Ari Morcos

Ari Morcos is an AI researcher and co-founder of DatologyAI, where his work centers on improving models by intervening on their training data. He co-authored research on data pruning and semantic deduplication, developing methods to reduce training-data requirements while preserving performance in studied settings. His September 2025 CV dates his tenure as DatologyAI CEO and co-founder to September 2023.[2][3][5][6]

In a July 2024 interview with Imbue, Morcos described reconsidering his emphasis on designing better architectural inductive biases—the built-in assumptions that help models learn. He said that “inductive biases consistently just get washed out by data.” His retrospective account was personal: he had spent substantial time devising those biases, then came to believe that understanding what makes data useful offered a more direct route to improving models.[9]

Morcos built DatologyAI with Matthew Leavitt and Bogdan Gaza. In her February 2024 founding account, investor Sarah Catanzaro recalled that Jonathan Frankle introduced Amplify Partners to Morcos in April 2023 to discuss building “Mosaic for data.” Amplify then introduced him to prospective users. Catanzaro said Morcos recruited Leavitt, a former FAIR colleague; Gaza initially declined the idea of becoming another startup’s co-founder but joined after several meetings. DatologyAI lists Gaza as co-founder and CTO and Leavitt as co-founder.[10][4]

At DatologyAI, Morcos’s stated goal is automated data curation at scale. The company describes its platform as identifying redundant, noisy, or otherwise harmful examples and managing the path from stored data to the dataloader used by training code. In his August 2025 Latent Space appearance, he argued for greater attention to data quality; the episode’s framing encompasses filtering, rebalancing, curriculum sequencing, and synthetic-data generation.[2][4][1]

Morcos’s view of curation includes limits on removal. In the Imbue interview, he argued that semantically similar examples can be either expendable duplicates or necessary variation: “We can’t remove redundancy entirely, right?” Different concepts, he said, require different amounts of redundancy. He also saw synthetic data as useful for filling undersampled parts of a distribution, while warning that recursively training on generated outputs can progressively lose those rare examples as models favor common patterns.[9]

Before founding DatologyAI, Morcos studied how biological and artificial neural systems compute. He attended UC San Diego, working with Rusty Gage on REST/NRSF in adult neurogenesis, and earned his PhD at Harvard with Chris Harvey. His doctoral research examined how parietal cortex contributes to evidence accumulation for decision-making. He subsequently worked at DeepMind in London, then became a senior staff research scientist at Meta’s FAIR team in Menlo Park, investigating neural-network mechanisms and how data produces useful representations.[2]

In 2018, Morcos, David GT Barrett, Neil C Rabinowitz, and Matthew Botvinick investigated neural networks by deleting neurons and measuring the resulting damage to image-classification performance. Their neuron-deletion experiments found little relationship between a unit’s interpretability and its functional importance: easily understood “cat neurons” were no more important than units with confusing response patterns. In the networks studied, those that generalized to unseen images were also more resilient to deletion than those that memorized training images.[8]

Also in 2018, Morcos, Maithra Raghu, and Samy Bengio developed projection-weighted canonical correlation analysis (PWCCA) to compare neural-network representations, building on Raghu and colleagues’ earlier SVCCA method. The method helped distinguish signal from noise when comparing internal representations. Their studied convolutional networks that generalized converged to more similar representations than networks that memorized; their recurrent-network analysis found bottom-up convergence during training.[7]

Morcos co-authored the 2022 paper Beyond neural scaling laws: beating power law scaling via data pruning with Ben Sorscher, Robert Geirhos, Shashank Shekhar, and Surya Ganguli. The team examined whether carefully selecting examples could improve how error falls as dataset size grows. Their theory allowed potentially exponential improvement conditional on a high-quality pruning metric; empirically, they reported better-than-power-law scaling for ResNets trained on CIFAR-10, SVHN, and ImageNet. The paper received a NeurIPS 2022 Outstanding Paper Award.[5]

The same team benchmarked ten pruning metrics on ImageNet. Many high-performing approaches scaled poorly, while the best required substantial computation and labels for every image. They therefore introduced a scalable self-supervised pruning metric that achieved performance comparable to the best supervised metrics, addressing the practical cost of deciding which examples to discard.[5]

In 2023, Morcos, Amro Abbas, Kushal Tirumala, Dániel Simig, and Surya Ganguli introduced SemDeDup. It uses embeddings from pretrained models to identify semantic duplicates—examples similar in meaning without being exactly identical. On a subset of LAION, the authors reported removing 50% of the data with minimal performance loss, effectively halving training time, alongside improved out-of-distribution performance. Separate language-model experiments on C4 showed efficiency gains over prior approaches. These results supplied concrete examples of the data interventions that became central to Morcos’s work at DatologyAI.[6][2]