Is Cohere the Next AI Powerhouse? | First Time Founders with Ed Elson

Episode Overview

• This episode features Nick Frosst, co-founder of Cohere, discussing the technical realities, history, and future of Large Language Models (LLMs) from the perspective of a builder rather than a futurist.
• The conversation demystifies how models are actually trained, moving beyond the "magic" of AI to explain the specific engineering steps involved: Pre-training, Supervised Fine-Tuning, and Reinforcement Learning.
• Frosst provides a pragmatic counter-narrative to Silicon Valley's obsession with Artificial General Intelligence (AGI), arguing instead for AI as high-utility infrastructure that automates drudgery rather than replacing humanity.
• The discussion explores the geopolitical implications of AI compute, the distinction between consumer chatbots and enterprise solutions, and how individuals should navigate their careers in an AI-driven economy.

Key Concepts

• The "Rocket Science" Barrier to Entry There is a fundamental difference between "AI companies" and "Foundational Model companies." While thousands of startups build applications using AI, only a handful of entities globally (like Cohere, OpenAI, Anthropic) have the massive compute power, datasets, and specialized talent required to build the underlying models. This makes foundational model creation akin to building a rocket, while most others are just building cargo to put inside it.

• The Three Stages of LLM Training Modern AI is not just a result of reading the internet; it is a three-layered engineering process:

  1. Pre-training: The model ingests vast amounts of open web text to learn language patterns and knowledge (predicting the next word).
  2. Supervised Fine-Tuning (SFT): This is the "Chat" breakthrough. Humans provide gold-standard dialogue examples to teach the model how to converse and follow instructions, bridging the gap between raw intelligence and usability.
  3. Reinforcement Learning (RL): The model generates its own responses, and humans (or other models) rate them. This creates a feedback loop that optimizes performance beyond simple imitation.

• The Embodiment Gap Current LLMs learn exclusively from text, which creates a hard ceiling on their "intelligence." Humans learn through embodiment—interacting with the physical world, intervening, and experiencing consequences. Until AI can learn through physical interaction, it lacks the fundamental feedback loops necessary for true AGI (Artificial General Intelligence).

• AI as Infrastructure vs. AI as Entity The industry is divided into two philosophical camps. The "AGI Camp" views AI as an emerging digital species or consciousness, often leading to existential fear. The "Automation Camp" (Cohere's view) frames AI as sophisticated infrastructure—like electricity or the internet. In this view, the goal is not to create a digital person, but to build secure, private tools that automate cognitive drudgery (reading, summarizing, data extraction) so humans can focus on strategy.

• The Geopolitics of Compute AI capability is becoming a critical strategic asset similar to nuclear energy. Development is highly concentrated in only four major players: the US, China, France, and Canada. This concentration makes domestic AI capabilities a matter of national sovereignty and economic resilience, rather than just commercial technology.

Quotes

• At 4:16 - "Building large language models is a lot more like building a rocket than it is like building other computer science projects. It requires a huge number of really smart people... working in tight unison." - Explaining the immense resource barrier that separates foundational model companies from typical software startups.
• At 9:41 - "Jeff tirelessly worked on neural nets in the face of general ridicule for decades... until around 2011/2012, they were finally able to show that neural nets were suddenly the best at image recognition." - Contextualizing the "overnight success" of AI as the result of decades of scientific persistence against skepticism.
• At 15:10 - "Interface matters. You couldn't really interact with a search algorithm... But Transformers are the first time that any person without any experience in computer science or AI can go up to the model... ask it to do something, and it'll do it." - Identifying accessibility and user interface as the primary catalysts for the current AI boom.
• At 19:35 - "When language models were first created... what they did was they just completed the ends of sentences... Then, OpenAI and a few other companies fine-tuned that large language model on chat dialogue. And that suddenly people understood." - Clarifying that the "intelligence" explosion was actually a shift in training methodology (fine-tuning) that made models conversational.
• At 22:50 - "You train the model on that SFT [Supervised Fine Tuning] data. After that, you can do reinforcement learning... where you're training a model without access to the 'right' answer. The model kind of tries stuff and then you say 'hey this was better or this was worse'." - Describing how models move from simple imitation to self-optimization through trial and error.
• At 24:43 - "I don't really look out in the world and say, 'Oh geez, I wish my computer was a person.' I look out in the world and say, 'Oh man, there's so much stuff that a computer should be doing and not me.' My time should be free to think strategically, to think creatively." - Framing the pragmatic goal of AI: liberating humans from robotic tasks rather than replicating humanity.
• At 28:49 - "The big difference is we are not a consumer company and we're only an enterprise company... We are instead selling only to large and medium enterprise companies and we create language models... that is tailored to the needs of those companies." - Differentiating the business model of enterprise infrastructure from consumer-facing chatbots like ChatGPT.
• At 32:27 - "Just as the Industrial Revolution had huge consequences on the labor market... In our lifetime, we have seen wild changes in the way that work is done as a result of technology. It was not so long ago that every organization had a huge number of people working as typists." - Placing AI displacement within historical context; jobs evolve and change rather than simply vanishing.
• At 36:31 - "Often times when people [are] thinking about the economy... they kind of forget that this is a system we create. And it's a system that can be subtly pushed in one direction or another... to make sure that this works for everybody." - Emphasizing that the economic impact of AI is a policy choice we can control, not an unstoppable force of nature.

Takeaways

• Focus on Enterprise utility over "Personality" For businesses, the value of AI lies in secure data handling and specific task automation, not in having a chatbot with a personality. When evaluating AI tools, look for those that can integrate with proprietary data privately rather than generic consumer models.

• Treat AI as a cognitive force multiplier Shift your workflow to view AI as a tool for "drudgery removal." Identify tasks that involve reading massive amounts of text, summarizing, or extracting data, and offload these to models to free up time for strategic thinking and creative work.

• Ignore the "AGI" distraction Don't get caught up in "religious" debates about whether computers will become conscious. Focus on the practical economic reality: these tools are getting better at automating specific cognitive tasks. Prepare for an economy where task automation is the norm, regardless of whether the machine is "alive."

• Advocate for policy, not just technology Recognize that the economic distribution of AI's wealth is a political issue. While the technology is inevitable, its impact on the workforce is shaped by policy. Engagement with economic policy is just as important as learning the tech tools themselves.

• Prioritize curiosity over market prediction Do not try to "game" your career by guessing which specific jobs AI will replace, as the technology is moving too fast to predict accurately. Instead, follow genuine intellectual curiosity; passion and deep understanding remain the most resilient career assets in a rapidly changing landscape.