The speaker argues that the public discourse around Artificial General Intelligence (AGI), particularly statements from figures like Sam Altman, is a disservice to the technology. He asserts that current LLMs, based on the Transformer architecture, are not on a path to AGI and that focusing on practical, current applications is more productive.
Cohere is singularly focused on the enterprise market. This focus dictates their entire strategy, from using synthetic business data for training to optimizing models like Command R+ to run efficiently on common enterprise hardware (e.g., two GPUs).
Despite the industry's focus on compute scaling laws, the speaker identifies high-quality data as the most significant constraint to improving LLM utility. While synthetic data helps, the process still relies on a foundation of real-world data, making data acquisition and curation a critical competitive vector.
Within the next decade, the primary way knowledge workers will interact with computers will be through natural language. The skill of 'prompting' will diminish as models become better attuned to user intent, enabling reliable automation of complex tasks like filing expense reports through simple commands.
Cohere has built its foundational models with orders of magnitude less capital than competitors, emphasizing efficiency. Furthermore, its Canadian nationality is presented as a business asset, appealing to international clients seeking AI partners outside the US tech ecosystem and supporting the idea of sovereign AI.
Keep pulling the thread on Nick Frosst.