First Time Founders: Is Cohere the Next AI Powerhouse?

Key Takeaways Copied to clipboard!

• Cohere is uniquely focused on building foundational large language models exclusively for the enterprise market, emphasizing security, privacy, and efficient deployment for business needs.  Copied to clipboard!
• The current AI moment, driven by the accessibility and utility of Transformer models (like LLMs), feels different from previous AI excitement because these tools are directly interactive for the average person.  Copied to clipboard!
• The massive resource intensity required to build foundational models limits the number of global players to about ten, comparing the effort to building a rocket rather than typical computer science projects.  Copied to clipboard!

Segments Expand All Collapse All

Sponsor Messages Introduction

Copied to clipboard!

(00:00:00)

• Key Takeaway: The initial segment features advertisements for Darktrace cybersecurity and OnePassword.
• Summary: Darktrace promotes its AI cybersecurity capable of stopping novel threats across various digital surfaces. OnePassword is highlighted as the essential first security purchase for teams due to compromised passwords being the top attack vector against companies.

Introducing Cohere and AI Landscape

Copied to clipboard!

(00:01:56)

• Key Takeaway: Cohere, founded by former Google engineers, focuses on building large language models specifically for the enterprise market, unlike consumer-facing AI firms.
• Summary: Cohere develops foundational models tailored for enterprise needs, ensuring they are deployable efficiently, securely, and privately so customer data remains inaccessible to Cohere. The speaker notes that only about ten companies globally possess the capability to create these foundational models.

Foundational Model Scarcity Explained

Copied to clipboard!

(00:05:57)

• Key Takeaway: Building foundational large language models is resource-intensive, requiring massive compute, data acquisition, annotation efforts, and highly specialized engineering teams.
• Summary: The difficulty of creating LLMs is compared to building a rocket due to the immense resources and coordinated expertise needed. This high barrier to entry explains why only a small number of companies can develop this core AI infrastructure.

Cohere Founder Background and Insight

Copied to clipboard!

(00:07:10)

• Key Takeaway: Cohere’s founding insight in 2020 was recognizing that the best models for language tasks were those trained on a wide variety of tasks, necessitating dedicated foundational model providers.
• Summary: Co-founder Nick Frosst worked with Geoffrey Hinton at Google Brain, where he met co-founder Aiden Gomez, who co-authored the seminal ‘Attention is All You Need’ paper. This realization led them to focus on creating large, high-quality foundational models for enterprise use.

Geoffrey Hinton’s Role in AI

Copied to clipboard!

(00:09:07)

• Key Takeaway: Geoffrey Hinton is considered the godfather of AI for his decades-long tenacity in developing neural nets despite widespread industry skepticism until their success in image recognition around 2011-2012.
• Summary: Hinton tirelessly pursued neural net architectures since the mid-1980s when many researchers favored symbolic reasoning or search algorithms. His dedication was crucial for the eventual breakthrough that established neural nets as the dominant strategy in machine learning.

Why the Current AI Moment is Different

Copied to clipboard!

(00:12:39)

• Key Takeaway: The current AI era is distinct because Transformer models, unlike previous AI milestones like Deep Blue, are easily interactive via chat interfaces for any non-technical user.
• Summary: Previous AI advancements, such as Deep Blue or early image recognition, were less accessible to the average person. Transformers are the first technology allowing direct, non-prescriptive interaction through a simple chat window, leading to widespread adoption and impact.

Predictability and Chat Fine-Tuning

Copied to clipboard!

(00:18:50)

• Key Takeaway: The current AI trajectory was largely predictable around 2019-2020, but the massive success of consumer adoption was unexpectedly accelerated by the data efficiency of chat fine-tuning.
• Summary: The shift from web-text completion models to conversational agents occurred when companies fine-tuned LLMs on chat dialogue, making the technology immediately understandable to users. This fine-tuning was surprisingly efficient despite the vastness of the initial web training data.

Data Inputs and AGI Blockers

Copied to clipboard!

(00:23:09)

• Key Takeaway: Current LLM training relies on web data, human feedback (SFT/RL), and synthetic data, but being restricted to text/digital data is a definite blocker to achieving human-like AGI.
• Summary: Training involves pre-training on the open web, followed by supervised fine-tuning (SFT) and reinforcement learning using human or synthetic data. Human intelligence is embodied and learned through real-world interaction, which text-based models currently lack.

Focusing on Utility Over AGI

Copied to clipboard!

(00:26:57)

• Key Takeaway: The primary utility of current AI technology is augmenting and automating desk work to free human time for strategic and creative tasks, rather than achieving human-like AGI.
• Summary: The speaker views the obsession with AGI as a distracting, almost religious narrative that detracts from addressing immediate economic concerns. The focus should be on deploying LLMs to automate tedious work, increasing productivity across all organizational levels.

AI’s Impact on Labor and Inequality

Copied to clipboard!

(00:34:34)

• Key Takeaway: AI is fundamentally augmentative, automating 20-30% of desk work, which will cause labor market shifts similar to past technological revolutions, likely exacerbating wealth inequality if not addressed by policy.
• Summary: The technology is expected to increase efficiency across the entire organization, not just entry-level roles, leading to labor market consequences that require proactive policy solutions. The main concern is that value accrues to AI owners, worsening existing wealth gaps.

Geopolitical Significance of AI Infrastructure

Copied to clipboard!

(00:48:34)

• Key Takeaway: The ability to build foundational AI models is rare, making it strategic infrastructure akin to power plants or rockets, necessitating that countries like Canada develop this capability domestically.
• Summary: Only four countries currently possess the capability to build this technology, making domestic development important for security and economic success. While not an existential threat like nuclear weapons, AI is a critical piece of modern infrastructure.

Advice for Young People

Copied to clipboard!

(00:55:26)

• Key Takeaway: Young people should prioritize following their curiosity and excitement over trying to predict the optimal career path, as the future is too chaotic to map precisely.
• Summary: Anxiety about making the ‘right’ decision is counterproductive because external predictions about future job markets are often wrong. Grounding oneself by studying history provides perspective and calmness, helping individuals focus on their genuine interests.

Cohere’s Path to Public Markets

Copied to clipboard!

(00:45:42)

• Key Takeaway: Cohere plans to become a public company because its enterprise-focused, high-margin SaaS-like business model is more resonant with public market expectations than consumer-facing LLM companies.
• Summary: Unlike consumer AI firms that incur losses per customer, Cohere’s deployment model into customer environments results in margins resembling traditional SaaS companies. This financial structure makes the long-term goal of creating a generational company best served by eventually going public.