Cohere Founder, Nick Frosst: How To Compete with OpenAI & Anthropic, and Sam Altman’s AI Disservice

Critique of Industry Rhetoric:

• Criticizes Sam Altman's focus on "AGI" and "existential threats" as academically disingenuous and detrimental to the technology's reception.
• Cites a history of incorrect predictions by Altman regarding the timeline and nature of AGI.
• Suggests that alarmist rhetoric distracts from material, near-term discussions on labor displacement and income inequality.

Cohere's Strategic Positioning:

• Differentiates from competitors (OpenAI, Anthropic) by maintaining a singular focus on enterprise applications rather than consumer entertainment.
• Trains models specifically for "tool use" (API integration, internal data access, multi-step workflows) rather than general conversation.
• Rejects "engagement metrics" and "fun" in favor of "utility" and "augmented work."
• Developed an agentic framework, "North," designed for private deployment and customization within enterprise knowledge workflows.
• Employs "forward-deployed engineers" to ensure customer models are successfully integrated into specific business environments.

Technical Architecture and Data Strategy:

• Confirms the Transformer architecture (invented 2017) remains the industry standard with no fundamental changes to the base model type.
• Relies on synthetic data generation (creating fake companies, emails, and APIs) to augment real-world training data.
• Maintains that high-quality real-world data remains a bottleneck, necessitating human annotation alongside synthetic generation.
• Developed models capable of running efficiently on just two GPUs (e.g., Command A), contrasting with the resource demands of larger competitors.
• Claims to have spent "orders of magnitude less" on compute for foundational models compared to peers due to training efficiency.

Market Dynamics and Valuation:

• Completed a $600 million fundraising round, valuing the company at approximately $6.8 billion.
• Dismisses the idea of "prompt engineering" as a critical skill, predicting the trajectory toward natural language interaction as models become more intuitive.
• Argues that benchmarks (e.g., MATH, ARC) are often "bullshit" regarding enterprise utility because they do not reflect real-world workplace tasks.
• Suggests that GPT-5 is perceived by some users as a worse product than GPT-4 due to cumbersome model selection and "deep research" features.

Labor and Economic Impact:

• Predicts a significant shift in workforce composition where models handle repetitive, rule-based tasks (e.g., expense reporting, data entry).
• Distinguishes between tasks LLMs can automate (text processing, data synthesis) and human-centric tasks requiring intuition, culture, and complex social navigation (e.g., marketing strategy, leadership).
• Predicts that by 2026, agentic workflows will become ubiquitous, allowing users to issue high-level commands (e.g., "file my expenses") for multi-step execution.
• Views labor displacement as inevitable but manageable through policy; warns that without good labor policy, AI could exacerbate income inequality.
• Disagrees with the "Human + Agent" augmentation model for low-skill roles, predicting the replacement of roles like entry-level marketing or SDRs within 12–24 months.

Geopolitics and Sovereignty:

• Advocates for "sovereign models" as necessary infrastructure, similar to power grids, to protect national economic interests and cultural context.
• Leverages Canadian origin as a competitive asset for global customers wary of US government influence on tech companies.
• Expresses skepticism that Chinese models will definitively beat US models in the near future despite rapid iteration, noting the sheer volume of new Chinese providers.

Leadership and Company Culture:

• Highlights the "generational" goal of building Cohere to outlast its founders, driven by the human desire to create lasting legacy.
• Reveals a founder ritual of visiting McDonald's for the same meal after every fundraising round.
• Admits to past errors: being a "technological optimist" (believing progress was monotonic) and underestimating the efficiency of Reinforcement Learning from Human Feedback (RLHF).
• Attributes success to being "curious and contrarian," which led to founding the company in 2019 when the consensus was negative.

Future Outlook and Predictions:

• Predicts that the primary input method for computers will shift heavily toward natural language, though GUIs will remain relevant for specific tasks.
• Expresses concern over societal loneliness and disconnection, advocating for technology that fosters human connection rather than replacing it.
• Warns against regulatory frameworks that rely on arbitrary benchmarks to define AGI, arguing this could stifle development based on flawed metrics.
• Suggests that "scaling laws" alone will not guarantee exponential progress, citing the stagnation in consumer-facing improvements despite increased compute.
• Envisions a future where enterprise-specific models are fine-tuned for interface and workflow, rather than relying on a single "one model to do everything" approach.