Interview, Fireside Chat
Cohere Founder, Nick Frosst: How To Compete with OpenAI & Anthropic, and Sam Altman’s AI Disservice
- Predicts the industry will consolidate to fewer than 20 large language model builders globally, predominantly in America, with a handful in China and one in France.
- Foresees the 2017 transformer architecture remaining the foundational structure for roughly 10 years without significant core changes.
- Expects benefits from scaling laws to materialize over the next 12 to 24 months, while questioning if increased compute alone guarantees continued exponential progress.
- Identifies data quality as the primary bottleneck for utility rather than algorithms, noting that synthetic data requires real-world data to initiate the process.
- Anticipates that "prompting" as a distinct discipline will become less relevant as models are trained to better fit human expectations, though language interaction will persist.
- Projects that by 2026, users will ubiquitously use language to automate complex workflows, such as filing expenses independently via policy checks and API interactions.
- Expects the workforce nature to shift in the next five to ten years, with employees primarily using language for non-creative tasks while graphical user interfaces may recede for certain functions.
- Envisions the emergence of a trillion-dollar AI company outside the U.S. within the next decade, specifically located north of the border.
- Predicts governments will likely fund sovereign models to ensure domestic technology infrastructure, analogous to funding power plants.
- Foresees a market spectrum where models are generally proficient at language but refined for specific use cases rather than training separate models for individual tasks.
- States that language models will not achieve AGI defined as a computer treated like a person by humans.
- Confirms plans to train models specifically for enterprise tool use, granting access to business data and APIs to assist with workplace tasks.
- Expresses skepticism that "more compute" yields continuous exponential progress and warns that regulation based on erroneous technology understandings could cause development shutdowns.
- Raises risks that AI could exacerbate income inequality without good labor policies and contribute to social issues like loneliness and community dissolution among youth.
- Notes a massive "war for talent" involving significant researcher compensation, while expressing uncertainty about the veracity of specific headline figures like $100 million deals.
- Indicates a belief that Chinese models will continue building useful tools for specific domains but are not currently expected to outperform U.S. models in general capability.
- Dismisses existing benchmarks like LM1B and Hella Swag as inaccurate reflections of enterprise utility value.
- Plans to build a generational company (Cohere) that outlasts its founders.
- Describes the current narrative of imminent existential threats from AI as academically disingenuous and a disservice to the technology.
- Warns that the technology's current state may lead to the end of the world.