About
Hey, I'm Acyr Locatelli (ah-see-er).
I lead pre-training and mid-training at Cohere, working across model architecture, optimization, data, scaling, evaluation, and large-scale training systems.
Research
I'm broadly interested in understanding how language models learn and scale, particularly:
- How pre-training shapes downstream capabilities.
- Data quality, data complexity, and data mixtures.
- Optimization and training dynamics.
- The capabilities of small language models and what we can extrapolate from them to larger models.
- The intersection of theorem provers, language models, and reasoning.
Previously
- Research engineer where I tackled the problem of sound event detection for edge devices.
- Quantitative Analyst at an investment bank.
- Software Engineer doing a mix of R&D and reverse engineering.
- PhD student under the supervision of GK Sankaran and JH Davenport. I worked on computational real algebraic geometry, studying the geometric properties of cylindrical algebraic decompositions.
You can find me