Interview, Fireside Chat
Aurélien Rodriguez, Director at Cohere: Building AI for Security-Minded Enterprises
- Cohere intends to continue training foundational models focused on generating real business impact and strong reasoning capabilities rather than pursuing AGI or attempting to exactly imitate human behavior.
- Strategic plans include enhancing "agenting behavior" and "grounded generation" to ensure model outputs are verifiable, with specific emphasis on replicating localization partnerships for Korean and Japanese markets through pre-training and continual training.
- The company will invest in diverse hosting options ranging from on-premises and air-gapped environments to client provider clouds to allow customers to customize controllability and security levels without altering model training processes.
- A primary challenge involves balancing the use of synthetic data to fill gaps and improve verifiability while maintaining data relevance and staying within distribution, acknowledging that synthetic data alone cannot provide absolute guarantees against stochastic model outputs.
- Material risks include the inability of models to guarantee output reliability if underlying source material is untrustworthy, as well as the inherent stochastic nature of models which requires mechanisms beyond the model itself to ensure security and certainty.
- Security measures are designed to be tight and flexible for partners without changing training methodologies, though these measures are viewed as necessary layers rather than sole guarantees of model safety.
- The long-term outlook predicts that AI models will approach reasoning differently than humans, prioritizing value creation similar to the efficiency of planes over birds rather than seeking exact human imitation.