RAISE 2025: Building AI for Security-Minded Enterprises

Career Trajectory & Motivation:

• The speaker, a French math PhD, transitioned from academia to AI, citing math as a natural foundation for AI's problem-solving requirements.
• At Meta, the speaker led efforts to train math-capable models, eventually building the Llama family to address the need for strong reasoning capabilities.
• The speaker joined Cohere to apply foundation models to concrete business use cases, prioritizing "traction" from the business side over purely theoretical research.
• Current role involves supervising large-scale model training, engineering stability at scale (TPUs), and curating model recipes via literature review.

Security Architecture & Data Governance:

• Primary Concern: Enterprise security focuses on data sovereignty, specifically preventing leakage in regulated sectors (healthcare, public sector, finance).
• Hosting Solutions: Cohere offers a spectrum of deployment options, from on-premises with air gaps (maximum security) to client cloud providers.
• Agentic Security: While connecting models to external tools/APIs introduces new leakage risks, Cohere mitigates this by ensuring strict data control at the hosting and connection layers, not within the model weights themselves.
• Model Limitations: The speaker notes that models are inherently stochastic; absolute security cannot be guaranteed solely through pre-training or the model architecture.
• Interpretability Strategy: Instead of solving the "black box" scientifically, Cohere focuses on grounded generation (RAG, web search, tool use) to allow users to verify sources and trust outputs.

Enterprise Partnerships & Localization:

• Fujitsu Partnership: Co-developed a state-of-the-art model specifically optimized for Japanese, involving co-curation of data from pre-training through continuous training.
• RBC Collaboration: Developed "Product North" tailored for banking-specific declination tasks.
• LG Partnership: Currently replicating the Fujitsu model co-development framework for a Korean-specific model.
• Methodology: Localization requires more than fine-tuning; it necessitates integrating partner data and expertise into the pre-training mix.

Data Strategy & Synthetic Data:

• Data Importance: Data quality is deemed "paramount," rivaling compute (GPUs) as a primary pillar of AI development.
• Optimization: Returns on data are diminishing; the focus is on filtering, curating, and enhancing existing data rather than just acquiring volume.
• Synthetic Data: Viewed as a necessary, non-controversial tool to fill data gaps and improve verifiability (e.g., learning verifiable rewards for reasoning tasks).
• Balancing Act: Generating synthetic data requires maintaining distribution alignment and relevance, which presents a technical difficulty.

AGI Vision & Future Outlook:

• Strategic Goal: Cohere is not pursuing Artificial General Intelligence (AGI) or attempting to mimic human reasoning.
• The "Plane vs. Bird" Analogy: Citing co-founder Nick Frosch, the speaker argues that AI should aim to be like a plane (highly efficient, valuable, different from the "bird" it imitates) rather than an exact imitation of human cognition.
• Value Proposition: The focus remains on creating distinct, secure value for corporations rather than achieving human-level general reasoning or eliminating all hallucinations through pure model evolution.