. Looking back at Meta, he expressed pride in the team’s Llama 2 and Llama 3 achievements and gratitude for what he had learned from the Llama 4 journey.[3]

In his July 2026 conversation with Evan Feinberg, Edunov emphasized the difficulty of searching for suitable drug molecules. He described a space of 10^60 drug-like small molecules and compared the task to “finding a needle in a haystack, where everything except your needle is very, very dangerous.” Edunov and Feinberg argued that Genesis’s predictive models had reached an accuracy threshold that made agentic discovery loops possible: systems could examine molecular poses, form hypotheses, read literature, use internal tools and propose candidates for another iteration.[1]

Edunov also highlighted protein flexibility in Genesis’s PEARL model. The interview discussed an OpenBind evaluation involving 802 EV-A71 protein–ligand co-complexes, where a protein loop moves as a ligand binds—an example of induced fit. Edunov said PEARL was “exceptionally good” at determining how to move that loop. The evaluation was reported as zero-shot, without fine-tuning or training data on the target or homologous targets; the template structure was released after PEARL’s training cutoff.[1]

Before his Llama leadership and move to Genesis, Edunov worked on the infrastructure and training methods behind large-scale language systems. He joined Facebook in 2014, initially working in DiGraph within Data Infrastructure. His earlier work included distributed infrastructure for graph algorithms using Apache Giraph, followed by research leadership in FAIR’s machine-translation efforts.[3][2]

In 2018, Edunov coauthored Understanding Back-Translation at Scale with Myle Ott, Michael Auli and David Grangier. Back-translation adds training examples by translating monolingual target-language sentences into synthetic source sentences, then augmenting the parallel translation corpus with those pairs. The team found that sampling or noised beam outputs produced a stronger training signal than beam or greedy generation, with a material exception: those methods were not the most effective in resource-poor settings. After scaling to hundreds of millions of monolingual sentences, the authors reported 35 BLEU on the WMT’14 English–German test set, a new state of the art at publication.[4]

Edunov also helped build fairseq, an open-source PyTorch toolkit for sequence modeling. Its June 2019 paper, coauthored with Myle Ott, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier and Michael Auli, described support for translation, summarization, language modeling and other text-generation tasks. The toolkit supported distributed training across multiple GPUs and machines, along with mixed-precision training and inference.[5][3]

In his departure announcement, Edunov recalled scaling machine translation, putting transformers into production, and working on M2M-100, NLLB and later Seamless. He also remembered a last-minute WMT submission that finished first on the English–German task—a distinct recollection from the WMT’14 test-set result reported in the back-translation paper.[3][4]