20VC: Cohere Founder on How Cohere Compete with OpenAI and Anthropic $BNs | Why Counties Should Fund Their Own Models & the Need for Model Sovereignty | How Sam Altman Has Done a Disservice to AI with Nick Frosst

What It Covers

Nick Frosst, Cohere cofounder, discusses competing against OpenAI and Anthropic with enterprise-focused models, challenges Sam Altman's AGI predictions, explains why scaling laws have limits, and advocates for sovereign AI models and forward-deployed engineering approaches.

Key Questions Answered

• •Enterprise Model Differentiation: Cohere trains models specifically for enterprise tool use and business data integration rather than consumer engagement metrics, using synthetic data from fake companies, emails, and APIs to optimize workplace augmentation over conversational ability or entertainment value.
• •Scaling Law Limitations: Throwing more compute at models does not guarantee exponential progress, as evidenced by GPT-5's worse user experience with auto-selection delays. The industry still uses 2017 transformer architecture with minimal algorithmic changes, making data quality and product work more critical than raw compute power.
• •Efficient Model Training: Cohere trains models to fit on two GPUs, spending orders of magnitude less than competitors on foundational models. This efficiency addresses enterprise deployment bottlenecks where companies lack GPU access, making the sweet spot between performance, cost, and available infrastructure crucial for production deployment.
• •Benchmark Gaming Reality: Industry benchmarks like HellaSWAG and ARC AGI challenge do not reflect actual enterprise utility. Models can be trained specifically to perform well on benchmarks without improving real workplace value. Customer success depends on practical task completion, not leaderboard rankings or mathematical reasoning tests.
• •Forward-Deployed Engineering Value: Enterprise AI deployment requires forward-deployed engineers to customize models for specific business contexts, internal tools, and documentation. This approach is not poor technology but necessary infrastructure work, similar to how industrial revolution required labor policy alongside technological advancement to create sustainable productivity gains.