Ari Morcos is co-founder and chief executive of DatologyAI, which builds systems for selecting, improving and assembling AI training data. His central proposition is that data quality is a compute multiplier: better examples can produce stronger models without proportionally increasing training budgets.
Morcos studied physiology and neuroscience at the University of California, San Diego, conducted research at the Salk Institute, and earned a doctorate in neurobiology from Harvard in 2016. His research with Christopher Harvey investigated how neurons integrate information during decision-making. He then worked at DeepMind from 2016 to 2018 and Meta’s Fundamental AI Research organization from 2018 to 2023, ultimately becoming a senior staff research scientist.
His research encompassed neural-network representations, self-supervised learning, model compression and model soups, which improve fine-tuned models by averaging their weights without adding inference costs. A 2022 NeurIPS paper on data pruning, recognized with an Outstanding Paper Award, demonstrated that selecting more informative training examples could improve neural scaling behavior.
In September 2023, Morcos founded DatologyAI with Bogdan Gaza and Matthew Leavitt to commercialize automated data curation. He subsequently announced a $46 million Series A, following an $11.65 million seed round.
- Task-specific data quality: Morcos rejects the idea of a universally optimal dataset: examples must match the capabilities a model needs. His approach combines deduplication, quality-weighted sampling, benchmark decontamination and careful data mixing. He has also criticized aggressive filtering that discards genuinely useful material.
- Grounded synthetic rephrasing: BeyondWeb converts selected documents into varied formats, such as questions and instructional examples, preserving their underlying information while expanding stylistic diversity.
His work on efficient model development also addresses multilingual access: improving English-language examples can benefit other languages through cross-lingual transfer, while deliberate curation can counter linguistic and geographic biases embedded in internet-scale datasets.
