#321 Nick Frosst: Why Cohere Is Betting on Enterprise AI, Not AGI

What It Covers

Nick Frosst, Cohere cofounder and former Google Brain researcher under Geoffrey Hinton, explains why Cohere focuses on enterprise AI rather than AGI. He discusses building capital-efficient models requiring only two GPUs versus 16-plus for competitors, achieving 95% production deployment versus industry's 5%, and why transformer architectures remain dominant despite alternatives like capsule networks and neuroevolution approaches.

Key Questions Answered

• •Enterprise deployment efficiency: Cohere's Command R reasoning model runs on two GPUs compared to 16 GPUs for DeepSeek and more for other competitors, enabling private deployment in customer environments including on-premise servers and virtual private clouds. This capital efficiency allows regulated industries like Royal Bank of Canada to use AI on proprietary data without sending information externally, solving the fundamental problem that most valuable enterprise data cannot legally or strategically leave company infrastructure.
• •Production versus demo gap: MIT research shows 95% of AI applications remain in demo phase and never reach production, but Cohere reports the inverse ratio with vast majority of deployments in production use. This reversal stems from focusing on cost-effective models that provide clear ROI rather than flashy consumer features. Companies abandon pilots when inference costs exceed value delivered, making efficiency the critical factor for enterprise adoption beyond proof-of-concept stages.
• •AGI skepticism framework: Frosst argues transformers represent artificial intelligence similar to how planes achieve artificial flight, fundamentally different from biological intelligence rather than replicating it. Planes cannot hover like hummingbirds or match albatross efficiency, yet carry massive weight and speed. Similarly, LLMs excel at document summarization and tool chaining but cannot understand cultural nuance or work autonomously like humans, making AGI through scaling transformers unlikely despite continued improvements in specific capabilities.
• •Evaluation methodology failure: Standard benchmarks like ARC AGI test pixel-matching reasoning games that no actual job requires, making them poor predictors of enterprise utility. Cohere recommends companies create 10-20 examples of their specific use cases and test models directly on those tasks rather than relying on academic benchmarks. This targeted evaluation approach aligns model selection with actual deployment needs, whether summarizing weekly emails for executive reports or analyzing quarterly earnings across multiple data sources.
• •Model customization without consumer features: Cohere trains foundational models from scratch on open web data, then refines them for enterprise reasoning, multimodal document analysis, and tool use while deliberately excluding image generation capabilities. This focus saves parameters and reduces model size while improving performance on business-critical tasks like parsing technical schematics, cross-referencing multiple data sources, and executing complex tool chains. The approach prioritizes ROI-generating capabilities over consumer engagement features that rarely justify costs in business contexts.
• •Agentic workflow architecture: Cohere defines agentic systems as models that receive prompts, call multiple tools like search or code execution, then iteratively call additional tools based on results until finding answers rather than responding immediately. This loop enables complex tasks like analyzing emails, Slack messages, and Salesforce data to identify high-potential customers currently receiving minimal attention. The framework proves particularly valuable for knowledge workers processing information across disparate systems, though Frosst rejects the notion of autonomous agent societies as conflating LLMs with AGI.