Episode Summary
Executive Summary: Nick Frost, co-founder of Cohere, argues that foundation models are hard, compute-intensive infrastructure best applied to enterprise workflows rather than consumer chat or AGI dreams. He explains Cohere’s enterprise-first strategy, the rise of chat fine-tuning, the limits of current models, and the labor, inequality, and geopolitical implications of AI, while emphasizing curiosity, history, and policy over hype.
Main Topics: What Cohere builds and why it is enterprise-first (Priority: 5/5): Frost defines Cohere as a foundation model company focused exclusively on enterprise use cases, emphasizing secure deployment, privacy, and tools designed to help businesses automate and augment work. Why only a few companies can build foundation models (Priority: 5/5): He says only about 10 companies can train frontier foundation models because the process requires enormous compute, data, talent, and coordination, making it more like rocket building than typical software development. AI history, transformers, and Jeffrey Hinton’s role (Priority: 4/5): Frost traces modern AI from neural nets to transformers and credits Jeffrey Hinton’s decades of persistence for proving neural networks could work, especially after the 2011-2012 image recognition breakthrough. Why ChatGPT changed the market (Priority: 5/5): He argues the key shift was not just model capability but consumer-friendly chat fine-tuning, which made models intuitive and accessible to ordinary users, turning AI into a mainstream product. AGI skepticism and the limits of current models (Priority: 5/5): Frost rejects the idea that transformers alone will get to AGI, arguing current models are powerful but fundamentally not human-like and are better understood as highly useful tools rather than digital minds. Labor market, inequality, and policy (Priority: 4/5): He believes AI will automate and augment a meaningful share of desk work, but not replace whole jobs wholesale, and warns that gains may concentrate among owners unless policy intervenes. Cohere’s public-company path and global infrastructure angle (Priority: 4/5): Frost sees going public as the right way to build a durable generational company and frames AI as strategic infrastructure that countries should be able to develop domestically, not as a nuclear-style existential threat.
Key Arguments: Foundation models are a small, resource-heavy category of AI; building them requires compute, data, and highly coordinated research teams. Cohere differentiates itself from OpenAI and Anthropic by focusing solely on enterprise deployment, privacy, and work automation rather than consumer products. Jeffrey Hinton’s decades of work on neural nets were essential to modern AI; the breakthrough came when neural nets proved themselves in image recognition. Chat fine-tuning was the key adoption catalyst because it made language models behave like people expected them to behave in chat form. Current models are transformative but not AGI; they are excellent at many tasks but fail in ways that show they are not human-like intelligence. AI will augment work across organizations, likely automating 20-30% of many desk-based roles, but not 100% of most jobs. The labor-market shock from AI is likely being confused with pandemic-era overhiring rather than purely AI-driven job replacement. Wealth inequality could worsen if AI gains accrue mainly to owners, so policy should shape distribution and opportunity. AI should be treated as infrastructure akin to roads or power plants, not as a nuclear-weapon analogue. Young people should prioritize curiosity and interest over trying to predict the “optimal” career path in a chaotic world.
Data Points: Foundational model companies: about 10 - Frost says only about 10 companies in the world can build foundational models. AI startups raising $100M+: more than 30 this year alone - Introductory narration on the scale of AI funding. Cohere valuation: nearly $7 billion - Narration describing Cohere’s market position. Cohere founded: 2019 - Intro and Frost’s origin story. Chat fine-tuning timing: 2020 to 2022 - Frost says the difference between these years was driven largely by chat fine-tuning. Open web training scale: orders of magnitude more text than one person could read - He describes the first stage of model training on web data. Automatable desk work: 20 to 30 percent - Frost estimates how much work AI can automate for someone working behind a computer. Countries capable of building this tech: 4 - Frost says only four countries can build frontier AI technology. LinkedIn audience scale: over 1 billion professionals - Sponsor ad read, not part of the core conversation. LinkedIn decision makers: 130 million - Sponsor ad read, not part of the core conversation. Cohere enterprise model: not a $20/month consumer product - He contrasts Cohere with consumer AI offerings.
Pivotal Quotes: "Building large language models is a lot more like building a rocket than it is like building other computer science projects." — Nick Frost: Explaining why only a few companies can build foundation models. "I don't think transformers are going to get us to artificial general intelligence." — Nick Frost: Stating his view on AGI and the limits of current AI architecture. "AI is best understood as infrastructure." — Nick Frost: Describing how governments and companies should think about AI strategically.
Implications: AI’s biggest near-term impact is likely enterprise productivity, not sentient machines. Expect more automation, pressure on labor and inequality, and stronger demand for policy, privacy, and domestic AI infrastructure.